Today we are launching conversations to measure that whole journey.

The short answer

A Traaker conversation is a simulated multi-turn buying journey, led by one of your personas against the real AI engines, that measures turn after turn whether your brand is cited, recommended, or replaced by a competitor.

You pick a persona, a starting intent, a number of turns (up to 6), the engines, the places and the languages. Traaker replays the conversation at every launch and shows you:

• the presence rate: the share of turns in which your brand is cited;
• the turn where it drops out, when it disappears along the way;
• who gets recommended instead and which sources the engine relied on;
• a score per turn, computed with the same scoring as your prompts.

Conversations are set up in the « Conversations » tab, right next to your prompts, from the « Configure prompts » button on each axis page.

Why measure beyond the first answer?

Definition of conversational presence rate (Traaker): the share of completed turns in a conversation in which the tracked brand is cited in the engine's answer. A brand cited on turns 1 and 2 of a 4-turn conversation has a presence of 50%.

An engine's first answer is often an overview. The following turns narrow the choice: the engine rules options out, brings in others, and ends up recommending one or two. That is where the decision gets made, and where a brand that started well can lose its place.

Each run boils down to one readable outcome:

Presence on turn 1 is the bridge to your prompts: it is the same measurement, extended across the whole exchange.

A conversation takes a few settings

(Screenshots are from the French interface.)

The persona: who asks the questions

Definition of a persona (Traaker): the profile of the buyer who leads the conversation (their role, context and way of writing), reusable from one conversation to the next.

A persona is described with traits that genuinely change how it questions the engine:

An expert, challenging head of SEO does not ask the same questions as a busy executive discovering the topic. Both are among your customers, and the engines answer them differently.

The starting intent and the turns

The conversation opens on one of four intents: Learning, Solution search, Comparison or Reviews. You then set a ceiling of 1 to 6 turns.

By default, each turn is written by the persona from its profile and your Brand Knowledge, reacting to what the engine just answered. You can also script a turn with a question, an objection or free text:

• one content item on a turn: it is sent to the engine as is, word for word;
• several items: the best fit for the conversation in progress is chosen at run time.

Scripted turns let you ask exactly the questions that matter to you (your prices, a frequent objection, a specific competitor), while the other turns stay natural.

Engines, places, languages and runs

A conversation is measured on five engines: ChatGPT, Perplexity, Gemini, Google AI Overview and Mistral (Le Chat). It can be run in several places and measurement languages, just like your prompts.

The number of runs (1 to 10) sets how many times the same setup is replayed at each launch. An AI engine never answers exactly the same way twice: several runs give a steadier measurement than a single exchange (we explain why in GEO visibility is a distribution).

The place reaches each engine through its own parameter

A local question ("the best GEO tool in my city") only makes sense if the engine knows where the customer is. Traaker passes the conversation's place through each engine's location parameter whenever it offers one:

In the example below, the persona asks for "the best GEO tool in my city" from Paris, and Mistral answers for Paris and the Île-de-France region.

Every turn, read as in the engine's own app

A conversation's page shows each run the way it reads in the engine's app: the Google results page with its AI Overview on the first turn, then AI Mode; the Le Chat interface for Mistral; sources then answer for Perplexity. You read what your customer reads.

Around that screen:

• a date band lists the measured days; a day without a measurement is not shown, so you never read a false 0%;
• the engine switch shows the day's presence for each engine;
• observed intent, turn by turn shows how the conversation evolved (a solution search that turns into a comparison, then a request for reviews);
• « Sources, targets and competitors » gathers the cited domains, your recognized Targets (products, locations, experts) and the named competitors, for one run, one engine, one place, one day or the whole conversation.

A score per turn, with the same scoring as your prompts

Each persona message is classified like a prompt, then the answer is scored with the matching axis rule:

A comparison turn that pits other brands against each other without naming you is flagged as such, "a comparison without your brand", rather than scored 0: in a conversation, a customer can legitimately compare other players at some point. Brand detection relies on your Targets, so an answer that cites one of your product lines or products counts as a citation of the brand.

Cadence and consumption

Conversations follow the brand's cadence, like your prompts: one launch measures prompts and conversations together. Their consumption shows in the same window footer, and each conversation can also be rerun on demand from its page.

• Elite and Enterprise plans: a monthly volume of simulated conversations is included (150 per month on Elite, custom on Enterprise), see pricing.
• Agencies: conversations are paid in credits, at 2 credits per completed turn. A turn that fails on the engine side is not billed.

Action plan: your first conversation in five steps

1. Open « Configure prompts » from one of your brand's axis pages, then the Conversations tab.
2. Pick a persona close to your real buyer, or create one from « Manage personas ».
3. Start from the intent that precedes the purchase, most often Solution search or Comparison, with 4 to 6 turns.
4. Script the decisive turn: the pricing question, the objection your sales team hears most, or the competitor that wins deals from you.
5. Read the turn where you drop out, then open « Sources, targets and competitors » to see who takes your place and which sources the engine relies on.

A conversation's opening questions are often the ones you already track as prompts: our post on the new prompt system explains how to cover them by theme and by place.

Frequently asked questions

What is the difference between a prompt and a conversation? A prompt measures the engine's answer to a single question. A conversation chains up to 6 questions and measures whether the brand stays present on each turn, when it disappears, and who gets recommended instead.

Who writes the conversation's messages? The persona, based on its profile, your Brand Knowledge and the engine's previous answer. Turns you script are sent to the engine word for word, without rewording.

Which engines are conversations measured on? ChatGPT, Perplexity, Gemini, Google AI Overview and Mistral. On Google AI Overview, the first turn is a search with its AI Overview and the following turns take place in AI Mode, as for a real user.

Why run the same conversation several times? Because an AI engine never answers exactly the same way twice. Several runs of the same setup give a more reliable presence rate than a single exchange.

How does Traaker know my brand is cited? Each answer is matched against your Targets (brand, products, locations, people). An answer that names one of your product lines or locations counts as a citation of the brand.

Do conversations account for the customer's location? Yes. The place is passed to each engine through its location parameter when it offers one (city for ChatGPT, coordinates for Perplexity and Google AI Overview, country for Mistral) and through the persona's instructions for Gemini.

Which plans include conversations? Elite and Enterprise plans include a monthly volume of simulated conversations. Agencies use them with their credits, at 2 credits per completed turn.

Available today

Conversations are available since November 5, 2026 in the « Conversations » tab, next to your prompts. Discover the Traaker platform or see the plans.

Methodology note. The rules described here (at most 6 turns, 1 to 10 runs, five engines, per-engine location passing, per-turn scoring) are those in production as of November 5, 2026. Conversations are simulated: the persona is a language model playing a buyer and does not replace observing real customers. Engine answers vary from one run to the next, which is why presence rates are read across several runs and several days.

Today we are launching a new prompt system that answers those four questions on a single screen.

The short answer

Traaker's new prompt system brings together, in one window, everything used to measure your brand in AI engines: your prompts classified on four axes (Narrative, Authority, Referral, Comparison), their language and place variants, the brand's measurement cadence, your consumption for the month, and a coverage map that crosses your themes with your places.

Three deeper changes come with it:

• A fourth axis, Comparison, for the questions where a customer is weighing several brands.
• One cadence per brand, which launches every axis together and measures each prompt once per period.
• Automatic classification of every prompt (intent, criteria, theme or use case, target), which feeds the coverage map and the scores.

It has been available on every account since November 3, 2026, from the « Configure prompts » button on each axis page.

The four measurement axes

Definition of a measurement axis (Traaker): the family of questions a prompt belongs to, and therefore how the engine's answer is scored. Every prompt belongs to one axis, which Traaker assigns when the prompt is saved and which you can change.

Comparison: your rank against the brands named

When customers hesitate, they ask a comparative question. The engine most often answers with an ordered list and a preferred brand. The Comparison score follows that rank: 100 if you are ranked first, 50 second, 33 third, 25 fourth, and 0 if the answer compares other brands without naming you. The criteria from your Brand Knowledge (price, lead times, quality…) are judged separately for each answer, which tells you which point the comparison turns on.

Comparison has its own page and its own card on the dashboard. It stays outside the overall GEO score, so that the history of that score remains comparable over time.

Narrative: tone sets the band, fidelity places the score

The Narrative score now follows a readable rule. The sentiment of the answer sets the band (positive from 70 to 100, neutral from 35 to 65, negative from 0 to 30), and alignment with your Brand Knowledge places the score within the band. A positive answer therefore always ranks above a neutral one. When an answer drifts from your Brand Knowledge, the answer detail explains the gap in one sentence.

Creating prompts without a blank page

The add button opens a composer with several sources:

• Write: one question per line, as you picture it.
• Generate: Traaker suggests questions from your Brand Knowledge, for an axis, a theme and a language you choose. On Comparison, you pick the competitors to set against you.
• Persona: questions written the way one of your personas would ask them, when the brand has personas.
• Keywords: up to 50 keywords turned into natural questions.
• CSV: a one-column file, to bring in an existing list.

Nothing is saved before review. The verification step classifies each line, then shows you the axis, the intent, the criteria or theme, and the detected target (your brand, a location, a person). You correct what needs correcting, then you add.

Definition of classification (Traaker): the automatic labelling of a prompt when it is saved, which gives it an axis, an intent, criteria (Narrative and Comparison) or a theme and use case (Authority and Referral), and the target it is about. Classification is what places a prompt on the coverage map and what selects the scoring rule for its answers.

One question, several languages and several places

The same question can be asked in French in Paris, in English in Geneva and in Spanish in Barcelona. The new system handles that as a family: an original question and its variants.

• 34 measurement languages, independent of the brand's working language. A translation is created only after you approve it, with a check that brand and competitor names were kept.
• Up to 20 places per question: Worldwide, your locations, or any city, region or address you search for. The place is passed to each engine through its location parameter when it has one, and through the instructions otherwise, to come as close as possible to the answer a customer on site gets.

The list shows each family on one line, with its flags, its main place and the number of prompts it represents. A last-run column shows when each family was last measured.

One cadence per brand, consumption you can see

Definition of cadence (Traaker): how often the brand is measured, from every day to once a week. A launch covers every axis and every engine of the brand, and each prompt is measured only once per period.

The bottom of the window always shows what your configuration consumes:

• On plans with a quota, a « Prompts » bar shows your requests for the month. Its detail spells out the calculation: number of prompts × number of engines × launches in the month = monthly requests, compared with your quota. A change that would exceed the quota is greyed out with its reason (for example "This needs 1,800 requests per month, your quota allows 1,500") before it is even sent.
• New prompts are measured right away when the month still has room, otherwise at the next launch, and a banner tells you which.
• For agencies, the cadence is set per brand (from every day to once a month, or « Manual »), and the bottom of the window shows the brand's credits per month. The « Launch » button lets you run every prompt or only those not yet measured today, with the credit cost of each choice shown before you confirm.

A prompt you no longer want to track moves to « Disabled » in one click, and comes back just as easily. Its past results stay available.

Coverage: see what you measure, and what is missing

The Coverage tab crosses, for one axis, your themes, use cases or criteria with your places. Each cell gives the number of prompts measuring that pair. A hatched cell marks a pair you do not measure yet: one click opens the generator already set to that theme and place.

It is the view that answers "am I measuring my renovation offer in Lyon?" without scrolling the list. It also counts conversations, on plans that include them.

Conversations, in the same window

On plans that include conversations (multi-turn buying journeys led by your personas), the Conversations tab sits next to the prompts. You create a conversation by choosing a persona, a starting intent, the engines, the languages and the places, and you can script some turns with your prompts, your objections or your pre-purchase questions. Their cadence and consumption appear at the bottom of the same window as the prompts.

Action plan: your prompts in five steps

1. Open « Configure prompts » from a brand's Narrative, Authority, Referral or Comparison page.
2. Look at Coverage before adding anything: the hatched cells are your blind spots.
3. Generate from a hatched cell, review the classification, add.
4. Extend your key questions to the languages and places where your customers buy.
5. Set the cadence at the bottom of the window, reading the consumption it implies.

To measure precisely what the engines cite (your products, your stores, your experts), pair these prompts with Targets. And to understand why a list of well-chosen questions beats a list of keywords, read our measurement of the queries AI generates.

Frequently asked questions

Are my existing prompts kept? Yes. Your prompts and their results history carry over as they are. Prompts not yet classified show as unclassified and become classified as soon as their text is edited and validated.

What is a request? A request is one prompt measured on one engine during one launch. A language or place variant counts as a prompt, since it is asked separately to the engine.

What is the difference between Narrative and Comparison? Narrative measures what the engine says about your brand when asked about it. Comparison measures your rank when the customer puts several brands in competition.

Why one cadence per brand? Because a launch then measures the whole brand at once, across all axes. Your axes are comparable with each other on the same date, and your consumption reads on a single line.

What happens if my configuration exceeds my quota? The change is greyed out with its reason before it is sent. If a launch no longer fits in the month, it is refused as a whole and a notification tells you, which keeps your results consistent.

Can I measure the same question in several countries? Yes, up to 20 places per question and in 34 languages. Each place is passed to the engine through its location parameter when it has one, and through the instructions otherwise.

Do agencies keep manual launches? Yes. The « Launch » button measures the whole brand, with the choice of running everything again or only what has not been measured today, and the credit cost is shown before you confirm.

Available today

The new prompt system has been live on every account since November 3, 2026. Open an axis page of your brand and click « Configure prompts », or discover the Traaker platform.

Method note. The figures in this article (four axes, 34 languages, 20 places per question, the Comparison scale and the Narrative bands) describe the rules in production on November 3, 2026. Request quotas and available engines depend on your plan. Scores remain measurements of samples of answers: a change from one day to the next is part of the engines' normal behaviour.

The keyword did not disappear. It changed hands. The user no longer types it; the engine writes it, sometimes dozens of them, before composing an answer. Google calls this query fan-out. Several platforms now offer to extract those machine queries and resell them to you as a new keyword list to work on.

We tested whether that list means anything. 224 calls to the Gemini API, 3 commercial intentions, 8 equivalent paraphrases of each, 2 models, and 4 hypotheses whose refutation criteria were fixed before data collection. Two of the four failed. We publish them as failed, which is what makes the rest contestable.

Three results, all measured:

• Two runs of the exact same prompt, word for word, share only 3 to 9% of their machine queries. On one model, 62 calls produced 237 distinct queries.
• On the model you are using today, half the calls issue no query at all. The answer comes straight out of the model's weights. On one of our three intentions, no search happened across 24 calls.
• On local intent, 94% of queries name a specific venue. And when you remove the search tool, the model already names those same venues.

So a keyword list extracted from an AI engine describes one execution and the model's prior beliefs. Not your customer's intent. The full paper, raw data and code are at the bottom of this article.

One intention, two phrasings, no shared words

On 1 October 2026, two searches run seconds apart:

Not one word in common beyond the "pizz-" root. Two different geographic resolutions. The same address on top.

A string-matching engine, the 2010 kind, would have treated these two lines as two separate markets: two search volumes, two difficulty scores, two pages to write. An engine that encodes the query and the index into the same vector space treats them as two neighbouring points. The unit of work is no longer the string, it is the intention.

That is where keyword thinking starts to cost money, and there is a figure for it. Semrush matched the prompts observed in its clickstream panel, over a billion rows across 200 million US users, against its own base of 27 billion keywords: over most of the October 2024 to February 2026 window, 65 to 85% of prompts match no existing keyword (Semrush). No volume, no CPC, no difficulty. Nothing to buy.

Incidentally, the same study corrects a widely circulated figure. The "AI prompts are 23 words versus 4 on Google" claim (and its "60 versus 3.4" variant) counts all prompts, including coding, translation, writing and brainstorming, none of which ever trigger a web search. Separate the two populations and the reading flips:

The high average comes from prompts that search for nothing, and it is falling. The ones that matter for AI visibility are 8.7 words long, and they are getting longer. Building a strategy on "prompts are long, think conversational" therefore repeats a reading error.

What the engine actually does: query fan-out

Definition of query fan-out: the mechanism by which a generative engine breaks the user's question into subtopics and issues several web searches of its own, whose results then feed its answer.

Those are not our words, they are Google's, which describes AI Mode as "breaking down your question into subtopics and issuing a multitude of queries", and Deep Search as able to issue "hundreds of searches" (Google).

On the ChatGPT side, Nectiv extracted those machine queries across more than 8,500 commercial prompts in nine verticals: 31% of prompts trigger a search, 2.17 searches per searching prompt, 5.48 words per query, and local searches most (59%) (Nectiv). A year later, on roughly 4,000 replayed prompts: 7.61 queries per prompt, a maximum of 29, and 64% of queries use a site: operator (Nectiv).

Hold on to that last figure, we come back to it: a site: query interrogates a domain already chosen. The engine is not looking for who could answer, it is checking someone it already has in mind.

The keyword has therefore gone from an input you picked to an output of the engine you never see, and which appears in no Search Console. Hence the offering built around it: platforms promising to "turn an AI query into keywords". That promise assumes two things. That these queries are stable for a given intention, and that they exist at all. We checked both.

Our test: 3 intentions, 8 paraphrases, 224 calls

Why Gemini. It is, to our knowledge, the only mainstream engine whose official API returns its own search queries: generateContent with the google_search tool exposes groundingMetadata.webSearchQueries, the exact list of queries issued, and groundingChunks[], the sources kept (Gemini documentation). ChatGPT's queries are visible in the JSON returned by its web interface, but no official API exposes them. Every figure below is therefore measured on Gemini, and we do not generalise it to other engines.

The material. Three commercial intentions in French, chosen to cover three different search surfaces: a CRM for a small business (B2B SaaS), the best Neapolitan pizzeria in Paris open on Sunday evening (B2C local), an electric bike for a 15 km daily commute at around €2,000 (B2C considered purchase).

Each intention has eight paraphrases of 21 to 25 words (mean 21.4) preserving every constraint: direct question, lexical synonyms, reversed constraint order, informal register, imperative form, narrative context first, negatively phrased constraint, and a paraphrase of the head term itself. The last one is reported separately: it is the hardest test, and a sceptic could argue it is not a true paraphrase.

The control group, which is the whole point. Each prompt is replayed 3 times (5 times in a second power run). Comparing two paraphrases means nothing unless you know how much two runs of the same prompt already differ. Without that baseline, you publish sampling noise and present it as an effect. It is the same requirement that led us to take apart a popular AEO case study: a number with nothing to compare it to says nothing.

The hypotheses, with their refutation criteria, fixed before collection:

We had committed to publishing nothing from a hypothesis that failed. Two failed. What follows explains why they failed, and why the conclusion comes out stronger.

Result 1: machine queries are not reproducible

We set out to show that two equivalent phrasings produce different queries. That is true, trivially: the Jaccard index between the queries of two paraphrases is 0.011, essentially no overlap.

Except the control tells a different story.

Rerun the identical prompt: 91 to 97% of the queries change. On gemini-2.5-flash, 62 searching calls produced 237 distinct queries, which is nearly one fresh query per call. With 70 control pairs on the pizzeria intention, the gap falls to 0.049, 95% confidence interval 0.018 to 0.073. On both models the entire interval sits below our 0.10 threshold.

H1-bis is therefore refuted, and that is a better argument than the one we were after. We wanted to say "rephrase, and the keywords change". The reality is harsher: change nothing at all, and the keywords change anyway. The "AI keyword" is not a property of the query, nor of the intention. It is an artefact of the execution.

The direct operational consequence: there is no "right phrasing" from which to extract truer keywords, and a list extracted from one call is not reproducible by the person you present it to. If you buy a fan-out report, you are buying one draw.

At a different level of measurement, paraphrasing does move something: at the word level rather than whole strings, the gap rises to 0.115 on local intent. What those queries share are the prompt's constraints (Sunday evening, opening hours, price, booking), not shared queries.

Result 2: half the time there is no query to extract

On gemini-3.8-flash, 49% of calls issued no search query at all. The breakdown by intention is telling:

On the B2B intention the model never searched. It answered from its weights, roughly 4,000 characters of recommendations, naming products, without consulting a single web page. There is no keyword to extract there, no source to influence in the moment, and no trace to audit. Visibility is settled entirely upstream, in what the model memorised during training.

The rate depends on the model: on gemini-2.5-flash only 1.6% of calls skip search. So "half of prompts never search" cannot be stated as a general law. "On the model users have today, one call in two" can. And this specific point converges with what Nectiv observes on ChatGPT, where 31% of commercial prompts trigger a search, and where local is the vertical that searches most.

This is the first conceptual flaw in fan-out offerings: they measure the retrieval path, and on a substantial share of prompts that path does not exist.

Result 3: on local intent, the fan-out verifies a shortlist, it discovers nothing

Here are queries actually issued on the pizzeria intention:

peppe pizzeria paris dimanche soir reservation
popine paris dimanche soir reservation
guillaume grasso pizzeria paris dimanche soir
agata pizzeria paris reservation dimanche soir
"da michele" bastille paris pizza prix dimanche

94% of that intention's queries (66 of 70) name a specific restaurant or a ranking. Strip the name and the remaining words are the prompt's own constraints. On the bike intention, 57% of queries name a brand or model.

These queries are not looking for "where to eat Neapolitan pizza in Paris". They are checking whether Peppe is open on Sunday evening. Which leaves the question of where the names come from.

We replayed the same prompt with no search tool at all, with the model's thought summaries enabled. With no web access, the model already names twelve venues: Peppe, Popine, Guillaume Grasso, Dalmata, La Manifattura, Agata, Iovine's, Popolare, Bricktop, Da Michele and others. Turn search back on: all five venues queried belong to that list, and across the full 72-call run, 9 of the 12 venues queried are in it. One thought summary, during a call with search enabled, reads: "First, I need to consider my top candidates."

So the fan-out formula on local intent is:

machine queries = [the shortlist the model already has] × [the prompt's constraints]

And here the argument becomes structural rather than statistical. Even if machine queries were perfectly stable, extracting them would teach you nothing about demand. You would be reading the candidate list the model already held, filtered by the user's constraints. A ranking of recurring n-grams in those queries, "reviews", the current year, "comparison", describes the model's behaviour. It is not an inventory of intents.

The 64% of site: queries measured on ChatGPT say exactly the same thing by another route: interrogating a named domain presupposes having already chosen it.

Which moves the working question. It is no longer "which keywords to target", it is "how do we get into the shortlist", meaning: exist in what the model memorised, and in the sources it consults to check itself.

What we set out to prove, and what turned out to be false

Our starting hypothesis was elegant: queries diverge but cited sources converge, so the intention is the stable object and the lexical path is just noise. It is false, and how it failed is worth telling.

The naive test seemed to confirm it. Cited domains do overlap far more than queries: around 0.23 to 0.34 against 0.01. But that test discriminates nothing. We ran it on simulated data where every paraphrase draws its queries at random from one shared pool, so with no effect to detect at all: domains come out at 0.47 against 0.05 for queries, and the test declares "confirmed". The reason is mechanical: there are far fewer possible domains than possible query strings, so domains overlap more even under pure noise.

So we replaced the test, before running the study, with a difference-in-differences: the paraphrase effect on queries minus its effect on domains, each measured against the noise of the same prompt replayed. Verdict: −0.018 on one model, −0.068 on the power run, 0.010 on the other model. Paraphrasing moves sources at least as much as queries. H2 is refuted.

Worse for the elegant hypothesis: on local intent, paraphrasing moves the shortlist itself. The overlap of restaurants named per call goes from 0.555 between two runs of the same prompt to 0.395 between paraphrases. The sources follow. This is not "stable answer, unstable path", it is the reverse: query strings are noise, and where phrasing acts, it acts on the answer.

We will therefore never write "sources converge while queries diverge". Our own data refute it. And that nuance has an uncomfortable commercial consequence: on local intent, how your customer phrases the question changes the list you need to be on.

What remains true of SEO

None of the above says SEO is useless. GEO (Generative Engine Optimization) is not a replacement, it is an extension, and two floors remain entirely technical.

The retrieval floor. No crawl, no citation. When the engine checks its shortlist it reads pages, and yours need to be reachable and readable by AI engine crawlers, which do not have the same capabilities as Googlebot (crawler comparison).

The extraction floor. The queries we observed are constraint checks: opening hours, price, availability, booking, compatibility. If the answer to a constraint sits in an image, behind a JavaScript tab, or buried mid-paragraph, the engine does not find it and the candidate drops off the list.

What changes is the object of the work. Not a list of strings to target, but a presence to build and an absence to measure.

Action plan

1. Stop buying AI keyword lists. If a report presents fan-out queries as keywords to target, ask for the control: the same measurement replayed twice on the same prompt. Without that number, the report describes one draw.
2. Work in intentions, not strings. An intention is a set of constraints (who, where, when, how much, instead of what). List your customers' constraints, not their phrasings.
3. Sample instead of measuring once. Since two runs of the same prompt diverge, a one-off measurement of your visibility is unreliable: GEO visibility is a distribution, not a score.
4. Measure per model, not on average. Our two models do not behave alike: 49% of calls without search versus 1.6%. A visibility figure aggregated across engines hides the essential.
5. Treat entering the shortlist as the objective. Which means: mentions on third-party sources the engine consults, consistent product naming everywhere, presence in your category's rankings and comparisons.
6. Make every constraint checkable in one line. Opening hours, price, coverage, compatibility, lead times, in text, on a crawlable page. That is what the engine comes to verify.
7. Watch the intentions that never search. On your B2B topics, test whether the engine consults the web at all. If not, your lever is the model's memory, which builds over months.

Traaker measures that presence per engine, per intention and over time, with the sources actually cited. See how the platform works.

Frequently asked questions

What is query fan-out? It is the mechanism by which a generative engine breaks the user's question into subtopics and issues several search queries of its own before answering. Google describes it for AI Mode, and Nectiv measures 7.61 per prompt on average on ChatGPT, with a maximum of 29.

Can you extract the keywords an AI uses and work them like SEO keywords? No, for three measured reasons. Those queries are not reproducible: two runs of the same prompt share only 3 to 9% of them. A substantial share of prompts issues no query at all (49% on the model we tested). And on local intent, 94% of queries name a candidate the model already had in mind, so they describe its beliefs rather than demand.

Does that mean keywords are dead? The keyword as a buyable targeting unit is dead: 65 to 85% of prompts match no keyword in Semrush's 27-billion base. Vocabulary itself stays useful for a different reason: the engine verifies constraints, and those constraints must be written in plain text on your pages.

Why do "best pizza Paris" and "number 1 pizzeria in the capital of France" return the same answer? Because matching no longer happens between strings but between vectors. The engine encodes the query and the indexed content into the same semantic space and compares positions, not words. Two phrasings of one intention therefore land in the same area. One caveat though: this does not make phrasings perfectly interchangeable. Our data show that on local intent, paraphrasing moves the list of candidates kept (overlap 0.555 versus 0.395).

Do you still need technical SEO to be cited by AI? Yes. No crawl means no retrieval, and therefore no citation. The queries we observed are checks on precise constraints: if the answer is not machine-readable, the candidate drops off the list. Technical SEO has become a floor, no longer a differentiator.

Do these results hold for ChatGPT and Perplexity? They are measured on Gemini, the only engine whose official API publishes its own queries. Only two points converge with published ChatGPT data: the frequent absence of search, and local being the vertical that searches most. Existing ChatGPT data measure neither reproducibility nor paraphrase effects, so we do not transpose the rest.

What should you track instead of a keyword ranking? Presence: on what share of a sample of intentions are you cited, by which engine, with which sources, and where are you absent. That is a distribution to sample over time, not a position to check.

Paper, data and code

• Full paper (7 pages: method, verdicts, limitations): Paraphrase and Fan-Out: What Gemini's Queries Reveal (and Don't Reveal) About Intent
• Raw data and scripts: ZIP archive containing the 224 call records (prompt, variant, run, queries, domains, answer, model, timestamp), the no-tool probe, the protocol, the collection script and the analysis script.

Replaying the study takes a Gemini API key, Node, and about four minutes per 72-call run. Verdicts may differ from ours: engine behaviour changes over time, which is exactly why we publish the files.

Method note and limitations. This is a reproducible demonstration of mechanism, not a population estimate: 3 intentions, 24 French prompts, 224 calls, 2 models, one geography and one point in time (1 October 2026). One engine, Gemini, because it is the only one whose official API publishes its queries; no mechanical transposition to ChatGPT or Perplexity. webSearchQueries is the engine's own declaration about itself, not a network capture, and the API with the google_search tool is not the consumer Gemini application. Thought summaries are written by the model, and the API does not timestamp them against searches: the order between candidate selection and search is therefore inferred, not observed. The semantic equivalence of the paraphrases is a human judgement, made auditable by an imposed constraint set and a named transformation grid, not proven. The entity lexicon used to classify queries was built by hand from the raw data. The CRM intention provides no data for H1 and H2 on the primary model since it never searched, and the bike intention has few control pairs. Intervals are 95% confidence intervals obtained by bootstrap over variants. Descriptive measurement, no causality claimed.

Today we are launching Targets to fix that. You declare exactly what you want detected in AI engine answers, and Traaker tells you which of those things was cited, in which answer, and in which source.

The short answer

A target is a verifiable identifier of your business that Traaker looks for word for word in AI engine answers: a brand name and its aliases, a product page URL, a store address, an editorial domain, a spokesperson's name. You can declare up to 500,000 signals per brand, across six target families, and organise them with up to 20 tags per target.

Detection is deterministic: text matching rules, not a language model passing judgement. Two direct consequences. The same content yields the same verdict from one run to the next, so a change in your score is a real market movement and not model noise. And the time to analyse an answer does not depend on how many targets you have: looking for 500,000 signals costs the same as looking for one.

The problem: "cited" means nothing

Every GEO measurement tool asks the engine the same question, then looks for your brand name in the answer. It is the easiest measurement to build, and the one that teaches you the least.

Three blind spots, and they are expensive:

• The brand name is the wrong object. You don't sell a name, you sell 310 products, 40 stores, a brand magazine and three experts who appear on television. Knowing your name shows up doesn't tell you whether your product page was recommended or a comparison site talking about you.
• Homonyms silently inflate scores. A brand called "Square", "Atlas" or "Pause" is cited everywhere, all the time, with no connection to it whatsoever. A tool that looks for a substring in a text cannot tell the difference.
• The geographic false positive is the worst of the three. An answer saying "the DESSANGE salon is on rue du Marché aux Herbes, and Toni&Guy just opened on the same street" contains your brand and your street. It also contains a competitor. Counting that answer as a citation of your establishment is a measurement error that makes you believe you are winning a territory you are losing.

Classic SEO made do with a position in a results page. An answer engine returns no position: it returns a sentence, with an object inside it. So the measurement has to be about the object.

What a target actually is

Definition of a target (Traaker): an entity you declare once, with the exact values under which it can appear in an AI answer. Traaker compiles each value into signals, the normalised forms that are genuinely searchable (accents folded, apostrophes canonicalised, URLs reduced to a comparable form, phone numbers in E.164). A signal is the unit the measurement consumes, and it is what your quota counts.

One target produces several signals. An establishment named "Jacy'z Hotel & Resort" generates a handful on its own, because ChatGPT will write the curly apostrophe where your back office typed the straight one, and because the "&" disappears from some restatements. That compilation work is what separates a declared target from a genuinely detectable one.

The six target families

Products and Categories sit together under a "Catalogue" tab, which gives five tabs on screen for six families.

The 500,000 signal cap is the default on every plan. It is per brand, not per account: an agency running fifteen brands gets fifteen times that budget, with no chance of one high-volume brand starving the others.

Persons: 50 titles, 13 languages

The Persons family deserves a note, because nobody else measures it. A hairdresser, a lawyer, a surgeon or a chef is cited by title as much as by name, and the title changes language with the engine. Traaker ships 50 titles covering courtesy, professional, military, religious, noble and political forms, generated across 13 languages (French, English, German, Spanish, Italian, Portuguese, Dutch, Arabic, Persian, Hebrew, Japanese, Chinese, Malay). You pick the title, Traaker proposes the name forms, you tick the ones that apply.

A French doctor cited in English in a Perplexity answer gets detected, because the English form was generated the moment you created the target.

How to fill 500,000 targets without typing them

A 500,000 signal cap only matters if filling it isn't a project. Three routes, fastest first:

1. From your sitemap. You give your domain, Traaker finds the sitemap (seven detection tiers, from robots.txt to llms.txt) and harvests up to 50,000 URLs. It splits product pages from category pages on its own, then shows the breakdown in an "Analyse" step before you commit to anything. You see "310 products, 1,402 categories" and then you decide.

2. By CSV. A downloadable template per family, up to 5,000 rows per batch. A second import merges with what exists instead of creating duplicates: you can replay the same enriched file without cleaning up behind yourself, and the coordinates of already-geocoded establishments are preserved.

3. By hand, in a dialog per family. That is the normal route for your brands, your establishments and your spokespeople, which number in the dozens rather than the thousands.

On performance, consolidating a partition runs at roughly 2.5 seconds per 100,000 targets on our infrastructure. Past 100,000 targets, the write is queued and hands control straight back to you instead of spinning.

Deterministic detection, not an AI guessing

This is the most important architectural choice, and it is what explains the two properties stated above.

Traaker does not ask a language model "is this brand cited?". It folds the answer (accents, case, apostrophes), splits it into word sequences, and compares those sequences against your set of signals. The result is a yes or a no, reproducible, explainable, and whose cost is proportional to the length of the answer rather than to the number of your targets.

The co-occurrence rule, and why it exists

An establishment poses a problem the other families don't: its name alone does not identify it. Forty salons carry the same retail name, and a street is shared with whichever competitors sit on it.

So Traaker requires two cumulative conditions before crediting an establishment: the brand and the street must appear within 80 characters of each other, and in the same sentence. The window alone was not enough: "DESSANGE is a well-known chain. Toni&Guy just opened on rue du Marché aux Herbes" puts the two halves 47 characters apart, crosses a full stop, and is about a competitor. A pair means "this street next to this brand," and two sentences are not that, however close they sit.

The consequence is visible on screen: every establishment shows its precision level and what actually identifies it. A city shared by two of your salons is not kept as a key, because it does not let anyone tell which one the AI meant.

A single search fills in the name, address, phone, website and Place ID from the Google listing. You retype nothing.

Content: seeing who picks you up as a source

The Content family answers a different question from the others: when an engine cites a source about you, which one is it?

You declare your domains and editorial sections, and Traaker groups them by root domain rather than by subdomain. blog.yourbrand.com, pro.yourbrand.com and yourbrand.com all land in the same group, because three rows for a single outlet makes the table unreadable without teaching you anything.

Each piece of content is attached to a brand, which lets its citation be coloured in that brand's colour inside the answer. On an account tracking several retail names, you tell your own outlet from a competitor's at a glance.

Tagging: slicing the measurement the way you are organised

A target accepts up to 20 tags, 40 characters each, in any family. The vocabulary is free-form: there is no taxonomy to maintain before you start.

What tags are for:

• Filtering your targets server-side, across the whole base and not just the page on screen.
• Slicing a measurement by range, by region, by network (company-owned versus franchised), by campaign.
• Renaming without breaking anything. A renamed tag applies retroactively to every target carrying it, with no reindex and no delay. Tags are never searched for in answers: they are analysis labels, not signals.

A tag is therefore the analysis axis you add after the fact, when the question comes up, without having to redo your setup.

What this changes in your reports

Targets are not one more configuration screen, they change how you read your measurements.

Inside engine answers, every detected target is highlighted in its brand's colour, in the prose as well as in the list of cited sources. Each engine's panel shows the families it recognised. You no longer read a percentage, you read the sentence that produced it, with the exact word that triggered the detection.

That is also what makes the result contestable, in the good sense: when a number surprises you, you can go back to the citation and check. A score without that traceability is an act of faith.

Who it's for

• Chains and multi-location retailers: the Establishments family with its co-occurrence rule is the only honest way to measure local AI visibility. Knowing "your brand" was cited in an answer about Brussels does not tell you whether it was your Avenue Louise salon or the Marché aux Herbes one.
• E-commerce and marketplaces: 310 products or 100,000, the cost of measurement doesn't move. Targets complement the E-commerce module by pushing detection down to each product page's exact URL.
• Media brands and publishers: the Content family measures your content as a cited source, which is the conversion that matters when your model is editorial.
• Practices, clinics, professional services: the Persons family measures how well known your practitioners are, title and language included. It is often the person who gets recommended, not the organisation.
• Agencies: the signal quota is per brand, so one large client does not consume everyone else's budget.

Action plan: your Targets in five steps

1. Declare your brand: its name, the aliases genuinely in use (a former name, an unaccented spelling, a local abbreviation) and its domain. With no "Brands" target, nothing gets highlighted in answers.
2. Import your catalogue from your sitemap. Look at the breakdown in the "Analyse" step, correct the product / category split if needed, then commit.
3. Add your establishments with a Google search, then check each one's precision banner. Fill in the street wherever you know it: without it, two stores in the same city cannot be told apart.
4. Add three tags at most to start with. A region, a network, a range. You will add more when a specific question comes up.
5. Run a visibility measurement and read the highlighted answers before you read the scores. That is where you will see whether your targets are precise enough.

Budget around ten minutes for steps 1, 3 and 4, and one minute for step 2 if you have a sitemap.

Frequently asked questions

What is the difference between a target and a signal? A target is the entity you declare (a store, a product, a person). A signal is one of the textual forms under which that entity can appear in an answer. A target always produces several signals, and it is the signal count that your 500,000 quota measures.

Is the 500,000 cap per account or per brand? Per brand. It is the only quota in Traaker that works this way. An agency tracking fifteen brands gets fifteen independent budgets, and a brand with a large catalogue does not penalise the others.

How many establishments can I declare? 100 by default, and that limit is raised on request for chains with more. Unlike signals, this cap is account-wide.

Does an AI decide whether I am cited? No. Detection rests on text matching rules applied to the normalised answer. That is what guarantees the same content always yields the same verdict, and that the measurement does not drift when a model provider ships an update.

What if my brand has a generic name? That is precisely the case Targets handle. A value that is too short or too generic is flagged as you type, and for establishments detection requires co-occurrence with the street rather than the name alone. You can also lean on your domains and URLs, which are never ambiguous.

Do I have to retype everything if I already have a catalogue in Traaker? No. Sitemap import and CSV import merge with your existing targets instead of creating duplicates, and the coordinates of already-geocoded establishments are preserved from one import to the next.

Are tags searched for in engine answers? Never. A tag is an analysis label used to filter and segment your reports. That is exactly why a tag can be renamed at any time, retroactively, with no reprocessing.

How long does analysing one answer take? A few tens of milliseconds, and that time does not depend on how many targets you have declared. A brand with one target and a brand with 500,000 signals pay the same analysis cost, because the work is proportional to the length of the answer.

Available today

Targets are live now on every plan, in the GEO Strategy page of your workspace. If you want to see what it looks like on your own brand before diving in, book a demo: we import your catalogue and your establishments live during the call.

To go further on measurement method, read why GEO visibility is a distribution and not a score and how to diagnose content that never gets cited. Full platform feature details are on the Platform page.

Method and limitations. The figures in this article describe the feature as shipped: 500,000 signals per brand and 100 establishments per account are defaults, raisable case by case; 50,000 URLs is the harvest cap of one sitemap import; 5,000 rows is one CSV batch; the 2.5 seconds per 100,000 targets and the few tens of milliseconds per answer are measured on our infrastructure and will vary with the length of the answers analysed. Deterministic detection guarantees the reproducibility of the verdict, not its exhaustiveness: a rephrasing that none of your signals covers remains a miss, which is exactly why the alias list stays editable at any time. The establishment co-occurrence rule is deliberately strict, which makes it cautious rather than generous: it would rather not credit an ambiguous mention.

The four major AI engines do not share a crawling architecture, and that directly changes what is visible on your site. OpenAI runs four separate bots (including GPTBot for training and OAI-SearchBot for search) and is gradually building its own index while still leaning on Bing ([1]). Anthropic takes the lightest approach: Claude outsources retrieval entirely to the Brave Search API ([2]). Perplexity has built the most independent infrastructure, an index of over 200 billion URLs ([3]), but also faces the most serious suspicion of bypassing robots.txt ([4]). Google has no separate AI crawler at all: Googlebot does everything, with a dedicated token, Google-Extended, to control training use ([5]). And the most consequential technical fact: only Googlebot executes JavaScript. GPTBot, ClaudeBot, and PerplexityBot do not ([6]).

OpenAI: four bots, and a homegrown index still under construction

OpenAI runs the most complex ecosystem, with four separate bots, each independently controllable via robots.txt ([7]):

Notably, since the GPT-5 launch in August 2025, OAI-SearchBot now generates roughly 3.5x more log events than GPTBot ([8]), a sign that search now outweighs training collection in OpenAI's crawl activity. GPTBot alone generated approximately 569 million requests across Vercel's network in December 2024 ([9]).

ChatGPT Search remains a hybrid architecture: queries from Enterprise and Edu workspaces still route through Bing ([10]), but OpenAI is building its own index in parallel, with separate verticals for news, shopping, PDFs, and academic papers, identified by Peec AI through experiment names such as prefer-index-over-serp-v3 ([1]). The retrieval stack reportedly runs on three layers: a discovery index, a reading cache that keeps copies of previously fetched pages, and a small set of pages opened live at query time ([11]).

Anthropic: the lightest approach, the most dependent on Brave

Anthropic runs three bots: ClaudeBot for training, Claude-User for real-time retrieval, and Claude-SearchBot for search index optimization, each independently controllable via robots.txt ([12]).

Claude's defining trait is that it has no search index of its own. When web search is enabled, Claude relies entirely on the Brave Search API, confirmed in March 2025 when Anthropic added "Brave Search" to its subprocessor list ([2]). Claude's web fetch tool then retrieves full page content on demand, with HTML preprocessing that strips scripts and styles ([13]).

It is also the engine with the strongest citation precision, thanks to its Citations API, which chunks source documents into sentences before passing them to the model, enabling exact-sentence-level attribution ([14]). Anthropic reports this feature improves citation recall by up to 15% compared to a hand-built, prompt-based citation implementation ([14]).

ClaudeBot also had its own notoriety moment: in July 2024, iFixit reported roughly one million requests in 24 hours, a crawl rate fast enough to trigger internal alerts ([15]).

Perplexity: the most independent infrastructure, the most contested

Perplexity has invested most aggressively in proprietary search infrastructure: an index of over 200 billion unique URLs, described as an "exabyte-scale index," backed by tens of thousands of CPUs and hundreds of terabytes of RAM ([3]). The company handles approximately 200 million queries per day ([3]) and was already serving roughly 30 million cited answers per day by May 2025 ([16]).

Two bots are declared: PerplexityBot for indexing and Perplexity-User for user-initiated fetches ([17]). The problem is what is not declared. Cloudflare documented "stealth crawling" behavior: user agents modified mid-crawl, and a switch to unannounced IP ranges, which led Cloudflare to de-list Perplexity as a verified bot ([4]). Multiple publishers have taken the matter to court, including the New York Times, Reddit, Dow Jones, and the BBC ([18]).

Google: one bot for everything, one token for AI training

Google takes the opposite approach from the other three: no separate AI crawler at all. Googlebot fetches everything, and Google controls training use through a dedicated robots.txt token, Google-Extended, introduced in September 2023 ([5]). Blocking Google-Extended prevents pages from being used for model training, but does not affect regular search indexing.

A newer fetcher, Google-Agent, handles real-time, user-triggered requests inside Gemini products, and behaves more like a standard web browser than a search crawler ([19]). Gemini relies on a six-stage grounding pipeline that rewrites the query, reranks results, and blends them into the model's context before generating a response ([20]).

The JavaScript gap: the technical fact that decides your visibility

This is the most consequential limitation across the entire ecosystem. GPTBot, ClaudeBot, and PerplexityBot do not execute JavaScript, a finding established by Vercel's analysis of over half a billion AI crawler requests ([6]). ChatGPT-User and ClaudeBot sometimes fetch JavaScript files (11.5% and 23.8% of requests respectively) but never execute them ([6]). Googlebot is the sole exception: it renders JavaScript through a headless Chrome-based pipeline, making it the only AI-adjacent crawler that can see fully rendered single-page application (SPA, an app whose content is built in the browser rather than delivered as complete HTML) content ([21]).

In practice, this means:

• Any content injected client-side, including JSON-LD structured data added through Google Tag Manager, is invisible to GPTBot, ClaudeBot, and PerplexityBot ([22]).
• On an SPA-based site, these crawlers only see the initial HTML shell: navigation skeletons, meta tags, empty placeholder containers ([6]).
• Content loaded via AJAX, infinite scroll, or anything dependent on client-side rendering stays inaccessible ([22]).
• Paywalled content is equally invisible: AI crawlers do not authenticate, they only see whatever the unauthenticated HTML response returns ([6]).

Server-side rendering (SSR, generating the full HTML on the server before sending it to the browser, rather than in the browser) or static HTML are therefore essential for AI engine visibility ([23]).

Who blocks what: blocking rates by crawler

Nearly 4 in 10 top websites now actively block at least one AI crawler, according to Zyte's State of Web Access 2026 analysis of 11,100 top landing pages ([24]):

(Source: [25])

The ranking is not random. OAI-SearchBot is the least blocked, likely because it sends referral traffic back to publishers, while GPTBot, purely dedicated to training with no traffic in return, is the most blocked ([24]). Among news publishers specifically, blocking rates run far higher: 79% block AI training bots and 71% also block AI retrieval bots ([26]). Cloudflare reports that over 2.5 million websites have chosen to completely disallow AI training through its managed robots.txt feature ([27]).

Volume gaps are just as stark. According to Vercel's network data from late 2024, Googlebot generated 4.5 billion requests per month, against 569 million for GPTBot, 370 million for ClaudeBot, and just 24.4 million for PerplexityBot ([9]). By early 2026, GPTBot and ClaudeBot accounted for 12% and 9.2% of global bot traffic respectively according to Cloudflare Radar data ([28]), and GPTBot's request volume grew 305% between May 2024 and May 2025 ([29]).

A side-by-side comparison of the four engines

(Synthesis of the qualitative positioning found across the studies cited above, notably [6] and [4].)

Gemini comes out ahead thanks to Google's index and JavaScript rendering, one of the two advantages no competitor shares. Claude offsets having no index of its own with the strongest citation precision, driven by its sentence-level Citations API. Perplexity matches the others on index scale but pays for it with the weakest robots.txt compliance.

The citation reliability problem, shared by all four

Despite these different architectures, no engine has solved the citation reliability problem. A 2025 study by the Tow Center for Digital Journalism at Columbia University, run across 1,600 test queries on eight generative search engines, found that these engines fail to retrieve correct information more than 60% of the time ([30]). A separate 2025 arXiv study found only 26.5% of bibliographic references were fully correct, versus 39.8% erroneous or fabricated ([31]). And source overlap between AI engines citing the same topic can run as low as 16% ([32]): for the exact same question, two engines can rely on evidence with almost nothing in common.

Over the same period, AI search visit volume jumped from 15.6 to 27.4 billion, a 42.8% rise between Q1 2025 and Q1 2026 ([33]). Audience is growing faster than answer reliability is improving.

Timeline: the dates that matter

• September 2023: Google introduces the Google-Extended robots.txt token to separate training from indexing ([5]).
• June 2024: WIRED and developer Robb Knight independently document Perplexity ignoring robots.txt and using spoofed user agents.
• July 2024: iFixit reports roughly one million ClaudeBot requests in 24 hours ([15]).
• March 2025: Anthropic officially confirms it relies on the Brave Search API for Claude's web search ([2]).
• Summer 2025: Cloudflare documents Perplexity's stealth crawling and de-lists it as a verified bot ([4]).
• August 2025: GPT-5 launches; OpenAI's crawl activity triples in the following weeks ([8]).
• October to December 2025: Reddit, then the New York Times, sue Perplexity for copyright infringement ([18]).
• March 2026: Google rolls out Google-Agent, a fetcher separate from Googlebot for real-time Gemini use cases ([19]).

FAQ

Can ChatGPT, Claude, and Perplexity execute JavaScript? No. GPTBot, ClaudeBot, and PerplexityBot do not render JavaScript; they only see the raw HTML returned by the server. Only Googlebot, through its headless Chrome rendering pipeline, executes JavaScript ([6], [21]).

Why doesn't Claude have its own search index? Anthropic chose not to build its own search infrastructure and instead delegates result retrieval entirely to the Brave Search API, a choice publicly confirmed in March 2025 ([2]).

Does robots.txt actually block every AI crawler? Not uniformly. OpenAI, Anthropic, and Google respect robots.txt directives per bot token. Perplexity, however, has been documented bypassing those directives using spoofed user agents and undeclared IP ranges ([4]).

What is Google-Extended, and how does it differ from Googlebot? Google-Extended is a separate robots.txt token that controls only whether pages can be used to train Google's AI models. Blocking it does not affect regular search indexing, which is still handled by Googlebot ([5]).

Why are these engines' citations often wrong? Studies show retrieval failure rates above 60% on test queries, with links sometimes fabricated and source overlap between engines as low as 16% for the exact same question ([30], [32]).

Does Perplexity respect robots.txt? The evidence documented by Cloudflare and independent developers indicates it does not, at least not consistently: user agents modified mid-crawl, undeclared IP ranges, which led Cloudflare to de-list it from its verified bots list ([4]).

Which crawler generates the most traffic on a given site? It varies by industry, but at the global scale, Googlebot dominates by far (4.5 billion monthly requests per Vercel), followed by GPTBot then ClaudeBot, with PerplexityBot remaining the smallest volume among the four main AI crawlers ([9]).

Your action plan

1. Serve server-rendered or static HTML for any critical content. GPTBot, ClaudeBot, and PerplexityBot will never see what only exists in a client-side generated DOM.
2. Never inject your structured JSON-LD data through a client-side tag manager. It needs to exist in the raw HTML the server returns.
3. Check your robots.txt token by token, not just User-agent: *. GPTBot, ClaudeBot, PerplexityBot, Google-Extended, OAI-SearchBot, and Claude-SearchBot are each controlled independently.
4. Separate training bots from search bots in your server logs. Blocking a training bot (GPTBot, ClaudeBot) does not have the same visibility impact as blocking a search bot (OAI-SearchBot, Claude-SearchBot).
5. Don't expect paywalled or authenticated content to be seen. No AI crawler authenticates; only freely accessible public content counts.
6. Track what actually gets cited, not just what gets crawled. A technically perfect, fully crawlable site can still be absent from answers if the content lacks the evidence density these engines look for.

If you want to know precisely which engines actually see your content and which ones cite you, check out Traaker's method for measuring and improving your visibility in AI search.

Keep reading

These crawling architectures are only the first step. Once an engine can see your content, it still has to choose to cite it, which depends on factors specific to each engine: read our guides on how to get cited by ChatGPT, how to get cited by Claude, and how to get cited by Perplexity.

Methodology and limitations. This article synthesizes public data from third-party sources (Vercel, Zyte, Cloudflare Radar, Perplexity Research, Anthropic, OpenAI) collected at different dates and with different methodologies; the percentages and volumes should be read as orders of magnitude rather than strictly comparable measurements. AI crawler behavior evolves quickly, and some sources do not fully agree on the exact timeline of certain events (notably when Cloudflare first detected Perplexity's stealth crawling). Always verify the current behavior of your own robots.txt and server logs before making a blocking decision.

Sources

1. SearchEngineWatch
2. TechCrunch
3. Perplexity Research
4. Cloudflare
5. ipregistry.co
6. Adobe Experience League
7. developers.openai.com
8. Search Engine Journal
9. Vercel
10. OpenAI Help Center
11. Search Engine Land
12. anthropic.com/crawl
13. Anthropic Docs
14. Anthropic
15. 404 Media
16. The AI Engineer
17. Perplexity Docs
18. The Verge
19. MarkTechPost
20. WebSearchAPI
21. NuxtSEO
22. Search Engine Journal
23. NuxtSEO
24. Zyte
25. Zyte, 2026 analysis of 11,100 pages
26. BuzzStream
27. PCMag
28. Oncrawl
29. SEOptimer
30. CJR
31. arXiv
32. Frase
33. Contently

This isn't a forecast anymore: Instant Checkout launched on August 26, 2026 for US users on Free, Plus and Pro (OpenAI). With 700 million weekly users, ChatGPT just became both an acquisition channel and a transaction channel. Here's how to get in, step by step.

The short answer

Making your products buyable in ChatGPT takes five steps: make your product pages machine-readable, enable Instant Checkout via Stripe and the Agentic Commerce Protocol, check what the protocol covers (and what it doesn't yet), decide whether to layer on advertising, and above all measure whether AI engines recommend you at all before you worry about payments.

The thing most brands miss: in ChatGPT, product results are organic and non-sponsored, ranked purely on relevance to the user (OpenAI). You cannot buy your way into a recommendation. Payment solves the last step of the journey; getting into the shortlist is won upstream, and that's exactly what GEO covers.

What is commerce inside ChatGPT? Two layers you shouldn't conflate

OpenAI's conversational commerce rests on two distinct systems, and confusing them means investing in the wrong place.

Definition of the Agentic Commerce Protocol (ACP): a protocol co-developed by OpenAI and Stripe that passes order information to the merchant's backend from inside the conversation. The merchant remains the merchant of record: payments, shipping, returns and support all stay in their existing systems (OpenAI).

Step 1: make your product pages readable by an AI

This is the step nobody wants to do, and the one that decides everything. Instant Checkout is worthless if your product never shows up in the tiles.

ChatGPT's shopping experience presents products side by side with prices, reviews and key specs in comparison tables (OpenAI Help). To make it into that comparison, your page has to supply extractable facts, not marketing prose.

What to do concretely:

• Quantify everything quantifiable. Dimensions, weight, materials, compatibility, warranty length. A numeric spec makes it into a comparison table; "built to last" doesn't.
• Structure your product data (Product, Offer, AggregateRating, availability, current price). When a model compares five products, it favours the one whose attributes are unambiguous.
• Write for the real query, not the keyword. ChatGPT handles complex, contextual queries and supports iterative refinement: budget, preferences, constraints (OpenAI). Your page needs the vocabulary of the constraint ("for wide feet," "machine washable," "USB-C compatible"), not just the product name.
• Cover the comparison yourself. A "this model vs. that one" table on your own page hands the model exactly the material it's trying to produce.
• Don't neglect images. Photo-based search has been part of the experience since March 2026, extended to the Free, Go, Plus and Pro tiers (the-decoder).

For the full breakdown of what makes a product page get picked or ignored, see our GEO field guide for products.

Step 2: enable Instant Checkout (Stripe + ACP)

This is by far the easiest step.

1. Already on Stripe: enabling agentic payments can take as little as one line of code (OpenAI).
2. On Shopify: over one million Shopify merchants are rolling out, including Glossier, SKIMS, Spanx and Vuori. Etsy is already live.
3. On another processor: participate through Stripe's Shared Payment Token API or the Delegated Payments Spec.

On security, payments run through encrypted tokens, authorized only for specific amounts and specific merchants, with the user's permission. You never receive raw card data and you don't switch PSP.

What the buyer-side journey looks like: they type "Buy," then confirm order, shipping and payment, all without leaving the conversation.

Step 3: check the current limits before you promise anything to your leadership

Three constraints to know before you wire up a business case.

In other words: the technical integration is ready well before the volume is. Which is precisely the argument for doing it now: the entry cost is one line of code, and the attention inventory is built upstream, on organic discovery.

Step 4: decide whether to layer on advertising

ChatGPT Ads is a separate layer, and its economics have moved fast.

CPM was more than halved in nine weeks as inventory opened up. A self-service Ads Manager has been available since May 5, 2026 at ads.openai.com: advertiser signup, budgets, bids, creatives, reporting (OpenAI).

A format that has nothing in common with Google's: Adthena's analysis of nearly 850,000 AI search queries shows ChatGPT displays a single ad per response, versus Google's carousel, and blocks regulated sectors (legal, pharma) that Google still serves. One advertiser owns the attention: scarce inventory, therefore premium, and therefore under growing competitive pressure as CPMs fall and buyers arrive.

The trap to avoid: believing that paying improves your position in the recommendations. See the next step.

Step 5: understand "answer independence" (and what it means for your budget)

OpenAI's founding principle is called answer independence: ads do not influence ChatGPT's answers (OpenAI).

Architecturally, ads run on systems separate from the chat model, and advertisers have no ability to shape, rank or alter responses (OpenAI Help). Slots appear below the end of a response, labelled "sponsored," visually separated, with an "About this ad" link. Ads are excluded near sensitive topics (personal health, mental health, politics), are not shown to under-18s, and conversations are never shared with advertisers, who only receive aggregated performance data.

The strategic consequence is sharp:

Media budget buys visibility next to the answer. Only GEO work buys a place inside it. Those are two different budget lines, and the second one has no paid shortcut.

That's why Step 1 isn't optional. On the organic layer, the one that produces product tiles and recommendations, the levers are GEO levers: factual density, structured comparisons, third-party authority. See our dedicated guide: how to get cited by ChatGPT.

Why it's worth the effort: the number that should convince your leadership

Shoppers landing on retail sites via AI chatbots were 40% more likely to buy than those coming from search, email or social, according to Adobe's Prime Day 2026 data: the first time AI-referred traffic converted best (GeekWire).

That isn't a measurement fluke, it's a property of the channel. Unlike Google Search, which depends on commercially valuable keywords typed by the user, ChatGPT can infer commercial intent from extended conversational exchanges, even when the user never explicitly searches for a product (Business Insider). By the time a shopper clicks, they've already told the assistant their budget, constraints and use case. They arrive qualified.

One more market signal: Amazon CEO Andy Jassy publicly named horizontal AI agents (ChatGPT, Gemini, Copilot) as his single biggest competitive threat, not another retailer (Forbes).

The timeline: how fast this is rolling out

The pace of the rollout is the strongest argument against waiting.

• February 9, 2026: US advertising pilot (Free and Go tiers, logged-in adults).
• March 26, 2026: expansion to Canada, Australia, New Zealand, after finding no impact on consumer trust metrics and low ad-rejection rates.
• May 5, 2026: self-service Ads Manager opens.
• August 11, 2026: UK, Mexico, Brazil, Japan, South Korea live.
• August 18, 2026: 31 European countries (Germany, France, Spain, Italy, Sweden, Norway, Denmark, Netherlands, Austria…), bringing the total to 40 countries (OpenAI).
• August 26, 2026: Instant Checkout launches in the US.

From 1 to 40 countries in six months. Instant Checkout will follow the same trajectory: the question isn't whether your market gets covered, but whether your product pages are ready when it does.

Frequently asked questions

Do I need to change payment provider to sell in ChatGPT? No. The merchant remains the merchant of record and keeps its existing payment, shipping, returns and support systems. On Stripe, enabling it takes one line of code; elsewhere, the Shared Payment Token API or the Delegated Payments Spec let you participate.

Can I pay to appear in the product recommendations? No. Product results are organic and non-sponsored, ranked purely on user relevance. Ads are a separate system, shown below the answer and labelled "sponsored." That's the answer-independence principle.

What does Instant Checkout cost? 4% commission on sales made by Shopify merchants through the feature. Nothing to appear in organic discovery.

Is Instant Checkout available outside the US? Not yet. As of August 28, 2026, Instant Checkout is US-only. ChatGPT Ads, by contrast, covers 40 countries including most of Europe since August 18, 2026. Organic product discovery doesn't wait: it works everywhere ChatGPT is used.

Can shoppers buy several items at once? Not yet. The system supports single-item purchases; multi-item carts are announced for later.

Do advertisers see the conversations? No. Conversations are never shared with advertisers, who only receive aggregated, non-identifying performance data. Users can turn off ad personalization and clear the associated data (deleted within 30 days).

The 5-point action plan

1. Audit the machine-readability of your product pages: structured data, numeric attributes, constraint vocabulary, explicit comparisons. This is the one step with no paid shortcut.
2. Enable Instant Checkout via Stripe / ACP now: the entry cost is marginal, and a head start isn't something you can buy back later.
3. Measure your GEO visibility before you buy media. Knowing which purchase queries you're recommended on, absent from, or losing to a competitor drives everything else.
4. Test advertising on a bounded scope, keeping in mind that only one slot shows per response and it doesn't change the recommendation.
5. Re-run the measurement monthly. CPM halved in nine weeks, 40 countries in six months: the state of this channel in August 2026 won't be its state in January 2027.

Measure before you integrate

Wiring up Instant Checkout takes a day. Becoming the product ChatGPT recommends takes months, and starts with knowing where you stand.

That's what Traaker does: query the AI engines (ChatGPT, Gemini, Perplexity, Claude) on your segments' real purchase queries, measure whether you're recommended, score every product page against seven GEO criteria, and generate publish-ready rewrites for the pages too weak to get cited. See the E-commerce module announcement or the dedicated page.

The question is no longer whether conversational commerce will take off, but how fast you'll adapt to it, in a world where AI becomes the first point of contact between the consumer and the product.

Methodology note. The figures and dates in this article come from OpenAI's official announcements and press sources published between February and August 2026 (links inline). Revenue projections are investor communications and should be read as such: eMarketer estimates OpenAI is significantly below its 2026 target, with OpenAI, Microsoft, Google and Amazon combined projected at under one billion dollars in AI ad revenue for the year (Futurism). Engine behaviour changes fast; measure yours continuously.

Until now, Traaker measured that dynamic very well at the level of a brand and its showcase website. But a catalog of ten thousand SKUs isn't a ten-page showcase site, and "are you cited" says nothing about "which of your ten thousand products, on which query, against which competitor." Today we're launching the E-commerce module to answer exactly that.

"I've never seen that from any other platform." Fabrice, Head of SEO/GEO at an e-commerce platform

The problem we hadn't solved yet

AI engines have changed how people shop, and the digital shelf has gotten shorter. Where Google used to hand back a results page the shopper explored themselves, an AI assistant now filters on their behalf and keeps only a handful of products. Being present or absent from that short list has a direct effect on sales, but until now, no tool let an online retailer know:

• whether their products are being recommended by AI engines on their shoppers' real queries,
• which of their ten, hundred, or hundred-thousand product pages are clear enough for an AI to understand and cite,
• and above all, what to fix: not an abstract diagnostic, but a rewrite ready to publish.

Classic SEO answers "can I be found." GEO (Generative Engine Optimization) answers "am I understood, and chosen." For a product catalog, that's a fundamentally different question, at a scale that general-purpose GEO tools don't address.

What the E-commerce module does

Catalog import, no technical integration

Give it your site's URL: Traaker detects the sitemap and automatically imports your products and categories, up to 145,000+ SKUs, or via a simple CSV file. A controllable scope (include/exclude rules by category) and a refresh that detects new and delisted items on every sync.

A GEO score for every product page

Each page is scored on seven criteria that determine whether an AI can understand and recommend it: structured data, factual density, specificity, purchase intent, extractability, evidence & authority, Q&A coverage. Pages rank worst-to-best so you tackle the critical ones first, and every weak page gets a ready-to-publish rewrite (arguments, quantified specs, FAQ), exportable as CSV back into your PIM or CMS.

Shopper segments, not keywords

Describe a customer in plain language: who they are, what they're looking for, how they buy. Traaker derives their real purchase queries, in the languages and on the engines that matter to you. Each query is then matched to the closest product in your catalog via semantic search.

GEO visibility measurement

Traaker queries the AI engines (ChatGPT, Gemini, Perplexity, Claude depending on your plan) on each segment's queries and measures whether you're recommended, whether a competitor edges you out, or whether you're simply absent. You get a visibility score per segment with its trend over time, the queries where you're losing ground to a competitor, and the product families with weak presence.

Content grounded in your real products

On the queries where the analysis finds you absent, Traaker generates the content that puts you back in the running: a buying guide, a comparison, an editorial piece, with your real products placed inside the body of the article. Every link points to the exact URL of your product page, resolved by semantic search inside your own catalog. The model never invents a URL.

Who it's for

• E-commerce & DTC: your catalog is your revenue. Know which queries you're recommended on, and turn your weak pages into citable ones.
• Marketplaces: tens of thousands of SKUs. The GEO score scales and prioritizes the categories where AI visibility matters most for sales.
• Manufacturers & brands: even without selling direct, make sure AI describes your products accurately, and feed the engines pages that hold up.

A full loop, not just tracking

A tool that only tells you you're losing citations doesn't solve anything. The E-commerce module closes the loop: from measurement to the corrected page, all the way to published content that wins back the queries where you were absent. It's the same logic that structures all of Traaker, applied at the scale of an entire catalog instead of ten showcase pages.

The module is available today. Book a demo to see your own catalog imported and your first pages flagged for fixing, or check out the dedicated page for full feature details.

This shift has a name: Generative Engine Optimization, or GEO. It is to AI answers what SEO was to search results. And the good news is that a wave of recent academic research has moved GEO from folklore ("just sound authoritative") to something measurable. Here is what actually works, what backfires, and how it changes by engine and by industry.

First, understand what you are optimizing for

An AI engine does not rank a list. It retrieves a handful of candidate sources, then an LLM re-reads them and synthesizes a recommendation. That means there are three gates, not one:

Most people obsess over ranking and ignore retrieval and absorption. That is a mistake, because in retrieval-augmented setups the retrieval step is a hard gate. If your listing does not embed close to how real buyers phrase their needs, nothing downstream can save you. So the first move is unglamorous: write to the messy, multi-sentence way people actually ask, not to keywords.

The uncomfortable truth about brands

Well-known brands get recommended 100% of the time when nothing else distinguishes the options. Researchers ran a clean experiment: ten products, identical specs, only the brand name differs. The famous brand won every time. AI engines have a built-in "incumbent advantage."

But here is the twist that should give every challenger brand hope: that monopoly is fragile, and it collapses the moment one product has any distinguishing signal. The thresholds are startlingly small:

Across all conditions, product attributes explained about 82% of the ranking, while brand explained barely 1%.

The lesson: the real barrier for a smaller brand is not that it lacks a famous name. It is that it lacks differentiating information. Give the model something concrete to grab onto, and the playing field levels fast.

Evidence beats persuasion

If you take one thing from this article, take this: AI engines reward evidence and largely ignore salesmanship.

When researchers tested classic marketing copy, the persuasion tactics that work on humans fell flat on machines. Anchoring ("premium quality at a third of the price"), scarcity ("only 500 units, selling out"), and loss aversion ("don't let your skin degrade") each moved recommendations by only about 10 to 13%. The model shrugged them off.

What worked instead is that every winner is verifiable:

Notice the pattern. The engine is not looking to be charmed. It is looking to justify a recommendation.

Structure is a lever all by itself

You can lift your citation rate by about 17% without changing a single fact: just by restructuring the content. Headings, chunk sizes, bullet formatting, and emphasis all matter independently of what you say.

Concrete targets from the research:

• Keep paragraphs in the 150-to-300-word range. Longer, and the model's attention degrades in the middle. Shorter, and information fragments.
• Devote roughly a quarter to a third of your content to structured elements (lists, tables). Lists and tables parse far better than prose.
• Put your key claims at the start of sentences and sections, where models pay the most attention.
• Implement Schema.org markup rigorously. Think of your product page as an API that an AI agent has to "do business with." Clean price, spec, availability, and review data makes you easy to recommend and hard to skip.

The same content does not win on every engine

This is where GEO gets interesting. The four major engines behave differently, and a strategy that dominates one can flop on another.

The through-line across all four: they are all biased toward earned media (third-party reviews and authoritative publications) far more than Google is. PR, expert roundups, and review-site coverage are not a side tactic in GEO. They are the main event.

It also depends on your industry

The best writing method shifts by domain. From the research:

A word on the line you should not cross

The research is candid about something important: the single highest-performing tactic in experiments was fabricated authority, made-up clinical trials, invented expert endorsements, planted reviews. It works, for now. But it is being actively detected, defense systems now flag persuasion-heavy and fake-review text with high precision, and it is straightforwardly deceptive.

The durable strategy is the honest version of the same instinct. Cite real certifications. Quote real experts. Publish real specs and real review data. The features that make fabricated authority effective are the same ones that make genuine authority effective, and only one of them survives contact with a detection system or a regulator.

Your starting checklist

1. Write listings to match how real buyers describe their needs, in full sentences.
2. Load up on verifiable evidence: specs, numbers, real quotes, real citations.
3. Give the model something distinguishing: ratings, review counts, concrete advantages.
4. Structure for machines: comparison tables, tight paragraphs, bolded value props, clean schema.
5. Invest in earned media, per language: not just your own website.
6. Specialize by engine: video and retailer presence for Perplexity, deep owned content for Gemini, top-tier authoritative coverage for ChatGPT and Claude.
7. Measure across repeated queries and multiple engines: AI answers vary run to run.
8. Stay factual. It is the only edge that compounds.

The generative shopping era rewards brands that make themselves easy to justify, not just easy to find. Give the machine a clear, evidence-backed reason to pick you, and it will.

This article synthesizes findings from recent GEO research including Aggarwal et al. (KDD 2024), E-GEO, "How to Dominate AI Search," the Incumbent Advantage brand study, SCI-Defense, and GEO-SFE. Specific percentages reflect 2024-to-2026 experiments and will shift as engines evolve.

Here is what the research actually shows about how Claude sources information, and how to optimize for it specifically.

Claude cites less, but each citation counts more

Claude does not spray citations across an answer the way Perplexity does. It leans on a small set of sources and lets each one carry more weight in the final response. Studies of citation behavior found that the ChatGPT-and-Claude style of engine cites fewer pages, but the pages it does cite exert far more influence on the answer.

The practical implication: with Claude you are not competing to be one of twenty citations, you are competing for one of a handful of slots in a tight, trusted core. And the pages that win those slots share a profile:

• Longer and well-structured
• Dense with extractable evidence: definitions, numbers, comparisons, step-by-step detail

Thin content does not make the cut.

Earned media is everything, and here is Claude's unique gift

All AI engines favor third-party, authoritative sources over brand-owned pages. Claude takes this to an extreme. In study after study it was one of the most "earned-heavy" engines:

What Claude cites is the review site, the expert roundup, the authoritative publication that talks about you. If you want Claude's attention, invest in getting covered by the outlets it already trusts in your vertical (the Consumer Reports and TechRadar-tier sources), not just polishing your own site.

Now the gift. Claude has a property no other major engine shares: it reuses the same trusted authority domains across languages. Where ChatGPT completely swaps to local-language sources when you query in French or Japanese, Claude keeps leaning on its established (often English) authorities. For a brand with limited budget, this is huge, strong English-language earned media travels across markets on Claude. You do not have to rebuild your authority footprint language by language the way you do for ChatGPT.

Claude is the hardest engine to flip, so give it real signals

If you are a challenger trying to unseat an incumbent, Claude is your toughest room. When researchers tested how easily a better competitor could displace a famous brand, the contrast was stark:

And under heavy optimization pressure from every competitor at once, Claude "snapped back" to recommending the established brand almost 99% of the time. The researchers literally called it a "brand guardian."

You do not beat that with cleverness. You beat it with substance. Claude is moved by genuine, verifiable quality signals: real ratings, real review volume, concrete specifications, actual third-party validation. The more objectively you can distinguish your product, the more even Claude will come around.

The counterintuitive part: don't oversell

This is the single most important thing to understand about Claude, and it separates it from every other engine.

• With ChatGPT, authority language works in a straight line: pile on more credibility cues and you get more lift.
• With Gemini, authority saturates fast, so a little goes a long way.
• With Claude, authority follows an inverted-U curve. Moderate, credible signals help. Crank the intensity up, and it backfires.

The research is vivid on this. When one product stacked authority claims and social proof together:

Claude read the pile-on as "too good to be true" and pulled back. It has, in effect, a built-in hype detector.

So the tone that wins on Claude is confident but restrained. State your credentials once, clearly. Cite the one real clinical result, not five breathless ones. A page that reads like a trusted reference beats a page that reads like an infomercial. On Claude, understatement is a strategy.

Structure for how Claude actually reads

Claude retrieves and generates in an integrated, real-time way, pulling content in as it composes. That has direct formatting consequences:

• Write self-contained chunks. Because passages get pulled in individually, each section needs to stand on its own without relying on the paragraph before it. Front-load the key claim in every chunk.
• Keep hierarchy shallow. Deep nested heading structures suit batch-retrieval engines like Google's; for Claude, a flatter, cleaner structure works better.
• Favor lists, tables, and clear formatting. Structured elements parse more reliably than dense prose, and structure alone can lift citation rates meaningfully with no change to the actual facts.
• Go deep, not thin. Remember Claude prefers fewer, richer sources. A thorough, evidence-dense page beats three shallow ones.

Factuality is not just ethics here: it is tactics

On most engines, keeping content honest is the right thing to do and roughly neutral for performance. On Claude, restraint and factuality are actively rewarded. Its skepticism toward over-claiming, its resistance to fabricated-sounding authority, and its "too good to be true" reflex all mean that the honest, precise version of your content genuinely performs better than the hyped version. The engine's design happens to align with good practice. Lean into that.

Your Claude checklist

1. Win coverage in the small set of authoritative, third-party sources Claude already trusts in your category.
2. Lean on your English-language authority. It transfers across markets on Claude better than on any other engine.
3. Give it real, verifiable differentiators: ratings, review counts, concrete specs. Claude does not flip on charm.
4. Keep authority signals moderate and credible. Overselling backfires. Do not stack hype.
5. Write deep, evidence-dense, self-contained sections with clean structure and shallow hierarchy.
6. Stay scrupulously factual. On Claude, it is both the right call and the winning one.

Claude is the engine that makes you earn it. But it is also the engine that, once convinced, stays loyal. Build genuine, well-documented authority and present it with restraint, and you will find Claude quietly recommending you long after the hype-driven tactics stop working on everyone else.

To get cited by Perplexity, don't try to be the best page on your topic. Try to get into the wide pool of sources the engine sweeps on every query. Breadth of coverage and freshness beat the perfection of any single page.

The short answer

Perplexity cites broad and shallow. On average it references 16.35 sources per query (versus 6.88 for ChatGPT), and each source carries little weight individually (mean influence score of 0.065, roughly four times lower than ChatGPT). Its quality bar is also the lowest of the engines measured: the mean quality of the pages it cites is 0.30 out of 1, against 0.69 for Google AI Overviews and 0.73 for Brave.

The strategic consequence is direct: on Perplexity, selection matters more than depth. Being one of the many pages retained in the candidate pool beats being the perfect page the engine never found. That's the exact opposite of what wins on ChatGPT.

What makes Perplexity different from the other engines?

Perplexity is a generative search engine: it reformulates your question, retrieves candidate web pages, summarizes each one, then synthesizes an answer with inline citations. Three traits set it sharply apart from its competitors.

Sources: 2026 breadth/depth study (602 prompts, 21,143 citations); 2025 GEO-16 study (70 B2B SaaS prompts).

Three practical readings follow from this table:

1. The pool is large. With ~16 sources retained per answer, your odds of appearing are mechanically higher than on an engine that keeps only 7. But you have to get in.
2. The quality bar is low. A properly structured, relevant page stands a good chance, where Brave or Google would demand a much higher quality level. This is not an invitation to publish mediocre content; it's the signal that your energy should go to coverage and freshness, not to endlessly polishing a single page.
3. Freshness is the decisive lever. This is where Perplexity diverges most radically.

Why freshness is the number-one lever on Perplexity

On Google, freshness is a tiebreaker. On Perplexity, it's a primary ranking signal: one of the factors most correlated with citation (r = 0.68 in the GEO-16 study, the highest coefficient of all measured pillars).

The numbers are striking. The mean age of sources Perplexity cites, by topic velocity:

And content decays fast in Perplexity's index:

• Day 0–30: peak visibility
• Day 30–45: still strong
• Day 45–90: meaningful decay begins
• Beyond 90 days: largely invisible

This opens what researchers call the "76-day Lazy Gap": a new site, with no established domain authority, can earn citations on Perplexity through freshness alone, before having to compete with legacy sources. It's one of the rare doorways where David can beat Goliath.

What to do concretely:

• Publish machine-readable dates: datePublished and dateModified in JSON-LD, article:published_time in meta, <time> in the HTML.
• Refresh your high-potential pages every 60 to 90 days, but with substantive changes, not cosmetic ones. A modified date with no new content is an empty signal.
• On high-velocity topics, treat freshness as a continuous editorial discipline, not a one-off project.

Don't conflate two things: freshness gets a page cited, but it shouldn't dictate what to crawl first. An old cornerstone article can stay an authority goldmine. Freshness and GEO value are not synonyms.

Getting into the pool: the coverage strategy

Since Perplexity cites broadly, your goal is to be present on as many entry points as possible for a given query. Three levers.

1. Topical breadth over the single perfect page

Instead of pouring all your effort into one monumental pillar page, multiply well-targeted pages that each cover a precise intent. The engine sweeps wide; give it several angles to grab. A cluster of five specialized, fresh pages often outperforms a single massive, frozen one.

2. Earned-media presence

Like all generative engines, Perplexity heavily favors third-party content (press, Wikipedia, expert reviews, vertical platforms) over brand-owned content. Between 79% and 88% of citations go to official, news or vertical sources. The most-cited domains across engines: YouTube, Wikipedia, Reddit, Reuters, LinkedIn.

This means a perfect page on your own domain may never get cited if it's your only presence. To exist in Perplexity's pool, you need to be mentioned elsewhere: press coverage, listings on vertical directories, Reddit discussions, guest articles. This is exactly what Traaker's Earned Media axis measures.

3. Semantic relevance above all

The factor most correlated with citation, across all engines, remains the relevance score between your page and the query (r = 0.43). Before breadth, before freshness, there is intent alignment. A fresh, well-distributed but off-topic page won't enter the pool.

The citation-failure taxonomy confirms it: 62% of cases where a relevant page isn't cited are semantic-alignment problems, not writing-quality problems. The page exists, it's correct, but it doesn't precisely answer what the user is asking. (See our article on diagnostic GEO.)

What helps less than you think

Two widespread beliefs deserve nuance from the data.

Wrapping content in an FAQ doesn't help on its own. Formatting a page as questions-and-answers is a neutral-to-negative signal for absorption (−5.74% in the breadth/depth study) when there's no real evidence density behind it. The FAQ has one virtue (it gets recrawled about twice as often), but don't count on it to get hollow content cited. FAQ form without substance moves nothing.

Perfecting a single page is a poor investment on Perplexity. Given the low influence per source (0.065) and the permissive quality bar, over-polishing one page yields little. The same effort spread across coverage and freshness yields more. Save deep optimization for ChatGPT: see our ChatGPT guide.

The Perplexity action plan in 6 points

1. Date everything, machine-readable. JSON-LD datePublished + dateModified, meta article:published_time, visible <time>. Without a parseable date, you start with a handicap.
2. Refresh every 60–90 days with substantive updates on your high-potential pages.
3. Multiply targeted pages rather than one single pillar. Cover multiple intents.
4. Build your earned media. Press, Wikipedia, vertical directories, Reddit. It's often the decisive unlock when your domain alone isn't enough.
5. Align each page to a precise intent. Semantic relevance is the first filter: before freshness, before breadth.
6. Measure per engine. Perplexity, ChatGPT, Google and Gemini don't behave the same. An aggregate score hides the gaps that matter.

Measure before you optimize

You can't optimize what you don't measure, and on Perplexity, the trap is believing one beautiful page is enough. Traaker measures your visibility engine by engine: which queries Perplexity cites you on, which sources it prefers, where your earned media is too thin, and how your freshness compares to your competitors'.

The Perplexity strategy fits in one sentence: get into the pool, stay fresh, be present everywhere. Breadth first, depth second.

Methodological note. The figures cited come from observational studies (correlations, not proven causal relationships). They are the best public evidence available to date: read them as strong tendencies, not guarantees. Engine behavior evolves; measure yours continuously.

ChatGPT doesn't sweep wide: it picks few sources and leans hard on each. To get cited, you must pull off two things at once: be selected among a small set of pages, and be a source so dense in evidence that the model can't ignore it.

The short answer

ChatGPT cites few and deep. It references on average 6.88 sources per query (less than half Perplexity's 16.35), but each source carries heavy weight: its mean influence score is 0.271, roughly four to five times that of Perplexity or Google. In other words, when ChatGPT cites you, you genuinely shape the answer.

The strategic consequence is twofold, and that's what makes ChatGPT demanding: you must be both selected AND absorbed. Getting into the small set of retained sources isn't enough; you also have to be dense enough in evidence that the model leans on you rather than its six other candidates. This is the inverse of the Perplexity strategy, where getting into the wide pool is the whole game.

What makes ChatGPT different from the other engines?

ChatGPT stands out through two measured behaviors.

Source: 2026 breadth/depth study (602 prompts, 21,143 citations, 18,151 fetched pages).

Two particularities to remember.

1. It compresses under complexity. Faced with a multi-constraint query, ChatGPT reduces its source count: it drops to 3.4 citations, where Perplexity opens up to 17.7. It synthesizes more internally instead of retrieving broadly. The harder the question, the tighter the bottleneck, and the more you must be an unavoidable source to appear in it.

2. It reacts to expert framing, not to source requests. This is a counterintuitive point: where Google and Perplexity cite more when the user explicitly asks "cite your sources," ChatGPT cites more under an expert-role framing. Content that presents itself as an authoritative reference in its field has a specific advantage on this engine.

Lever #1: getting absorbed through evidence genres

Once selected, what determines your influence on the answer is the evidence genre your page contains. It's the most powerful factor identified in the research, and it's largely common to all engines, but it takes on its full meaning on ChatGPT, where each absorbed source weighs enormously.

Influence gain by evidence genre present in the page:

The most influential semantic roles are definition (0.153) and comparison (0.152); passages that merely reference a source without asserting anything are the weakest (0.053).

What to do concretely:

• Open your pages with a clear definition of the concept covered. It's both a high-gain genre and the most influential semantic role.
• Embed structured comparisons: "X vs Y" tables, side-by-side criteria. This very article contains several; that's no accident.
• Densify with sourced numbers. A quantified, attributed claim is worth far more than a generality.
• For technical content, show code and procedures. It's the highest-yield genre (+76.9%).

Lever #2: semantic alignment

After evidence genre, the second signal most correlated with influence is the semantic similarity between your content and the generated answer (r = 0.36), just behind the page-query relevance score (r = 0.43). Both beat raw page length.

The lesson: longer text doesn't help on its own. What helps is answering precisely, in the words and concepts the answer will mobilize. Dense, aligned content beats long, diffuse content.

This is consistent with the citation-failure taxonomy: most relevant uncited pages fail on semantic alignment, not writing quality. The right topic isn't enough, you need the right angle, the right entities, the right level of precision. (See our article on diagnostic GEO.)

Lever #3: authority framing

Since ChatGPT cites more under an expert-role framing, the positioning of your content matters.

• Sign and credit. Identified authors, visible qualifications, displayed expertise. E-E-A-T signals (Experience, Expertise, Authoritativeness, Trustworthiness) are recognized by the model.
• Adopt an assertive tone. Cut hedging language ("maybe," "it seems"). Backed assertions beat cautious phrasings.
• Lean on earned media. Like all generative engines, ChatGPT strongly favors third-party content (press, Wikipedia, expert reviews) over brand-owned content. An authority presence beyond your own domain strengthens your credibility in the model's eyes. This is what Traaker's Earned Media axis measures.

What helps less than you think

Wrapping content in an FAQ isn't enough. The question-and-answer format is a neutral-to-negative signal for absorption (−5.74% in the breadth/depth study) when there's no evidence density behind it. Form doesn't replace substance: a hollow FAQ doesn't get cited.

Raw length isn't a lever. The correlations show it clearly: semantic alignment and evidence genre beat word count. Lengthening a page without densifying its evidence is wasted effort.

Over-optimizing for breadth is a misread on ChatGPT. Given its low citation count and high influence per source, the "be everywhere" strategy that wins on Perplexity is secondary here. Save broad coverage for Perplexity: see our Perplexity guide. On ChatGPT, depth decides.

The ChatGPT action plan in 6 points

1. Open with a definition. High-gain genre (+57.3%) and most influential semantic role.
2. Densify with evidence: sourced numbers, structured comparisons, code and procedures for technical content.
3. Align semantically: match each page to the expected answer; precision before length.
4. Frame with authority: credited authors, assertive tone, visible E-E-A-T signals.
5. Build your earned media to strengthen credibility beyond your own domain.
6. Target complex queries by being unavoidable: ChatGPT compresses to 3.4 sources under constraint, so selection there is even more contested.

Measure before you optimize

On ChatGPT, the trap is the inverse of Perplexity's: being found is useless if you're not dense enough to be absorbed. Traaker measures your visibility engine by engine: which queries ChatGPT cites you on, how much your pages actually shape its answers, and where your evidence genres are too weak to carry weight.

The ChatGPT strategy fits in one sentence: be selected, be dense, be unavoidable. Depth first, breadth second.

Methodological note. The figures cited come from observational studies (correlations, not proven causal relationships). They are the best public evidence available to date: read them as strong tendencies, not guarantees. Engine behavior evolves; measure yours continuously.

A new natural experiment shows most of the lift in a famous case study had nothing to do with the optimization.

Here is an uncomfortable possibility. That impressive AEO case study you bookmarked, the one where ChatGPT referrals jumped 5x after some tweaks, probably measured the wrong thing.

A 2026 paper by Watanabe and Nakayashiki ran the numbers on a single high-traffic website and found that most of the "win" was the platform rising, not the optimization working. AEO, short for Answer Engine Optimization, is the practice of tuning your content so AI assistants like ChatGPT cite and recommend you. The study is worth reading in full (arXiv:2606.04362), but the short version is a useful gut check for anyone selling or buying AEO services.

What they actually did

Most AEO claims rely on third-party traffic estimators. This study did not. The authors used first-party analytics and raw server logs from one real domain, so the referral counts are measured, not modeled.

The setup is what makes it interesting. In January 2026 the team applied AEO changes to one section of the site and left the rest untouched. That untreated remainder became the control group, same domain, same brand, same time window, just no intervention.

This matters because of something the paper calls platform tailwind, the background growth in AI referral traffic that happens to everyone whether they optimize or not. If ChatGPT is sending more clicks to the whole web, your traffic goes up even if you do nothing. A proper test has to separate that tide from your own rowing.

The headline number is mostly tide

Here is the part that should make you skeptical of round-number case studies.

Total ChatGPT referrals to the studied domain grew 5.7x over the window. That is the kind of figure that ends up in a LinkedIn post. But the control pages, the ones that got no AEO work at all, grew 3.5x in the very same period.

So a large share of the "5.7x from AEO" story was happening to pages nobody touched. If you only looked at the treated pages and ignored the control, you would credit the platform's growth to your own intervention. That is the core mistake the paper is built to expose.

So what did the intervention actually do?

To isolate the real effect, the authors used an interrupted time-series model. In plain terms, that means fitting the traffic trend from before the change, projecting it forward, and measuring how far the treated pages broke from that expected path after January.

The estimated true effect of the AEO work was about 1.82x, with a 95% confidence interval of 1.31 to 2.54. A confidence interval is the range the real number probably sits in. So the honest read is "somewhere between a 1.3x and 2.5x lift," not a clean 1.82.

That is still a positive result. It is just a far more modest one than 5.7x, and the gap between those two figures is the whole point.

Why the authors won't call it proven

The paper is careful, and you should be too. The authors ran a placebo-in-time test, which means pretending the intervention happened on a date when nothing actually changed, then checking whether the model "finds" a fake effect anyway. If it does, your method is too twitchy to trust.

That test returned p=0.16. Without getting lost in statistics, that value is not low enough to rule out chance. The pre-intervention period was short and noisy, which makes any before-and-after comparison shakier.

In the paper's own framing, the effect is suggestive, not conclusive. Treat it as a promising signal from one site, not a law of nature.

One piece of good news

There is a quiet win buried in the results. A common fear with AEO is that restructuring content for AI answers will hurt your traditional Google rankings.

That did not happen here. Google organic clicks to the treated pages held steady, and the pages stayed indexed. Optimizing for the answer engine did not cost them the search engine, at least in this case.

What this means for you

• Always run a control. Compare your optimized pages against untouched pages on the same site over the same window. Without that baseline, you cannot tell your work apart from the platform's growth.
• Discount raw multiples. When a case study brags about 5x, ask what the unoptimized pages did. If they grew 3x on their own, the real story is much smaller.
• Measure with first-party data. Server logs and your own analytics beat third-party estimators, which model traffic rather than count it.
• Expect modest, uncertain lifts. A genuine effect in the 1.3x to 2.5x range is plausible and worth pursuing. A clean, confident, huge number usually means the measurement was sloppy.
• You probably won't tank your SEO. Optimizing for AI answers did not reduce Google clicks or indexation in this study, so the "AEO will break my search traffic" fear is not automatic.

The takeaway

None of this says AEO is worthless. It says the way we measure AEO is often broken, and broken measurement inflates the wins.

The discipline the paper models, a control group, a real baseline, and honesty about uncertainty, is the difference between marketing you can act on and marketing you just retweet. Before you trust any AEO success story, including your own, ask one question first: compared to what?

Source: Watanabe & Nakayashiki (2026), "Disentangling Answer Engine Optimization from Platform Growth: A Log-Based Natural Experiment on ChatGPT Referral Traffic," arXiv:2606.04362.

Primarily based on Schulte's "Don't Measure Once: Measuring Visibility in AI Search (GEO)" (April 2026), and its companion paper "Quantifying Uncertainty in AI Visibility: A Statistical Framework for Generative Search Measurement" (March 2026).

Papers: arXiv:2604.07585 | arXiv:2603.08924 | arXiv:2604.25707

The Problem: Your GEO Measurement Is Lying, and You Don't Know It

You run a query against ChatGPT. Your brand gets cited: top 1, paragraph 2. You log it in a spreadsheet, run the next one. You compile 30 queries. You draw a conclusion about your "AI visibility."

Here's the catch: if you re-run the exact same 30 queries 24 hours later, you'll get a different result. Not marginally different, significantly different. Some brands disappear. Others appear. Cited sources change. Order changes. Tone changes.

That's the central observation from a wave of academic research published in spring 2026: AI search is not deterministic, and measuring it as if it were produces wrong numbers.

This collides head-on with traditional SEO practice. For 25 years, measuring visibility meant opening a ranking, reading a position, archiving it. Position 3 is position 3. It doesn't become position 7 between two scrolls. Generative search breaks that property, and most GEO tools on the market haven't yet folded it into their measurement methodology.

The Thesis in One Line

GEO performance must be characterized as a distribution, not as a single point.

That's the formulation in the April 2026 reference paper, and it's the idea every other recent piece of work (academic and industrial) corroborates in its own way.

Practically, if you want to answer "what's my citation share on ChatGPT for this query?", the right answer isn't a number. It's a range: "between 18% and 34% with a 95% confidence interval, measured over N runs across K days." Anything shorter than that is, statistically, noise dressed up as signal.

Three Sources of Variance, and Why They Compound

1. Per-run variance (intra-day)

Same query, same engine, two runs 30 minutes apart: responses differ. The LLMs powering ChatGPT, Gemini, Perplexity or Claude use stochastic sampling (temperature > 0) at generation time. The retrieval pipeline (which pages get fetched before the model answers) introduces a second layer of randomness.

The companion statistical paper (March 2026) quantifies this via bootstrap analysis: across domain-vs-domain comparisons, many apparent differences fall inside the noise floor. In other words: if you observe 22% for brand A and 19% for brand B on a single run, you have not statistically shown that A is ahead of B. You may simply have rolled different dice.

2. Cross-engine variance

The industry has nailed this down with very clear numbers:

• ~11% domain overlap between ChatGPT and Perplexity citations on the same queries (Digital Bloom Report, surfaced via Geneo)
• ~14% URL overlap between Google AI Mode citations and the historical top 10 organic (SE Ranking, surfaced via Geneo)

Translation: ranking well on Perplexity tells you almost nothing about ChatGPT. And ranking well in traditional SEO does not guarantee that you'll appear in AI Mode. These are not three views of the same ranking: they are three different rankings that must be tracked and measured separately.

3. Temporal variance (drift)

Models change. Corpora change. Retrieval pipelines change. A brand that dominates in March may have slipped by May because OpenAI updated its model routing, because a competitor domain published content that got picked up everywhere, or simply because the internal ranker was retrained.

These three variances do not cancel out: they compound. A measurement taken Tuesday morning on Perplexity is not the same as a Thursday-evening one, nor the same as an identical slot on ChatGPT, nor the same as the following week's reading.

The Hidden Structure: Everything Follows a Power Law

The companion paper (arXiv:2603.08924) adds an important finding about the shape of citation distributions: they follow a power law. A small number of domains capture the vast majority of citations; the long tail captures almost nothing.

The industry has confirmed this at scale. An analysis by Trakkr.ai of 1.3 million AI citations spanning 60,209 domains finds exactly the same pattern: citation frequency follows a power law.

Two practical consequences for measurement:

1. Mean metrics lie. The mean of a power-law distribution is dragged by outliers. If you report "the average domain is cited 4.2 times on this query," you are describing a domain that doesn't exist. Prefer the median and percentiles.
2. Gaps between brands are huge in absolute terms, and tiny in the long tail. Moving from rank 80 to rank 60 in the distribution barely changes your citations. Falling out of the top 20, or entering it, changes everything. The useful unit of measurement is not linear.

Citation ≠ Absorption

A third April 2026 paper (arXiv:2604.25707) adds a layer that naive measurement misses: being cited doesn't mean being used.

AI engines distinguish two stages:

• Selection: your URL appears in the list of sources the engine consulted
• Absorption: your content actually fed the generated answer (the produced sentences lean on what you wrote)

A page can be selected and cited at the bottom of an answer without any of the produced sentences reflecting its content. Conversely, a page can be absorbed and reformulated without an explicit citation. Measuring only URL presence in the citation list misses half the picture.

That's why serious GEO measurement needs at least two layers:

1. Selection frequency: rate at which your domain appears in citations across N runs
2. Absorption depth: how much of the generated text reflects your content (lexical overlap, semantic similarity, or explicit in-line citation marking)

Operationalization: How Many Samples?

On the practitioner side, Geneo's measurement guide turns the academic thesis into actionable rules. The standard that's emerging:

• 3 to 5 same-day runs per query per engine, to bound intra-day variance
• Longitudinal tracking over weeks, to detect drift
• Explicit multi-engine coverage (at minimum ChatGPT, Perplexity, Gemini), because cross-engine overlap is too low to extrapolate from any single one

And a metric set adapted to the distributional nature of the signal:

Geneo even proposes concrete alerting thresholds:

• Jaccard overlap < 0.35 for two consecutive days → high instability, do not draw business conclusions
• Drift Rate > 40% week-over-week → something has shifted on the engine or ecosystem side, trigger an investigation

This is the managerial translation of the paper: don't just measure, measure with an alerting policy that distinguishes noise from signal.

iPullRank's Three-Tier Framework

The agency iPullRank has published, in parallel, a conceptual frame that summarizes the new discipline well. Their formulation:

"Share-of-voice is no longer a static percentage of positions held, but a statistical distribution of presence over many trials. Measuring it requires repeated sampling, probabilistic modeling, and acceptance that visibility is not a single snapshot but rather a range of likely outcomes."

They recommend a three-tier measurement stack:

1. Input metrics: what content you produce, which external sources mention you, what your topical coverage looks like (you control everything)
2. Channel metrics: how AI engines treat your content: citation rate, absorption depth, propagation latency after publication (you measure)
3. Performance metrics: business impact: referred traffic, lead quality attributed to AI channels, conversion lifts on audiences exposed to your brand in AI answers (you attribute)

The classic first-generation GEO trap is to measure only tier 2 without connecting it to the others. You end up with a score that moves without being able to explain its cause (tier 1) or measure its consequences (tier 3).

Google's Counterpoint: "GEO and AEO Are a Myth"

In May 2026, Google publicly pushed back, arguing that AI Overviews and AI Mode use the same ranking systems as regular search. Their position: if you rank well in traditional SEO, you'll appear in AI answers. GEO would be SEO rebranded.

That claim is factually true within Google's narrow perimeter. It's still doubly insufficient.

First, the ~14% URL overlap measured between AI Mode and the top 10 organic says the opposite: even at Google, the gap between "well-ranked in SERP" and "cited by AI" is massive. The two systems share infrastructure; they don't do the same thing at answer time.

Second, and this is the central point, Google is talking only about Google. The argument doesn't touch ChatGPT, Perplexity, Claude, or Gemini outside AI Mode. Yet that's precisely where the measurement-variance problem is most acute, because those engines don't use Google's ranking, because their retrieval is different, and because their generation stochasticity is exposed much more directly to the user.

Conclusion: Google's stance is internally consistent for its own ecosystem but doesn't excuse anyone from rigorously measuring the other engines. If anything, it confirms the thesis: each engine has its own statistical regime.

Practical Synthesis: What to Do Tomorrow

If you manage a brand's AI visibility, here's the operational translation of the five sources:

1. Stop reasoning in "my position on ChatGPT for this query." That sentence is malformed. The right phrasing is "my average citation share on ChatGPT for this query across 5 same-day runs, with its confidence interval."

2. Sample 3 to 5 times the same query within the same time window, on every engine you track. A single measurement is forbidden for drawing conclusions.

3. Multi-engine by default. At minimum ChatGPT, Perplexity, Gemini. Cross-engine overlap is too low for any one to be representative.

4. Track stability, not just score. Jaccard overlap, drift rate, survival rate. A brand stable at 18% is in better strategic shape than a brand oscillating between 10% and 30% with an average of 20%: even if the latter has a better "score" on an isolated run.

5. Measure two layers: selection AND absorption. Citation is necessary but insufficient. Without absorption, your brand appears in a footnote nobody reads.

6. Define your alert thresholds before you measure. Jaccard < 0.35 two days in a row, drift > 40% week-over-week: not universal constants, but reasonable starting points to calibrate against your own history.

7. Document your protocols. A GEO benchmark with no measurement protocol description (number of runs, time window, engines covered, model version when capturable) is not comparable to another. Methodology is half the result.

The Bigger Picture

GEO has moved in two years from a nascent discipline (the foundational Princeton/IIT Delhi paper at KDD 2024) to a measurable one. The current phase is where measurement catches up with practice, and where we discover that many of the numbers shared so far are statistical fiction: averages reported without variance, single points treated as trends, cross-engine comparisons made as if it were the same ranking.

The shift is healthy. It forces tools, agencies, and in-house teams to publish protocols, expose confidence intervals, and stop promising certainties that the very nature of generative engines doesn't allow. That's exactly the bet we're making at Traaker: repeated sampling, multi-engine tracking by default, and an explicit separation between selection and absorption.

The message holds in one line: in GEO, a single measurement is not a measurement.

Primary sources: Schulte, "Don't Measure Once: Measuring Visibility in AI Search (GEO)," arXiv:2604.07585, 8 April 2026; "Quantifying Uncertainty in AI Visibility," arXiv:2603.08924, March 2026; "From Citation Selection to Citation Absorption," arXiv:2604.25707, April 2026.

Industry sources: iPullRank, "The Measurement Chasm: Tracking GEO Performance"; Geneo, "Ultimate Guide to AI Search Volatility Tracking"; Trakkr.ai, AI Citation Tracking (analysis of 1.3M citations across 60,209 domains); Google, public stance on AI Overviews / AI Mode (May 2026).

In February 2026, Google shipped WebMCP in a Chrome preview. Behind the acronym sits a W3C standard, co-led with Microsoft, that turns every website into an API callable by AI agents. After GEO (being cited), welcome to the era where you also need to be actionable.

From "answering humans" to "serving agents"

Since 2024, the GEO (Generative Engine Optimization) conversation has focused on one question: how do you get ChatGPT, Perplexity, or Gemini to cite your site in their answers? Statistics, citations, clear structure, earned media, the playbook is now well documented.

WebMCP shifts the problem one level up. When a user tells an AI agent "book me a flight from Paris to Lisbon tomorrow morning on site X" or "buy me those sneakers," citation is no longer the goal. Execution is. But today, agents trying to interact with a "normal" site have to simulate a human: parsing the DOM, clicking visual elements, guessing what each button does. It's slow, unreliable, and breaks at the slightest UI change.

WebMCP takes a different route: let the website declare itself to agents, as a set of typed tools.

What is WebMCP, concretely?

WebMCP (Web Model Context Protocol) is a W3C standard in progress that lets a web page expose its features (submitting a form, running a search, adding a product to a cart, filing a support ticket) as callable tools for an AI agent, directly from the browser.

The project started in summer 2025 as MCP-B ("MCP for the Browser"), created by a former Amazon developer. It is now led by Google (Chrome team) and Microsoft (Edge team), under the WebMCP banner, and shepherded at the W3C. A first draft specification was published in early 2026, and Chrome opened an Early Preview Program for developers.

The idea inherits from Anthropic's Model Context Protocol (2024), which standardizes how an LLM calls external tools. WebMCP is its natural extension to the browser: the web page itself becomes an MCP server, and tools run client-side, inside the user's tab.

Two APIs: declarative and imperative

The standard offers two ways to expose capabilities.

1. The declarative API (for HTML forms)

Just add attributes to an existing <form>. No JavaScript required.

<form action="/search" toolname="search_flights"
      tooldescription="Search available flights between two cities">
  <input name="from" toolparamtitle="Departure city"
         toolparamdescription="IATA code or city name" />
  <input name="to" toolparamtitle="Arrival city"
         toolparamdescription="IATA code or city name" />
  <input type="date" name="date" toolparamtitle="Departure date" />
  <button type="submit">Search</button>
</form>

To an agent, this form is no longer a set of pixels to interpret: it's a search_flights tool with three documented parameters.

2. The imperative API (for everything else)

For more dynamic interactions (actions triggered inside an SPA, filtering, cart updates…), developers use a new JavaScript interface:

navigator.modelContext.registerTool({
  name: "add_to_cart",
  description: "Add a product to the user's shopping cart",
  schema: {
    productId: { type: "string", description: "Product SKU" },
    quantity:  { type: "number", description: "Number of units" }
  },
  async execute({ productId, quantity }) {
    await cart.add(productId, quantity);
    return { ok: true, cartSize: cart.size };
  }
});

The W3C draft defines three main methods on window.navigator.modelContext:

• provideContext(...): full update (handy for SPAs)
• registerTool(...): add an individual tool
• unregisterTool(...): remove a tool

Functions can be asynchronous and delegated to web workers.

The three use cases Google highlights

In its Early Preview Program announcement, the Chrome team calls out three priority areas:

1. Customer support: the agent fills in a ticket with technical context (browser, current URL, error logs) without the user typing it all over again.
2. E-commerce: the agent finds a product, configures options (size, color, customization), adds to cart, and walks through checkout.
3. Travel: the agent searches for flights, filters results, books, all while talking to the user instead of clicking for them.

These are not the only use cases: any interface built around forms or structured actions is a candidate.

What WebMCP changes for GEO

Until now, a brand's presence in the AI ecosystem could be measured on two axes:

• Discoverability: do AI engines mention your brand when a user asks a relevant question?
• Accuracy: is what they say about you correct, up to date, positive?

WebMCP opens a third axis: actionability. When a user tells their agent "order me a pizza," three scenarios emerge:

1. No site in the market implements WebMCP → the agent tries its best to navigate visually, often fails, and ends up recommending a competitor.
2. A single vendor implements WebMCP → the agent systematically picks that site, because it's the only one it can operate reliably.
3. Several sites implement WebMCP → competition kicks back in, but now on the quality of the exposed tools (granularity, speed, clear descriptions).

It's the exact dynamic we saw with responsive design around 2012–2014: first a UX bonus, then a Google signal, then a baseline requirement. Semrush, in its recent coverage, draws the parallel explicitly: sites that move early will capture the emerging agentic traffic; laggards will lose ground.

For the brands we support on Hlight, the implication is direct: a complete GEO strategy in 2026 is no longer limited to editorial content. It must include an actionability audit, are your key user journeys exposable via WebMCP? Are your forms clean enough that a declarative annotation is all it takes?

What it doesn't solve yet

The standard is young, and several questions remain open:

• Security and privacy: the corresponding section of the W3C draft is currently empty. Risks around CSRF, XSS, or abuse combined with emerging features (Prompt API, Web AI) are not yet scoped.
• User consent: when can the agent call a tool without re-asking for confirmation? Guardrails still depend heavily on the MCP client (Claude Desktop, Chrome, etc.), not the protocol.
• Multi-browser support: Chrome opens the ball in preview, Edge follows, Firefox and Safari have not yet committed. Broad availability is expected mid- to late 2026 at the earliest.
• Discoverability of the tools themselves: nothing today says which sites expose WebMCP, nor how agents rank them. This looks suspiciously like a fresh ranking battlefield.

What to do now

Three concrete moves for teams building or running a website:

1. Audit key user journeys. What are the 3 to 5 actions users actually come to your site to do (buy, book, sign up, contact)? Those are the first candidates for WebMCP exposure.
2. Clean up your HTML forms. The declarative API costs almost nothing… provided you have clean <form> elements, explicit name attributes, and sensible labels. That's basic HTML hygiene that also benefits accessibility and SEO.
3. Sign up for Chrome's Early Preview Program. The Chrome team ships docs, demos, and spec changes as they go. Being there early means not redoing the work six months later.

And on the GEO side: keep investing in citation (statistics, sources, earned media), because an agent that doesn't know you will never call your tools, no matter how well exposed they are.

Sources

• webmcp.dev: WebMCP open source library
• Chrome Developers: WebMCP Early Preview Program
• Semrush: What is WebMCP?
• Silicon.fr, WebMCP: interface between web and AI agents
• Le Monde Informatique: The WebMCP API extends web apps to AI agents
• Codely: What is WebMCP and how to use it
• VentureBeat: Google Chrome ships WebMCP in early preview
• Forbes: Google ships WebMCP, the browser-based backbone for the agentic web

GEO Has a Diagnosis Problem

Imagine going to the doctor with knee pain. Instead of examining you, they prescribe a generic treatment: "Take some painkillers, exercise more, eat better." It might help, on average. But if your problem is a torn ACL, this advice is useless. Worse, exercising could make things worse.

This is exactly what first-generation GEO does.

The foundational GEO paper (2024) and its successors identified strategies that work on average: adding statistics, citing sources, improving fluency, adopting an authoritative tone. Some frameworks like AgenticGEO have even automated strategy selection with self-evolving AI agents.

But none of these approaches ask the fundamental question: why, precisely, isn't your content being cited?

This is the shift GEO research is taking in 2026, and it's exactly the approach we're adopting at Hlight.

The Starting Observation: 43% of Relevant Pages Are Never Cited

Recent GEO research has made a striking observation by analyzing existing benchmarks: 43% of thematically relevant web pages receive zero citations from generative engines.

For these pages, the question isn't "how do I increase my citation share from 15% to 20%." It's "why am I completely invisible?"

And when researchers applied classic GEO strategies to these invisible pages, they discovered something troubling: generic optimizations can actively harm niche content. On certain specialized topics, applying GEO "best practices" actually decreased the citation rate.

The problem is structural. Generic strategies are derived from aggregated patterns: what works on average across thousands of pages. But specialized content, niche topics, and underrepresented domains systematically deviate from these patterns. Applying the same recipe to them is like prescribing the same medication to everyone.

The First Taxonomy of Citation Failures

Perhaps the most important contribution of this new wave of research is also the simplest: a systematic classification of why pages aren't cited.

By analyzing 949 contrastive pairs (cases where two pages were retrieved for the same query, but only one was cited), researchers identified four categories of failure, distributed across the entire generative engine pipeline.

1. Technical Integrity (10.1% of cases)

The content doesn't even reach the language model. The causes:

• Access blocking: firewalls, 403 errors, login walls
• JavaScript failure: client-side rendered content that the crawler cannot render
• Unparseable content: corrupted text, binary data, empty strings
• Excessive noise: useful content is buried under ads, navigation, boilerplate

This is the equivalent of a patient who can't even make it to the consultation room. No content optimization can solve a crawling problem.

2. Semantic Alignment (62.2% of cases)

This is the dominant category, and the most nuanced. The content reaches the model, but it doesn't match what the query is asking for:

• Intent divergence: informational content for a transactional query (or vice versa)
• Contextual gap: the right topic, but missing the specific entities or jargon expected
• Outdated information: stale or temporally misaligned data
• Location mismatch: British regulations for an American query

62% of failures. In other words, in the majority of cases, the problem isn't that your content is poorly written: it's that it doesn't precisely answer what the user is looking for.

3. Content Quality (27.1% of cases)

The content addresses the right topic but presents it poorly:

• Informational poverty: too shallow to be citable
• Fragmentation: disconnected snippets that resist synthesis
• Excessive verbosity: key facts are diluted in filler
• Unstructured layout: dense prose where tables or lists would help

This is where classic GEO strategies are most relevant, but only if the diagnosis is correct.

4. Systemic Exclusion (0.6% of cases)

The content is good, relevant, well-presented, but faces structural disadvantages:

• Competitive redundancy: a higher-authority source (Wikipedia, for example) covers the same facts
• Window truncation: relevant content is buried too deep to fit within the model's context window

This last case is the most frustrating: no content optimization can solve it. If Wikipedia says the same thing you do, the AI engine will cite Wikipedia.

The Diagnostic Approach: Examine Before Prescribing

Armed with this taxonomy, the new generation of GEO tools (including Hlight) adopts a radically different approach: diagnose first, then repair in a targeted manner.

The Principle: Diagnose, Then Repair

For each uncited page, a diagnostic system follows an iterative cycle:

1. Diagnosis: Compare the page with the highest-ranked cited competitor. Identify precisely why the competitor was preferred, classifying the vulnerability according to the taxonomy.
2. Tool selection: Choose the appropriate intervention from a library of specialized tools, taking into account previous attempts (memory).
3. Repair: Apply the tool to a copy of the page.
4. Verification: Test whether citation is achieved. If not, re-diagnose and iterate.

This is fundamentally different from the "apply strategy X to everyone" approach. Each page receives a personalized treatment based on its specific problem.

Repair Tool Categories

A library of diagnostic tools covers four functional categories:

Information augmentation:

• Entity injection: surgically insert missing facts or entities at optimal points in the text
• Data serialization: convert narrative descriptions into structured HTML tables

Structural improvement:

• Structure optimization: transform "walls of text" into hierarchical content with headings, lists, and emphasis
• Noise isolation: separate useful content from boilerplate (navigation, ads, footers) via semantic tags

Content positioning:

• BLUF optimization (Bottom Line Up Front): extract key points and place them in a summary at the top of the page
• Content relocation: surface buried content via "TL;DR" or "Key Points" sections
• Intent realignment: rewrite the opening paragraph to directly address the query's intent

Persuasive refinement:

• Persuasive rewriting: adopt an authoritative tone, add social proof, counter-arguments
• Historical red-teaming: contextualize dated content by creating links between past and present

Memory Prevents Loops

A crucial detail: an effective diagnostic system maintains a per-query memory that records previous attempts. If a tool has already failed for the same type of vulnerability, it is excluded from the options. If a tool fails twice consecutively, it is globally excluded from the current path.

The system also has escalation protocols: if factual augmentation doesn't work, it switches to persuasive rewriting. If structural reorganization fails, it forces a BLUF summary at the top of the page.

The Results: +40% Citations by Touching 5% of the Content

The research results on the diagnostic approach speak for themselves:

Three major observations:

1. Surgical efficiency. The diagnostic approach modifies only 5% of the original content, compared to 25% for generic methods. And it achieves better results. This confirms that citation failures are rarely a problem of overall quality: most pages need targeted corrections, not a massive rewrite.

2. Content preservation. With a Jaccard score of 82.4% (compared to 18% for generic rules), the diagnostic approach preserves the essence of the original content. Generic methods, by rewriting 25% of the text, distort the content, which is problematic for creators who care about their voice and message.

3. Cross-method robustness. Optimized with one citation method (Attribute-First), the diagnostic approach also improves results with another method (In-Context): +14.3% citation rate. The repairs are fundamental, not engine-specific.

Generic Optimizations Can Be Harmful

One of the most important findings concerns the analysis by topic. On certain thematic categories, generic rules perform worse than doing nothing at all.

This is particularly visible on topics where the baseline citation rate is already high (such as health, around 80%). Generic rules, by massively rewriting the content, sometimes remove domain-specific information that was precisely the reason the content was cited in the first place.

The diagnostic approach, by contrast, shows consistent gains across all topics: precisely because it diagnoses before acting and only touches what needs to be touched.

The lesson is clear: in GEO, not optimizing can be better than optimizing blindly.

What Diagnostics Cannot Solve

The research also has the honesty to document its limitations. Even after diagnostic optimization, some queries remain uncited.

The analysis of these failures reveals a recurring pattern: competitive dominance. A university page on machine learning, even perfectly optimized, won't be cited over Coursera or edX for the query "best online machine learning courses." The AI engine has an internal bias toward sources with high domain authority, a factor external to the content itself.

This is an important conclusion for the ecosystem: if certain content is systematically disadvantaged regardless of optimization effort, the citation mechanisms of AI engines amplify certain voices at the expense of others. Creator-side optimization alone cannot guarantee equitable visibility.

The Broader Context: Structure Matters Too

This diagnostic shift isn't limited to a single tool. Other recent work converges toward the same conclusion: GEO optimization must be targeted and multidimensional.

Recent research on structural feature engineering demonstrates that document structure influences citation as much as its semantic content. By decomposing structure into three levels:

• Macro-structure: the overall architecture of the document (sections, hierarchy)
• Meso-structure: how information is divided (paragraphs, chunks)
• Micro-structure: visual emphasis (bold, lists, tables)

...researchers achieve +17.3% citation rate and +18.5% subjective quality across 6 generative engines, without modifying the meaning of the content. Just by restructuring.

Other work goes even further by arguing that the entire RAG paradigm is fundamentally limited for GEO. The concept of Semantic Entropy Drift mathematically models the inevitable decay of LLM confidence over time, which means that any textual optimization is by nature transient.

Practical Implications: What to Do Now

For Content Creators

1. Stop applying generic recipes blindly. "Add statistics everywhere" can harm your niche content. First identify why you aren't being cited.

2. Check technical integrity. 10% of failures come from the fact that the AI engine can't even read your page. Test rendering without JavaScript, verify that your main content isn't buried in boilerplate.

3. Align with intent. 62% of failures are semantic alignment problems. Does your page truly answer the question the user is asking? With the right entities, the right geographic context, up-to-date data?

4. Structure for the machine. Research confirms that structure (headings, lists, tables) helps AI engines extract and cite your content. A "wall of text" is your enemy.

5. Put the essentials first. The BLUF (Bottom Line Up Front) principle is one of the most effective tools of the diagnostic approach. If your key answer is in paragraph 15, the AI engine may never find it.

For Hlight Users

This is exactly the diagnostic philosophy we're integrating into Hlight. Rather than applying uniform transformations, our approach first analyzes why your content isn't being cited, then applies targeted corrections adapted to your specific situation. The result: more visibility, fewer modifications, and content that stays true to your message.

For Research

The citation failure taxonomy is a reusable framework. Future work can extend it, refine it, and most importantly validate it on commercial production engines (Google AI Overviews, Perplexity, ChatGPT Search).

The MIMIQ Benchmark: Evaluating Generalization

An important methodological contribution of this research is MIMIQ (Multi-Intent Multi-Query), a document-centric rather than query-centric benchmark.

Existing GEO benchmarks associate each document with a single query. In practice, a content creator can't anticipate users' exact queries. MIMIQ associates each page with 60 queries covering diverse intents, personas, and phrasings, with a train/test split.

This allows testing whether an optimization produces genuinely more citable content, or whether it over-optimizes for a specific phrasing. The diagnostic approach, thanks to its batch aggregation, performs well on this generalization test.

The Big Picture: From SEO to Diagnostics

GEO is undergoing a transition that echoes the evolution of medicine. We're moving from general practice ("take these vitamins, they help on average") to precision medicine ("here is your specific diagnosis, here is the targeted treatment").

First-generation GEO strategies (adding stats, citing sources, improving fluency) remain useful as basic hygiene. But for the 43% of pages that are completely invisible, they aren't enough. You need to understand why the AI engine ignores your content, and intervene at the right place.

Research shows that this diagnostic approach is not only more effective (+40% citations) but also more respectful of the original content (5% modifications vs 25%). It's a better outcome with less intervention: the hallmark of a correct diagnosis.

The message for content creators is clear: before optimizing, diagnose. The answer to "how do I get cited more by AI?" starts with "why am I not being cited today?"

Based on the paper "AgenticGEO: A Self-Evolving Agentic System for Generative Engine Optimization" by Yuan, Wang, Wang, Sun, Wang & Li (Beihang University). Published in March 2026.

Paper: arXiv:2603.20213 | Code: github.com/AIcling/agentic_geo

Quick Refresher: What Is GEO?

If you haven't read our previous article on the foundational GEO paper, here's the short version:

AI search engines (ChatGPT, Perplexity, Google AI Overviews) no longer show you a list of links: they read websites and give you a synthesized answer. Generative Engine Optimization (GEO) involves modifying your web content so that these AI engines are more likely to include and cite it in their responses. It's SEO, but for the AI era.

The original GEO paper (2024) showed that strategies like adding statistics, quotations, and references could increase visibility by up to 40%. But these were fixed strategies, identical for everyone: the same trick applied to every piece of content.

AgenticGEO asks the question: what if an AI agent could find the best strategy for each specific piece of content, and continuously improve?

The Problem With First-Generation GEO

The foundational GEO paper tested 9 strategies (adding citations, stats, improving fluency, etc.) and showed they work, on average. But when AgenticGEO researchers looked more closely, they found a critical problem:

Nearly half of all content couldn't be improved by any of the existing fixed strategies.

Think about it: telling a legal article to "add statistics" might help, but telling a personal essay to "add statistics" could make it worse. The original GEO approach was like a doctor prescribing the same medication to every patient, regardless of their symptoms.

Three specific problems:

1. Content diversity: Different articles require different optimization approaches. A recipe blog and a scientific paper have nothing in common.
2. Engine unpredictability: AI search engines are black boxes that constantly change. A strategy that works today on Perplexity might not work tomorrow, or on Google AI Overviews.
3. Feedback is expensive: To know if your optimization worked, you need to submit the content to the AI engine and check the result. Doing this thousands of times for each piece of content is impossible.

The Core Idea: An AI Agent That Evolves Its Own Strategies

AgenticGEO rests on a surprisingly elegant concept: instead of humans designing fixed strategies, let an AI agent discover, evolve, and select strategies automatically, and keep improving over time.

The system has three stages: Learn, Evolve, and Apply.

Stage 1: Learn (Offline Critic Alignment)

Before any online action, AgenticGEO trains a small, lightweight AI model called the Critic. Think of the Critic as an intern studying past examples to develop judgment.

Here's what happens:

• Take a set of content + query pairs
• Apply various rewriting strategies to each
• Run them through a generative engine and measure the results
• Train the Critic to predict: "Given this content and this strategy, how much will visibility improve?"

The Critic is built on a small language model (Qwen2.5-1.5B, tiny by current standards) with a scoring head on top. It learns to predict visibility gains without needing to query the expensive generative engine each time.

Why this matters: Once trained, the Critic can evaluate thousands of strategy-content combinations in seconds, instead of waiting for slow and expensive AI engine responses.

Stage 2: Evolve (Online Co-Evolution)

This is where it gets interesting. AgenticGEO uses an approach borrowed from evolutionary biology called MAP-Elites: a technique for maintaining a diverse population of high-quality solutions.

The Strategy Archive

Imagine a library of optimization strategies. Each strategy is a set of instructions for rewriting content (e.g., "add authoritative citations and restructure into clear sections" or "simplify language and add statistical evidence"). But unlike the 9 fixed strategies of the original GEO, this library:

• Starts with seed strategies (similar to those in the original GEO)
• Evolves new strategies by mutating and combining existing ones
• Keeps only the best AND most diverse: not just the single best performer

Diversity is crucial. MAP-Elites doesn't just keep the top performer. It maintains a grid of strategies that are each the best in their niche. One strategy might be the best for scientific content, another for opinion pieces, another for e-commerce product pages. The archive preserves them all.

The Co-Evolution Loop

Here's the cycle:

1. Select a parent strategy from the archive
2. Mutate it using an LLM "Evolver" (Qwen2.5-7B) that generates variations
3. The Critic scores the new strategy at low cost (no engine call needed)
4. Occasionally verify against the real generative engine (expensive but necessary)
5. Update the archive: if the new strategy is good AND sufficiently different, it earns a spot
6. Update the Critic: use the real engine feedback to keep the Critic calibrated

The Critic and the archive evolve together: hence "co-evolution." The Critic gets better at judging strategies, and the strategies get better at optimizing content. It's a virtuous cycle.

Reducing Engine Calls

Key innovation: the Critic acts as a gatekeeper. Instead of testing every new strategy against the real AI engine (expensive), the Critic pre-filters candidates. Only the most promising ones are verified with real engine calls.

The result: AgenticGEO retains 98.1% of its performance while using only 41.2% of the engine feedback that would otherwise be required. This is a massive cost reduction.

Stage 3: Apply (Multi-Turn Agentic Rewriting)

At inference time (when you actually want to optimize a piece of content), AgenticGEO doesn't just apply a single strategy. It launches a multi-step planning process:

1. The Critic analyzes your content and the target query
2. It selects the top 25 most promising strategies from the evolved archive
3. It picks the best one and applies it (rewriting via Qwen2.5-32B)
4. It evaluates the result and decides: "Is this good enough, or do I need to apply another strategy?"
5. It can chain up to 3 rewriting steps, each building on the previous one

This is the "agentic" part: the system makes autonomous decisions about what to do, evaluates its own work, and iterates. It doesn't just execute a template; it plans and adapts.

The Results: State of the Art Across the Board

AgenticGEO was tested against 14 baselines (including the original GEO strategies, AutoGEO, and other methods) on 3 datasets and 2 different generative engines (Qwen2.5-32B and Llama-3.3-70B as underlying GE models).

Key numbers:

• 46.4% average visibility gain compared to unoptimized content
• Outperforms all 14 baselines on every dataset/engine combination
• Cross-domain transfer works: strategies evolved on one domain (e.g., science) transfer well to unseen domains (e.g., law, health)
• Cross-engine transfer works: strategies optimized for one AI engine also improve visibility on another

That last point is particularly important. It means the system isn't "gaming" a specific engine: it's learning to make content genuinely better in ways that multiple AI systems recognize.

Why "Self-Evolving" Matters

Most AI optimization systems are static: you train them once, deploy them, and they gradually become outdated. AgenticGEO is designed to keep evolving:

• New content types? The archive evolves new strategies.
• Engine behavior changes? The Critic recalibrates with fresh feedback.
• New domains? Cross-domain transfer provides a starting point, and evolution fills the gaps.

The researchers provide a theoretical argument showing that the co-evolution process achieves sub-linear regret: the system's errors decrease over time at a mathematically guaranteed rate.

Comparison With Other Approaches

Practical Implications

For Content Creators

The immediate takeaway is that personalized optimization beats generic advice. "Add statistics everywhere" is correct on average, but the right strategy depends on your specific content, your domain, and the AI engine you're targeting. Tools built on AgenticGEO-type approaches could eventually offer content-specific recommendations.

For the GEO Research Community

AgenticGEO shifts the field from "which fixed strategy is best?" to "how do we build systems that discover and adapt strategies automatically?" This is a fundamentally different (and more scalable) research direction.

For AI Engine Developers

The paper raises an important question: as GEO tools become more sophisticated, will AI engines need to evolve their defenses? There's an emerging arms race between content optimizers and the engines that consume that content. AgenticGEO's evolutionary approach is particularly hard to counter because it doesn't rely on any single exploit, it continuously discovers new ones.

Under the Hood: Technical Stack

For those curious about the implementation:

• Critic model: Qwen2.5-1.5B with LoRA fine-tuning (2 epochs) + MLP value head
• Strategy Evolver: Qwen2.5-7B-Instruct
• Content Rewriter: Qwen2.5-32B-Instruct
• Generative engines tested: Qwen2.5-32B-Instruct and Llama-3.3-70B-Instruct
• Archive: MAP-Elites quality-diversity
• Inference: Top-25 strategy selection, up to 3 rewriting steps
• Training: Hybrid loss (Huber regression + pairwise rank-aware alignment)
• License: MIT (fully open source)

The Big Picture

AgenticGEO represents a shift in how we think about content optimization for AI search. The original GEO paper said "here are 9 tricks that help." AgenticGEO says "let's build a system that discovers its own tricks, picks the right one for each situation, and keeps getting better."

It's the difference between a cookbook and a chef. The cookbook gives you recipes. The chef understands ingredients, techniques, and diners' preferences, and invents new dishes as needed.

As AI search engines become the dominant way to find information, the ability to stay visible in their responses becomes existential for content creators. AgenticGEO suggests that the future of this optimization won't be a checklist of tips: it will be AI agents optimizing content for other AI agents.

Welcome to the meta-game.

Paper: Yuan, J., Wang, J., Wang, Z., Sun, Q., Wang, R., & Li, J. (2026). AgenticGEO: A Self-Evolving Agentic System for Generative Engine Optimization. arXiv:2603.20213

Based on academic research (Princeton, Beihang University) and field insights from Semrush, HubSpot, and Microsoft.

1. Include Statistics and Hard Data

The foundational GEO paper (Princeton, 2024) analyzed 10,000 real queries. Result: pages containing statistics and citations have 30 to 40% higher visibility in AI responses compared to content without data.

Language models love concrete numbers. They use them to build credible responses.

Action: add sourced data to every article. Prioritize your own research and exclusive data.

2. Structure Your Content for Extraction

AI engines don't read like humans. They extract information blocks. Chris Long (VP Marketing, Go Fish Digital) tested and confirmed: AI pulls clear snippets from highly structured content.

Action:

• Use bullet points and numbered lists
• Create subheadings in question form
• Add comparison tables
• Include FAQ sections

3. Earn Mentions on Third-Party Sources (Earned Media)

Research shows that AI engines have a systematic bias toward earned media (third-party, authoritative sources) over brand-owned content. This is one of the biggest shifts from classic SEO.

Action: aim for mentions in specialized press, citations in third-party articles, podcast appearances, and guest contributions.

4. Unlinked Mentions Matter

In classic SEO, a backlink is king. In GEO, brand mentions without links appear to carry significant weight. AI detects textual mentions even without hyperlinks.

Action: don't overlook brand mentions on Reddit, YouTube, forums, and social media. Even without a link, they feed your AI visibility.

5. Add Expert Quotes and Citations

Content that includes direct quotes from experts ranks better with AI engines. This strengthens E-E-A-T signals (Experience, Expertise, Authoritativeness, Trustworthiness).

Action: integrate quotes from recognized experts in your field. Cite your sources with links to the original studies.

6. Keep Your Content Up to Date

AI engines favor recent information. A 2022 article that hasn't been updated loses relevance against a competitor updated in 2026.

Action: regularly revise your key articles. Add the latest data and trends. Display the update date.

7. Think Server-Side Rendering (SSR)

AI crawlers struggle to execute JavaScript. If your content relies on client-side rendering (a React SPA, for example), it may be invisible to generative engines.

Action: use server-side rendering (SSR) or static site generation (SSG). Verify that your content is accessible without JavaScript.

8. Tend to Your Wikipedia and UGC Presence

Wikipedia accounts for a significant share of LLM training data. UGC platforms (Reddit, YouTube, forums) also have strong exposure in generative engines.

Action:

• If your brand is eligible, make sure you have an accurate and up-to-date Wikipedia page
• Be active on Reddit and YouTube in your niche

9. Make Your Content Machine-Readable

Microsoft's official guidelines for generative search emphasize: make your catalogs and content machine-readable.

Action:

• Use Schema.org markup (structured data)
• Offer structured formats (JSON-LD, tables)
• Avoid content locked in images or non-indexable PDFs

10. Measure Your AI Visibility

You can't optimize what you don't measure. New tools allow you to track your presence in AI responses.

Available tools:

• Semrush AI Visibility Toolkit: mentions, citations, share of voice by AI platform
• HubSpot AI Search Grader: free benchmark of your AI visibility
• Semrush Enterprise AIO: advanced tracking of sentiment and citations

Summary

Sources

• GEO: Generative Engine Optimization (arXiv:2311.09735)
• Semrush: GEO Guide
• HubSpot: Generative Engine Optimization
• Microsoft: AEO and GEO Guide (PDF)
• Chris Long on LinkedIn: GEO Tests

SEO is your ticket to the party. GEO is your invitation to the head table. (Neil Patel)

The False Debate: SEO vs GEO

Every time a new technology emerges, we hear the same thing: "SEO is dead." It was already the case with social media, with voice search, with featured snippets. And now with generative AI.

The reality is more nuanced. GEO doesn't replace SEO, it extends it.

Why SEO Remains Essential

Classic search engines aren't going anywhere. Google still processes billions of queries per day. Many users still prefer to click on links, compare sources, and navigate between sites.

Moreover, AI search engines rely on the same content indexed by Google. If your content isn't well-referenced in SEO, it has little chance of being found and cited by an AI.

SEO remains the foundation:

• Crawlability and indexation
• Heading and tag structure
• Page speed and mobile-first
• Backlinks and domain authority

What GEO Adds

GEO introduces new dimensions that classic SEO doesn't cover:

Proof by Example: HubSpot

HubSpot ranks #3 on Google for the keyword "CRM." And when you ask ChatGPT what the best CRM is, HubSpot appears as the top recommendation.

This isn't a coincidence. E-E-A-T signals (Experience, Expertise, Authoritativeness, Trustworthiness) transfer from one platform to another. Good SEO naturally feeds GEO.

The Lorelight Lesson

Lorelight was a GEO-focused startup that offered to monitor where your brand appears in AI responses. It shut down. SEO expert Aleyda Solis commented: tools that simply show your AI visibility aren't enough. You need to integrate GEO into a comprehensive search strategy that includes SEO.

Separating SEO and GEO into two distinct strategies doesn't make sense. Both share the same fundamentals.

The Integrated Approach

The right strategy in 2026:

1. Continue your SEO: it's the foundation everything rests on
2. Add a GEO layer: structure your content so it's extractable by AI, include data and citations, earn mentions on third-party sources
3. Measure both: track your classic rankings AND your visibility in AI responses (Semrush AI Visibility, HubSpot AI Search Grader)

In Summary

GEO is not a replacement for SEO. It's a natural extension for a world where users get their answers directly from AI. Brands that invest in both will be present everywhere: in the links, in AI responses, and in users' minds.

Sources

• HubSpot: GEO vs SEO
• Semrush: GEO Guide
• Neil Patel: GEO
• Aleyda Solis on LinkedIn: Lorelight Shutdown

In 2023, GEO was a research paper from Princeton. In 2026, it's a topic covered by The Verge, Semrush, HubSpot, and Neil Patel. A look back at a meteoric adoption.

From Academic Research to Practical Guides

In November 2023, researchers from Princeton and IIT Delhi published the foundational paper "GEO: Generative Engine Optimization." The idea: optimizing your content no longer for Google's classic results, but to be cited in AI-generated responses, ChatGPT, Perplexity, Google AI Overviews, Claude.

Two years later, the concept has left the labs. Here's what happened.

The Signs of Massive Adoption

The Verge Is Covering It

As early as August 2025, The Verge covered the phenomenon with a clear message from the experts interviewed: "Say hello to GEO." In January 2026, at the NRF Retail Expo, the outlet noted that SEO had "branched into new acronyms: AEO, GEO", a sign that the concept had entered the vocabulary of commerce and marketing.

Digital Marketing Giants Are On Board

• Semrush published a comprehensive guide and launched a dedicated toolkit (AI Visibility Toolkit) to measure visibility in generative engines
• HubSpot released a detailed guide and a free tool (the AI Search Grader) to benchmark your presence in AI responses
• Neil Patel published a GEO guide, stating that "GEO isn't the future, it's already the present"
• Microsoft published official guidelines for optimization in generative search

Academic Research Is Exploding

On ArXiv, the number of papers mentioning "Generative Engine Optimization" has exploded in 2026:

• AgenticGEO (March 2026): AI agents that automatically optimize content
• AgentGEO (March 2026): diagnosis and repair of citation failures
• SAGEO Arena (February 2026): first realistic end-to-end benchmark
• Pinterest GEO (February 2026): production deployment at the scale of billions of images, +20% organic traffic
• E-GEO (November 2025): first e-commerce benchmark with 7,000+ queries

Real-World Production Cases

Pinterest deployed GEO at scale: by fine-tuning VLMs to predict what users would search for (rather than simply describing images), they achieved +20% organic traffic and growth of several million monthly active users.

Why Now?

Three factors explain this acceleration:

1. Massive adoption of AI engines: ChatGPT surpasses 200 million users, Google AI Overviews reaches billions of people every month
2. Measurable impact on traffic: sites that don't appear in AI responses lose visibility, even if they rank well in classic SEO
3. Tooling is arriving: Semrush, HubSpot, and others now offer tools to measure and optimize your presence in AI responses

What This Changes for You

If you create content (whether you're a company, a media outlet, a blog, or a freelancer), GEO is no longer optional. AI queries use an average of 23 words (compared to 4 on classic Google). Users ask complex questions and expect synthesized answers.

The question is no longer "should I pay attention to GEO?" but "how do I integrate it into my existing strategy?"

Sources

• The Verge: ChatGPT Search and GEO (August 2025, January 2026)
• Semrush: GEO Guide
• HubSpot: Generative Engine Optimization
• Neil Patel: Generative Engine Optimization
• Pinterest GEO (arXiv:2602.02961)
• AgenticGEO (arXiv:2603.20213)

Based on the paper "SAGEO Arena: A Realistic Environment for Evaluating Search-Augmented Generative Engine Optimization" by Kim, Jeong, Kim, Lee & Lee (Yonsei University / Konkuk University). Published in February 2026.

Paper: arXiv:2602.12187

The Problem: GEO Lives in a Simplified World

Since the foundational GEO paper in 2024, researchers have proposed numerous strategies to improve content visibility in AI search engine responses. Adding statistics, citations, improving fluency: all of this works. In the lab.

But there's a fundamental problem: all existing benchmarks are cheating.

How? They skip the hardest steps. When Perplexity or Google AI Overviews answer a question, here's what actually happens:

1. Retrieval: the engine searches through billions of pages and selects a few hundred
2. Reranking: a model reranks these pages to keep the top 5-10
3. Generation: an LLM synthesizes a response from these selected pages

Existing benchmarks (GEO-Bench, AutoGEO, etc.) only test step 3. They feed the documents directly to the LLM and measure whether optimization improves citation. They completely ignore steps 1 and 2.

It's like testing a sprinter only on the last 10 meters of a 100-meter race. It says nothing about their ability to run the full race.

What SAGEO Arena Changes

SAGEO Arena is the first benchmark that tests the end-to-end pipeline: retrieval, reranking, and generation. It also introduces a key concept everyone was ignoring: structural information.

A Real Web Corpus at Scale

Unlike previous benchmarks that use raw text, SAGEO Arena built a corpus of 170,000 real web documents covering 9 domains. For each document, they extracted:

• The body text, the main content
• The structural information: title, meta description, heading hierarchy (h1, h2...), and schema markup

Why does this matter? Because in the real world, search engines heavily use these structural signals to decide which documents to retrieve and rank. Ignoring them means ignoring half the game.

A Complete and Modular Pipeline

SAGEO Arena integrates all three stages of the generative search pipeline:

Each stage is modular and configurable. Researchers can trace exactly how an optimization affects each stage of the pipeline.

The Evaluation Protocol

The process is simple but rigorous:

1. Run the pipeline with a test query
2. Select a target document from those that reach the generation stage
3. Apply a GEO optimization strategy to that document
4. Reinsert the optimized document into the corpus
5. Rerun the pipeline with the same query
6. Measure: did the document gain, maintain, or lose visibility at each stage?

The Shocking Finding: GEO Breaks Retrieval

This is the paper's most important result, and it calls the entire field into question:

Existing GEO strategies sometimes improve generation, but actively degrade retrieval and reranking.

Concretely: when you optimize your page's text with classic GEO techniques (adding stats, citations, etc.), your page can disappear from search results before even reaching the LLM. The optimized document is no longer retrieved by the search engine, so it has zero chance of being cited.

It's as if you improved your store's window display, but the GPS stopped directing customers to your street.

Why Does This Happen?

GEO strategies modify body text to please the LLM (step 3), but these modifications can:

• Alter the signals the retrieval engine uses to find the document (step 1)
• Change the perceived relevance for the reranking model (step 2)

Optimizing for one stage can sabotage the others. This is a problem nobody had measured before SAGEO Arena.

Structural Information: The Unexpected Hero

The second major finding: structural information (titles, meta descriptions, schema markup) is essential for maintaining visibility at the retrieval stage.

The researchers found that the two types of information play complementary roles:

In other words: structure gets your document into the race, and content determines whether it wins. Optimizing only content without touching structure is like having an excellent product with no marketing.

This is a direct bridge between traditional SEO (which focuses on structure) and GEO (which focuses on content). Both are necessary.

SAGEO: Stage-Aware Optimization

Armed with these findings, the researchers propose stage-aware SAGEO, an approach that adapts optimization to each stage of the pipeline:

• For retrieval: optimize structural information (titles, meta descriptions, schema markup)
• For reranking: optimize relevance and text quality
• For generation: apply classic GEO techniques (stats, citations, fluency)

This combined approach achieves the best overall visibility among all evaluated strategies. It's the first time a method jointly optimizes SEO and GEO.

Comparison With Existing Benchmarks

The difference is striking: SAGEO Arena is the only one covering the complete pipeline with real web documents.

What This Means for Content Creators

1. SEO Is Not Dead

Contrary to what many claim, traditional SEO remains crucial. Structural information (titles, meta descriptions, schema markup) is what enables your content to be found by AI search engines in the first place. Without it, your GEO optimizations are useless.

2. Don't Sacrifice Retrieval for Generation

If you rewrite your content to please LLMs, verify that your page is still properly indexed and retrieved by search engines. Content perfectly optimized for generation but invisible to retrieval has zero visibility.

3. Think in Pipelines, Not Silos

Effective optimization requires working on all three stages:

• Structure: to be found
• Relevance: to be well-ranked
• Content quality: to be cited

4. Schema Markup Matters

Structured data (JSON-LD, schema.org) isn't just a SEO bonus. It plays an active role in how AI search engines understand and select your content.

Limitations to Keep in Mind

• The benchmark uses open-source models to simulate the pipeline, not actual commercial engines (Google, Perplexity, etc.)
• The 170K document corpus is large but remains smaller than the real web
• Off-page factors (backlinks, domain authority) are not modeled

Despite these limitations, SAGEO Arena represents a major qualitative leap over previous benchmarks.

The Big Picture

SAGEO Arena is a paper that should make the entire GEO community think. Its central message is simple but powerful:

Optimizing only for the last stage of the pipeline is insufficient, and potentially counterproductive.

The future of optimization for AI search is neither SEO alone nor GEO alone, but an integrated approach that understands and optimizes each stage of the generative search pipeline. SAGEO Arena finally provides the tool to measure this.

For content creators, the lesson is clear: keep your SEO fundamentals solid (structure, metadata, schema) while applying GEO techniques (stats, citations, fluency) to your content. One without the other isn't enough.

Paper: Kim, S., Jeong, W., Kim, S., Lee, S., & Lee, D. (2026). SAGEO Arena: A Realistic Environment for Evaluating Search-Augmented Generative Engine Optimization. arXiv:2602.12187