GEO by engine: how each AI builds its answer

The AI response is not a citation: it’s a construction

I track 15 sites a week. They all ask me the same question. « How do I get cited by ChatGPT or Perplexity? »

The question starts from a false assumption. Answer engines don’t cite. They build.

Neil Patel just published an engine-by-engine breakdown. His angle is clear. Most GEO content stops at citation mechanics. He goes further. He describes how ChatGPT, Claude, Gemini and Perplexity select, synthesize, then formulate their answers.

A client called me on a Tuesday morning. He had invested 8 000 $ in content. 800 references. Zero semantic structure. His organic traffic was capped at 4 000 sessions per month. The AIs cited him 0 times.

The content held up. The architecture didn’t.

8 000 $ invested in product pages. Not one question answered. 800 pages that say « here is what we sell. » None that say « here is what you need to know. »

I see this case every week. A site full of pages describing products. No page answering a question. No page providing a definition, a table, a comparison. Language models look for evidence. Not promises.

When I restructured that catalog around semantic cocoons, results came fast. 47 targeted queries went from 3 % to 22 % citation rate. No advertising. Just an architecture that speaks the language of answer construction.

Selecting is no longer enough. Formulating is where the brand wins.

Neil Patel’s breakdown lands at the right time. No generalities. He details what differs between the four tools. In my view, it’s the only useful read in GEO today.

Because you don’t need to be everywhere. You win by understanding where your content will be reformulated.

Your page is read by an answer engine like a hurried human. It looks for the shortest path. If your content forces it to sort, it moves on. If it forces it to connect, it reformulates something else. It’s that simple.

The rest of this article leaves backlinks and domain authority aside. It focuses on answer construction. And on what each engine expects from you.

Avant d’être cité, votre contenu doit traverser le pipeline de construction du moteur. Chaque IA enchaîne les mêmes trois temps : elle sélectionne, synthétise, puis formule la réponse visible.

The 3 stages of an answer: selection, synthesis, formulation

An answer engine works in three stages. Selection: it gathers a set of sources. Synthesis: it extracts the key facts. Formulation: it writes the visible answer.

Neil Patel puts it well: each tool weights these three stages differently. ChatGPT bets heavily on synthesis. Claude excels at long-form rewriting. Gemini favors selection within the Google ecosystem. Perplexity puts citation front and center.

I tested this on a 945-page site. Before, 0 pages served to formulate a direct answer. Every page served to sell. Result: the AIs reformulated competitors. After restructuring, 14 « answer » pages captured the majority of citations across 47 queries.

Why? Because those 14 pages respected the three stages. A factual title helps selection. A table at the top of the page helps synthesis. An assertive, ready-to-repeat sentence helps formulation.

Three points to check.

• Selection: is your page the most obvious source? Does the title answer the exact question?
• Synthesis: does your content give the facts directly? Are the data accessible without reading the whole thing?
• Formulation: can your opening sentence be picked up verbatim? Is it affirmative, factual, without superlatives?

Most company content fails on the third point. They write « Our innovative solution optimizes your performance. » No language model will repeat that sentence. It contains no fact. No data. No precision.

Conversely, « 47 queries achieved 22 % citation rate after restructuring » gets picked up easily. It contains verifiable information. That’s what engines formulate.

Something else I observe. Many sites optimize selection but neglect formulation. They polish title tags. They forget the first paragraph. A language model reads the first paragraph as a summary. If it’s vague, the answer will be vague.

Reading comfort matters. When cognitive load drops, the probability of being picked up rises. We’ll come back to this with the DOSE framework.

ChatGPT, Claude, Gemini, Perplexity: four different logics

Four engines, four ways of working.

Neil Patel decoded the mechanics of each.

These differences matter. They change how you structure your content.

On ChatGPT, brand recognition and short answers make the difference: blocks that are easy to copy. On Claude, bet on long, well-built resources with visible logic. On Gemini, stay consistent with Google’s Knowledge Graph and structure your data. On Perplexity, multiply fresh sources and source every claim.

An e-commerce client with 800 references saw Claude cite him more on technical PDF pages. Perplexity cited him on updated comparisons. ChatGPT, never. The reason? No page gave a short answer. We built a network of 12 FAQ pages. 6 weeks later, ChatGPT started picking him up.

Neil Patel says it: the right reflex is to understand the engine serving your audience, and build for it first. Chasing all four at once leads nowhere.

ChatGPT. It favors known sources. If your domain is young, appear on third-party sites. And place an answer sentence at the top of the page.

Claude. It reads long documents. White paper, complete guide, case study: everything works, provided you have explicit subheadings. Each subheading must be an assertive sentence.

Gemini. It relies on entities. Structured data, pages linked to the right Knowledge Graph entities. A summary table at the top of the page works wonders.

Perplexity. It loves to cite. Every number must have a visible source. Every claim must be sourced. Content with 7 sources performs better than content with 0.

4 engines, 4 editorial workstreams

Want to take action? Here’s what I deploy with my clients.

For ChatGPT. Write opening paragraphs of 2 to 3 sentences. Start with the answer. Then elaborate. If you write « What is a semantic cocoon? », your first sentence must be: « A semantic cocoon is an architecture of pages linked by a pillar keyword. » Nothing else. The formulation is ready.

For Claude. Produce a structured document over 2 000 words. Add a table of contents, factual subheadings, tables. Claude extracts better from organized content. I observed a pickup rate 2,4 times higher with this format for a B2B client.

For Gemini. Polish your semantic markup. Use bullet lists, comparison tables, definitions. Every important entity should appear on the page. Link your internal pages with explicit anchors.

For Perplexity. Date your content. Update the numbers. Show your sources. Articles without dates are ignored more often. A comparison updated 3 weeks ago outperformed content frozen for 2 years across 6 tracked niches.

A frequent trap: trying to optimize for all four engines on the same page. You dilute the message. In my view, one page per intent, with a primary orientation, is better. Internal linking does the rest.

For a SaaS client, we built 9 ChatGPT-oriented pages, 6 Claude-oriented pages, and 4 Perplexity-oriented pages. Overall citations rose 14 % in 9 weeks, without massive production.

Selection, synthesis, formulation. Three levers. Four engines. One rule: make the model’s job easier.

Cognitive load: the DOSE angle that makes the difference

The DOSE framework, which I learned with Guillaume Attias at BMO Academy, reveals a point often overlooked in GEO. Clarity isn’t only for SEO. It acts on the reader’s chemistry. And on the model’s.

DOSE stands for four neurochemical levers: dopamine, oxytocin, serotonin, endorphin. Here, the dominant lever is serotonin. Serotonin rewards comfort. Clear, structured, unambiguous content creates that comfort. The language model, for its part, seeks the shortest path. Less cognitive load, higher probability of being picked up.

When your page is confusing, it increases cognitive load. The engine has to sort. Connect. Interpret. Result: it reformulates less, or it reformulates something else. When your page is crystal clear, it almost copies.

An example. A service page written for a B2B site began with « For 15 years, we have supported companies in their digital transformation. » 0 citations in GEO. We rewrote the first sentence: « An SEO agency structures your content to be picked up by AIs. » 6 weeks later, Perplexity cited the page on 12 queries.

The change was neurochemical. Not semantic. The clear sentence reduces load. It signals an answer. The other signaled a brochure.

Another field observation. Tables at the top of the page earn more citations than long introductory texts. They structure information. A table with 6 rows, 3 columns, with precise numbers. That’s immediate comfort.

DOSE is an editorial design tool, not a gadget. Before writing, ask yourself: « Does this sentence reduce effort? » If the answer is no, rewrite it.

Neil Patel doesn’t talk about neurochemistry. His pages tell the same story though. Every engine favors the synthesis that requires the least interpretation from it.

Is your content a comfortable armchair or a cold waiting room?

Taking back control: where to start

You don’t need to redo everything. Start with an audit of pages that should be answers.

Check your top 10 queries. For each one, find the page that answers. If the answer isn’t in the first paragraph, that’s your first project.

Then, test the three stages: selection, synthesis, formulation. Is the page selectable? Is the content synthesizable? Is the formulation repeatable?

Finally, choose a priority engine. ChatGPT, Claude, Gemini or Perplexity. The one that matches your audience. And structure for it.

An e-commerce client started with 4 000 organic sessions per month. After 5 months of restructuring, with no advertising, organic traffic rose to 11 600 sessions. AI citations represented 0 % at the start of the project. They now represent 18 % of traffic sources.

The lever was clarity. Not content volume.

Answer engines read your site. They classify it, then they reformulate what they understand.

The method is 5 steps.

• Identify the 20 questions your customers ask most.
• Map each question to an existing or new page.
• Rewrite the first paragraph with a factual answer.
• Add a summary table in the visible area.
• Repeat the process for the engine that matches your market.

At each step, cognitive load drops and the probability of being picked up rises.

Now, is your content designed to be reformulated?

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