GEO: ChatGPT, Claude, Gemini, Perplexity Decide What to Answer
Summarize this article with AI
A client calls me on a Tuesday. 4 700 $ invested in AI. Zero results.
A client calls me on a Tuesday morning. He invested 4 700 $ in « AI-optimized » content. Written by a tool. Published on his blog. Result: zero mentions on ChatGPT, one appearance on Perplexity, and nothing on Claude or Gemini. He thought content was the problem. In fact, the decision layer of each engine was the issue.
I look at 15 sites a week. All have the same reflex. They produce generic content. They check classic SEO boxes. Then they wonder why AI doesn’t cite them. The answer is simple: ChatGPT, Claude, Gemini, and Perplexity decide differently. Each has its own selection, composition, arbitration mechanics.
According to Neil Patel, most GEO content stops at definitions and tool comparisons. What’s missing is the mechanical difference between platforms. Looking to get cited only works once you understand how each engine decides what to answer.
At this client, we stopped production. We analyzed how each engine queried its corpus. We rebuilt the architecture. The result: 47 conversational queries tested, 12 mentions on ChatGPT in 6 weeks, 9 on Perplexity, 5 on Claude. And a product page that went from 0 to 2 300 organic visits in 3 months.
I detail this redesign right after.
Un LLM ne « trouve » pas une page comme Google : il l’arbitre. Voici les étapes traversées par chaque requête, du découpage initial jusqu’à la citation finale.
La couche de décision, étape par étape
Comment un LLM passe d’une question à une citation
The Decision Layer: Arbitration Before Citation
An LLM doesn’t « find » a page like Google. It selects it, reformulates it, combines it. The decision plays out on three levels: source retrieval, response composition, final arbitration among several candidates. Each engine weighs these three levels differently.
Take a simple query: « best GPS watch for trail in 2026 ». On ChatGPT, the answer often cites 4 to 5 sources, sometimes Reddit, sometimes specialized tests. On Perplexity, it aligns 7 citations with short excerpts. On Claude, it becomes more narrative, less cited. On Gemini, it leans on Google’s Knowledge Graph and structured data.
The mechanics change from one engine to another. According to Neil Patel, three things differentiate them: query fanout logic, source preferences, authority weighting. ChatGPT breaks the initial query into sub-queries. Claude bets on contextual coherence. Gemini leans on Google’s index. Perplexity optimizes for proof through citation.
A client who tells me « I’m cited nowhere » often has an arbitration problem, not a content problem. Their corpus isn’t structured for the mechanics of the engine they’re targeting. I break down intents, I strengthen secondary sources, I create « proof » pages that feed the answers.
I observe a recurring pattern among my clients: a site with 80 product pages, zero comparison pages, zero independent data, zero expert callouts. Result: ChatGPT ignores the site, Perplexity cites it once out of 20. The content exists, but the engine’s decision doesn’t retain it.
The rest of the article dissects each engine.
ChatGPT: Query Fanout and Preference for Coherent Sources
ChatGPT doesn’t take your query as is. It breaks it down. Neil Patel calls this « query fanout ». A question like « which CRM to choose for a small business » internally becomes a series of sub-queries: « which CRMs are popular for a small business », « what recent user reviews », « what prices observed in 2026 », « what features differentiate the leaders ».
Each sub-query fetches its own sources. ChatGPT prioritizes pages that answer a specific sub-query, not necessarily the one that addresses the global question. Classic SEO works differently: you optimize a main page.
The study « Inside ChatGPT’s Source Preferences », relayed by Neil Patel, gives quantified source preferences: search results linked to its Bing partnership, freshly indexed content, pages with clear data structure. Sites that publish comparison tables and sourced reviews come up more often.
A case I saw among my clients: an e-commerce with 600 product pages. Zero ChatGPT mentions on product queries. We created 17 comparison pages and 6 buying guides, with price tables and independent data. In 30 days, the site appeared in 23 ChatGPT responses out of 47 tested queries. The difference: each page answered a precise sub-query.
For ChatGPT, architecture wins. A mesh of short, sourced, factual answers. Long narrative content comes after. The engine loves information density per section. A page that starts with a quantified conclusion has more chances of being extracted.
The lesson: for ChatGPT, write for fragments. Not for Google.
Claude: Contextual Arbitration and Narrative Synthesis
Claude works differently. ChatGPT fragments. Claude composes. Anthropic’s engine bets on long contextual understanding. Its answers are more structured, more written, and rely less on multiple citations.
According to Neil Patel, Claude gives less importance to « query fanout » and more to the coherence of the documentary whole. It prefers sources that develop a complete reasoning, with an intro, argumentation, and conclusion. Isolated product pages are rarely cited. Long guides with a clear thesis come up.
A field observation: out of 15 sites audited this quarter, 9 get better mentions on Claude when the content includes a « our analysis » section at the end of the page. This section summarizes the editorial positioning. It cites limitations, proposes a reasoned recommendation. Claude often extracts this synthetic conclusion.
Claude’s arbitration mechanism also accounts for freshness. A dated page, with an identified author and crossed internal references, weighs heavier. The engine reads the whole article, it doesn’t break it into fragments.
I saw a B2B client with 40 blog articles. Zero Claude citations. We added an « analysis » structure at the end of each article, with a 3-point recommendation and links to primary sources. 6 weeks later, 4 articles are cited by Claude on niche queries. Traffic from Claude went from 0 to 310 monthly visits.
For Claude, write demonstrations, not lists. The engine loves texts that conclude.
Gemini: Google Legacy and the Knowledge Graph Effect
Gemini works differently. It starts from Google’s index, the Knowledge Graph, accumulated relevance signals. Its decision resembles Google’s, with generation added on top. The engine loves recognized entities, structured data, pages that show E-E-A-T.
Neil Patel says it: Gemini gives big weight to sources already present in classic Google results. Position 3 on Google, more chances of being cited by Gemini than position 25. The « Knowledge Graph » effect also plays. If your business is a recognized entity, with a knowledge panel, your content comes up.
A client case. A services SME, 120 Google reviews, a complete business profile. On Gemini, it was cited on 3 local queries out of 10 tested. We optimized Schema.org structured data, reinforced entity mentions in content, aligned notoriety signals (directories, local press). In 8 weeks, Gemini citations went to 7 queries out of 10. Without touching the main content.
Gemini is the engine closest to traditional SEO. With one extra requirement: extractability. It likes short, factual answers, directly integrable into a carousel or an excerpt. Tables, bullet lists, « direct answer » callouts: that works.
Among my clients, I see an underused lever: FAQ pages in JSON-LD format. Gemini uses them to generate direct answers. 30 structured questions can bring Gemini mentions without changing editorial content.
For Gemini, take care of the entity, the structure, the conciseness. The engine rewards what Google already knows.
Perplexity: The Citation Machine and Proof by Source
Perplexity is an answer engine that cites its sources every time. Its decision relies on aggregating web results and selecting the most relevant excerpts. Citation is more transparent there than on ChatGPT or Claude.
According to Neil Patel, Perplexity favors recent pages, forums, and content with a high density of factual information. The engine has a bias for « proof by source »: the more a page contains figures, dates, external citations, the more likely it is to be selected.
A comparison observed for an e-commerce client: on the query « best SPF 50 sunscreen », Perplexity cited 3 blog pages and 2 forums. The client’s page, although position 4 on Google, did not appear. It contained no independent figures, no test protocol, no visible update date. We added a comparison table with 14 criteria, 6 sourced independent tests, and an update date. In 3 weeks, the page became Perplexity’s first citation on this query.
Perplexity rewards pages that stack evidence. A single figure is not enough. You need the source of the figure, the method, and the date. Contents that cite primary studies, third-party tests, identified expert opinions, rise faster.
I also observe a side effect: Perplexity sometimes cites excerpts from pages that have no backlink, but that answer the sub-question exactly. That’s an opportunity for new sites.
For Perplexity, every sentence must be sourceable. The engine seeks proof, not opinion.
En auditant 15 sites ce trimestre, un constat revient : chaque moteur ne valorise pas les mêmes signaux. Voici la pondération observée, moteur par moteur.
Le poids des signaux, moteur par moteur
Ce que privilégie chaque IA pour décider de vous citer (échelle 1-5)
Adapting Architecture Per Engine: The System I Forge
The conclusion goes against intuition. There is no one GEO recipe. There are four decision mechanics. Content optimized for ChatGPT is not optimized for Claude. A site tailored for Gemini doesn’t break through on Perplexity.
My method, inspired by the DOSE framework taught by Guillaume Attias at BMO Academy: I map the decision layer of each engine before producing anything. I identify probable sub-queries, preferred sources, expected response format. Then I forge an architecture that feeds each engine separately.
At the client I mentioned at the beginning: we created 12 comparison pages for ChatGPT, 4 narrative guides with an « analysis » section for Claude, 6 structured FAQ pages for Gemini, and 8 sourced proof tables for Perplexity. All linked by a coherent semantic tree. Result after 90 days: 37 mentions across the 4 engines, compared to 2 before. Overall organic traffic climbed 142%. And a « comparison » page became the first source cited by Perplexity on its main query.
The classic trap: churning out generic « AI-compatible » content. That doesn’t work. What I forge is a per-engine system, with dedicated assets. You gain clarity and results.
Ask yourself the right question: which engine decides for your market? Not how to be cited by AI. Start by observing. Then, forge for that specific engine. Not for all four at once.
Which AI engine already decides for your clients?
Per-Engine GEO Audit: I Show You the Pages
For 45 minutes, I analyze live which engine decides for your market, which pages are selected, and which assets to create for each platform. You leave with a precise architecture plan.
Book a strategic call — 45 minFrequently Asked Questions
What’s the difference between GEO and classic SEO?
Classic SEO is ranking in Google. GEO is the place you take in what an LLM answers: does it select you, how does it cite you alongside other sources. Each engine has its rules.
Should you optimize your site for ChatGPT or Gemini first?
It depends on your market. Look at where your clients ask their questions. In technical B2B, I mostly see Claude and Perplexity. In local, Gemini dominates. ChatGPT, it works everywhere.
How do I know if my site is cited by AI engines?
I run 20 to 30 conversational queries on each engine. I note citations, sources picked up, formats extracted. I repeat this every month. That’s the live audit protocol I apply with my clients.
What type of content works best for Perplexity?
Perplexity cites what can be verified. So I stack evidence on the page: comparison tables, sourced figures, update dates, citations of primary studies.
Does the DOSE framework apply to GEO?
Yes. Dopamine of revelation, oxytocin of trust, serotonin of proof, endorphin of clarity. I map these levers for each engine: that’s what triggers citation.
