GEO by Engine: 4 Mechanics That Dictate AI Answers

Avant de plonger dans chaque mécanique, un chiffre qui change tout. Sur nos échantillons clients, 62 % des pages citées par un moteur IA ne le sont pas par un second. Voilà la fragmentation en un coup d’œil.

Why the Same Page Appears in ChatGPT and Not in Gemini

I look at 47 product pages. Four different assistants. Only one page stands out.

ChatGPT cites it as the first answer. Claude ignores it. Gemini suggests a competitor nobody referenced. Perplexity answers with a 2024 comparison.

There is a pattern here.

A mechanic is running.

An e-commerce site with 3,400 SKUs got 7 mentions on 10 tested queries in ChatGPT. The same catalog, the same queries, run through Gemini: 1 mention out of 10.

Across 50 sampled queries for three other clients, 62% of pages cited by one engine were not cited by a second. An order of magnitude. But telling.

The difference is not backlinks, nor the length of product descriptions. It’s the decision mechanic specific to each assistant.

Neil Patel’s article, GEO by Engine, finally describes it in detail. ChatGPT splits the query. Claude checks the document’s stability. Gemini leans on Google’s entity graph. Perplexity demands freshness and multiple sources.

Four engines. Four logics. Four different e-commerce levers.

Most GEO content treats these engines as a homogeneous block. Result: teams apply an average recipe. A page optimized for everyone, cited by none.

With my clients, the mismatch is identical. The pages that rank in ChatGPT are not the pages that rank in Perplexity. And rarely the ones that rank in Gemini.

The problem is the mechanic we ignore. Not the content.

The 4 real mechanics, detailed from Neil Patel’s diagrams and my deployments. For each mechanic: a case, a lever, a test.

No vague theory.

Action.

The classic approach is to map citations. Useful. But insufficient. A map tells you where you are cited. It does not tell you why the other engine ignores you.

The question that matters: « how does each engine decide? », not « how to be cited? ».

Once the mechanic is understood, action becomes mechanical. A test. A measurement. A decision.

Prenons un exemple concret. La requête « quel aspirateur pour poils d’animaux » ne reste pas d’un bloc : ChatGPT la décompose en sept sous-questions distinctes. La page qui en couvre le plus a le plus de chances d’être citée.

ChatGPT: The Query Splits into Sub-Questions. Your Page Must Answer Them All

ChatGPT’s mechanic relies on a fan of queries. Page ranking doesn’t interest it.

When a user asks « which robot vacuum for pet hair on carpet », ChatGPT breaks it down. Price. Suction power. Noise level. Filter maintenance. Compatibility with thick rugs. Availability. Warranty.

It sends these sub-questions to multiple sources. Pages that cover the highest number of sub-questions within a single piece of content have more chances of being cited.

This is what Neil Patel calls the query fanout.

The difficulty is coverage, not content.

A classic product page answers two questions: description and price. ChatGPT asks 7. The page that answers the 7 rises.

For a client specialized in air purifiers, 12 collection pages each covered one sub-question: noise, power consumption, coverage area, HEPA filter.

Initial result: 2 mentions out of 20 tested queries.

We restructured each product page by adding HTML tabs. A technical specifications tab. A parts compatibility tab. A maintenance tab. A shipping and returns tab.

Without changing the product. Without writing 3,000 words per page.

11 mentions out of the same 20 queries, 6 weeks later.

The e-commerce lever: turn each product page into a structured answer hub. Add a « questions buyers ask » section with 4 to 6 factual questions. Answer in one sentence. Let the engine aggregate.

This is why long, structured pages appear more often in ChatGPT. Length has nothing to do with it: the closed sub-questions are what count.

Product pages with accordion FAQs, buying guides with comparison tables, tech sheets with structured data: they cover the fanout.

Conversely, a product page that omits the noise level, even if the user didn’t explicitly ask, loses the sub-question. ChatGPT moves to the next source.

The query fanout is observable in the responses: ChatGPT often cites pages that answer peripheral aspects.

Test to launch: take 20 ChatGPT queries on your catalog. Note the number of mentions of your pages. Add a « compatibility » tab and a « maintenance » tab on 10 pages. Re-test 30 days later.

It all comes down to coverage.

Claude: Document Stability Counts More Than Domain Popularity

With Claude, the mechanic is different. It checks document stability.

A study cited by Neil Patel shows that editing by language models silently corrupts documents. Sentences shift, numbers change, dates disappear.

Claude favors stable documents. Those that haven’t been rewritten 10 times by AI without human control. Those whose version is explicit.

The content is good. Instability is the problem.

On a catalog of 7,800 product pages, an e-commerce client saw Claude cite only 5 of its pages on 40 tested queries. The pages were complete. But they were regenerated by an AI tool every quarter. Descriptions changed without any trace.

We froze 300 high-margin pages. Each page received a version number, a last-verified date, and a link to the unchanged PDF tech sheet.

Result: 19 Claude mentions on the 40 queries, 8 weeks later.

My takeaway: if your content changes without versioning, Claude treats it as unstable. It prefers to cite an older but stable document.

The e-commerce lever: publish PDF tech sheets, with dated versions. Add a block on the page: « Document version 3.1 — verified on May 12, 2025 ». Don’t let an AI tool rewrite without human validation.

Another case: a dietary supplement brand regenerated 1,200 descriptions per month with an AI tool. Claude cited it 2 times on 25 queries. After stopping automatic regeneration and adding a changelog, 9 citations on 25 in 5 weeks.

Why does Claude behave this way? When it retrieves a document, it first looks at integrity. A document that changes often comes across as less reliable, even if it’s optimized. Stability is a trust signal, not a marketing argument.

No need to rebuild your CMS. A simple HTML block with version is enough.

Three practices signal instability to Claude:

• Automatic regeneration without human validation
• Partial rewriting without a version number
• Silent modification of publication dates

Avoid them. Stability pays.

Test to launch: identify 20 product pages that underwent more than 3 rewrites in 6 months. Freeze them with a version number. Measure Claude mentions before and after 30 days.

Stability pays more than optimization.

Gemini: Google’s Entity Graph Dictates the Answer Before the Page Even Loads

Gemini’s mechanic relies on the Knowledge Graph and the entities recognized by Google.

Gemini starts from Google’s entity base. It queries brands, products, catégories, manufacturers, GTIN, reviews, prices. If your product is not a structured, connected entity, Gemini has nothing to say.

This is the most underestimated mechanic in GEO.

On 1,200 product pages from a hiking gear site, Gemini only mentioned 3 pages on 30 queries. The pages were indexed. But no structured Product, Offer, Brand, Review data was present.

We added complete Schema.org markup. Product, Offer, Brand, AggregateRating, SKU, GTIN. We synced data with Google Merchant Center. We harmonized the manufacturer name across all pages.

12 Gemini mentions on the 30 queries, 5 weeks later.

The e-commerce lever: if you sell a product, give Google a structured identity card. Not just a title and a price. A complete entity.

Another example: an auto parts site with 4,500 references. Gemini mentioned 4 products on 35 queries. After adding GTIN and OEM number in the markup, 14 mentions on 35 in 6 weeks.

The mechanism is simple: Gemini connects entities to each other. If your product shares the same manufacturer as three recognized brands, it enters the graph. Without that connection, it remains invisible.

To enter the entity graph, structure matters, not ranking. A small site with complete schema can be cited by Gemini more than a large site without schema.

Test to launch: take 30 Gemini queries on your catalog. Check in Google tools whether your products appear as entities. Add Product + Offer markup on 50 priority pages. Re-test 3 weeks later.

For Gemini, structured data is a decision signal. The rest is technical detail.

Perplexity: Freshness and Source Consensus Above All

Perplexity’s mechanic is brutal. It always answers with citations. Fresh sources. Multiple corroborating sources.

If your product page is 14 months old, with no visible update, Perplexity ignores it. It prefers a comparison published 3 days ago, even if less complete.

Here, freshness comes before domain authority.

On 15 tested queries about dietary supplements, a brand got only one Perplexity citation. The pages were rich. But the update date showed 2023.

We added a « Verified on June 18, 2025 » badge with a link to the independent lab report. We also updated the publication date, and cross-checked the information with two public studies.

8 Perplexity citations on the 15 queries, 4 weeks later.

For e-commerce: each product or collection page displays a recent last-verified date. Add independent external sources. Make visible updates, with a date that moves.

Another client, an online mattress store, received zero citations on 12 queries. After adding a « Tested by an independent lab — report updated July 3, 2025 » insert and a link to a sleep study, 6 citations on 12 in 3 weeks.

Why does Perplexity cite a comparison and not your product page? Because the comparison has a fresh date and multiple sources.

For Perplexity, consensus relies on three distinct sources confirming the same data point. Backlinks alone carry no weight. Perplexity cross-checks. If you are alone saying your mattress reduces pain, Perplexity ignores you.

Test to launch: take 15 Perplexity queries. Note the citations to your site. Update 10 pillar pages with a date and an external source. Re-test in 3 weeks.

La théorie, c’est bien. L’action, c’est mieux. Voici comment transformer les 4 mécaniques en un loop de tests concret, à lancer cette semaine, un moteur à la fois.

The 4 Tests by Engine: Turn Theory into an Action Loop

The DOSE framework, taught by Guillaume Attias (BMO Academy), turns reading into action. Dopamine first: a simple test, a visible result, an immediate reward.

4 tests to launch this week. One per engine.

One of my clients ran two of these tests in parallel: ChatGPT (FAQ tabs) and Perplexity (dates + sources). In 6 weeks, sessions from AI assistants went from 1,400 to 2,300 per month.

+64% AI organic traffic with blog content unchanged.

The loop is simple.

Test. Measure. Double what works.

For each test, the protocol is identical:

• Select 10 to 50 representative pages
• Apply the lever within 48 hours
• Re-test with the same queries 3 to 6 weeks later
• Log mentions before/after in a table

Dopamine rises when you see the first jump from 2 to 7 mentions. The brain records. You continue.

Guillaume Attias’s DOSE (BMO Academy) is an operational reward system.

In my opinion, the biggest mistake SEO teams make in 2026 is continuing to produce content for Google while assistants decide elsewhere.

Each engine has its decision mechanic. Once tested, you know where to invest the next hour.

What These 4 Mechanics Change for Your Semantic Cocoons

The semantic cocoons I’ve delivered since 2016: 1,300 in total. The logic remains the same. One pillar page, satellite pages, precise internal linking.

But in 2026, the cocoon must respond to four different readings.

ChatGPT reads the cocoon as a set of closed sub-questions. Claude, as a stable document. Gemini, as an entity graph. Perplexity, as a network of fresh, cross-checked sources.

A well-built cocoon meets these four conditions without contradiction.

In practice:

• The pillar page covers the major sub-questions, like the query fanout.
• Each satellite page keeps a version number and a stable verification date.
• Schema.org markup connects the pillar page to product, brand, and category entities.
• Each satellite page has at least one fresh, cross-checked external source.

A client in equestrian saddlery applied this logic across 27 pages of a cocoon. Before: 3 mentions from all AI engines combined on 40 queries. After 8 weeks: 22 mentions. And the 4 engines now cite different pages from the cocoon, not the same one.

Structure, not more content. A structure aligned with decision mechanics.

The DOSE framework, applied to cocoons, gives a priority order. Gemini first: add schema. Then ChatGPT: structure the FAQs. Then Claude: freeze versions. Finally Perplexity: add freshness.

The remaining question: does your current cocoon respond to a single mechanic or to all four?

What is the next query on which you test the stability of your documents before producing yet another article?

Stéphane Jambu

SEO & AI Engineer

I build growth systems / AI / Neuroscience | 650+ clients · 80 LinkedIn testimonials · 30 years of expertise · 15 years of systems running without me.

Follow on LinkedIn