How to Get Your Products Recommended by AI: A Field Guide to Generative Engine Optimization
Your customers have quietly changed how they shop. Instead of typing "best running shoes" into Google and scrolling ten blue links, they now ask ChatGPT "I'm a marathon runner with flat feet, what shoe should I buy?" and get a shortlist of three, with reasons. Perplexity, Gemini, and Claude do the same. The recommendation is the storefront now.
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:
| Gate | Question it answers | What it depends on |
|---|---|---|
| Retrieval | Does your page even get pulled into the model's working set? | How closely your listing embeds to how real buyers phrase their needs |
| Ranking | Once there, does the model pick you for the shortlist? | Your distinguishing product attributes |
| Absorption | Does your content actually shape the answer, or just get cited in passing? | Structure and evidence density |
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:
| Distinguishing signal | Threshold that flipped the recommendation |
|---|---|
| Rating edge | as small as 0.075 stars |
| Review count | 1.6× higher |
| Price | 7.3% lower |
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 — every winner is verifiable:
| Technique | Effect | Why it works |
|---|---|---|
| Statistics and hard numbers | Highly effective, most future-proof | "6061-T6 alloy, 3.2mm thick" beats "durable aluminum" |
| Quotations from credible sources | Up to +40% visibility — the single best move in the foundational GEO study | Gives the engine something to cite |
| Citations and references | Strong lift | Authoritative sources justify the pick |
| Comparison tables & pros/cons | +43% extraction accuracy | Machines parse structure, not prose |
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.
| Engine | Personality | Favors | Fastest way in |
|---|---|---|---|
| ChatGPT | Authority-driven | Earned, third-party sources; almost never social. Swaps its entire source ecosystem by language | Top-tier earned coverage — from local-language publishers |
| Perplexity | Challenger's friend | The widest mix; YouTube is frequently its single most-cited source, plus retailer listings and community discussion | Good video reviews + accurate retailer listings |
| Gemini | Most brand-friendly | Brand-owned pages — deep, well-structured content on your own domain pays off. Saturates quickly on authority language | Rich owned content; don't over-claim |
| Claude | Conservative "brand guardian" | Established names; reacts to authority on an inverted-U curve — moderate credibility helps, aggressive over-claiming backfires as "too good to be true." Reuses trusted English sources across languages | Moderate, genuine authority; strong English sources travel well |
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:
| Industry | What wins | Exploitable edge |
|---|---|---|
| Electronics & industrial | Precise specifications — a numbers game | Detailed spec sheets |
| Skincare & experience goods | Brand reputation and credibility signals | Genuine certifications, real clinical evidence |
| Automotive | Review sites rule; cited content is often stale | Freshness — newer video and retailer content (Perplexity especially) |
| Finance | Convergence on a few evergreen review hubs | Getting cited there beats optimizing your own site |
| Local services | Fragmented; aggregator directories dominate | Needs its own playbook entirely |
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
- Write listings to match how real buyers describe their needs, in full sentences.
- Load up on verifiable evidence: specs, numbers, real quotes, real citations.
- Give the model something distinguishing: ratings, review counts, concrete advantages.
- Structure for machines: comparison tables, tight paragraphs, bolded value props, clean schema.
- Invest in earned media, per language — not just your own website.
- Specialize by engine: video and retailer presence for Perplexity, deep owned content for Gemini, top-tier authoritative coverage for ChatGPT and Claude.
- Measure across repeated queries and multiple engines — AI answers vary run to run.
- 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.