How do you do generative engine optimization (GEO)?
Published on
Generative engine optimization, or GEO, is working to be cited in the answers of ChatGPT, Claude, or Gemini, and it starts with a measurement, not a recipe. You need to know which questions your customers ask the AIs, what the AIs search for and what they cite; write pages that carry the question in their title and answer it in developed paragraphs; then be present on the third-party sources they consult when the question does not name you. In our measurements, with Claude, a paragraph of 50 to 100 words is cited 3.7 to 4.8 times more often than a short paragraph; but a page the AI does not go looking for is never cited, however well written.
For what type of question? The word GEO has several meanings: it also refers to a magazine and, in marketing, to geographic targeting, and the AIs themselves get it wrong. This page is about Generative Engine Optimization, optimizing for AI answers. And it all depends on your customer’s question: if they name you, your site is the first lever; if they do not, third-party sources are.
Generative engine optimization (GEO) starts with a measurement
The first step is to ask the AIs your customers’ questions, as they phrase them, and to record three things: the searches the model writes itself, the pages those searches return, and the ones it cites. This information cannot be guessed. On our questions asked in French on September 15, 2026, ChatGPT ran ten searches per answer at the median, Gemini three, Claude two, and 95% of ChatGPT’s answers included at least one search in English.
You also need to tell contexts apart. When a question names a brand, the AI looks on that brand’s side: on troubleshooting questions about business software, the documentation of the brand asked about makes up 58% of ChatGPT’s citations. When it names no one, the AI picks third-party sources. The same company can be well cited in the first case and absent from the second; a measurement that mixes the two says nothing useful.
Writing the page: the question in the title, the answer at the start
The page is chosen on the user’s question. With ChatGPT, a page whose title repeats the terms of the question is cited 1.3 to 1.9 times more often depending on the corpus; with Claude, a URL that carries the question is cited 1.2 to 1.7 times more. An unambiguous title also protects against a misreading: on the question “Is what we read about GEO true?”, all three AIs first understood it as being about the magazine or an organization.
The answer comes next, in the first lines, with the words of the question. The rest of the page develops it: where the subject comes from, what recognized sources say, what more is known, what is not known. A thin page is discarded more often: ChatGPT discards pages under 300 words, and slow pages, more often.
Writing the passages: paragraphs that can be cited on their own
The cited passage is chosen on the AI’s words, not your customer’s. With Claude, whose API indicates the exact text cited, a paragraph that repeats at least a third of the words of the model’s search is cited 5 to 8 times more often than a paragraph that contains none of them, and a paragraph of 50 to 100 words 3.7 to 4.8 times more than a paragraph of 10 to 25 words. A paragraph that states, with a verb such as “is”, “allows”, or “must”, is cited 1.7 to 2.7 times more often.
These words go in the body, never in the title: with Claude, the terms the AI added in its searches lowered citations on our first corpora when they appeared in the title, and had no effect on the following ones. Each paragraph must also stand on its own, since the AI cites an excerpt, not the whole page. Google, for its part, writes that there is no need to break content into small pieces or to write in a particular way for AI: complete paragraphs, using the subject’s real vocabulary, serve both.
Being present where AIs search
This is the step that optimizing a page does not cover. ChatGPT writes the name of a specific source in 72 to 91% of its searches depending on our corpus, and cites that named publisher 1.8 to 2.6 times more often. AIs mostly cite third-party sources: according to Muck Rack, in May 2026, 13.7% of the citations of ChatGPT, Claude, and Gemini went to the content of the company targeted by the question. A page the AI does not go looking for is not read.
Doing GEO on your own site alone therefore leaves out the questions that do not name you. For those, you need well-built pages on the sources AIs consult in your field: this is what we call GDO, distributing content where AIs search. The two complement each other, since a third-party page must also be well written to be picked up.
Checking again over time
An AI answer changes from one time to the next. When we asked the same questions again about an hour later, only 25% at the median of the pages ChatGPT cited the first time were cited again. Presence is therefore measured as a frequency, over many questions asked several times and on several AIs, which do not read the same web: across our five corpora combined, only 1.8% of cited domains were cited by all three.
Freshness plays a small part: with Claude, a page less than six months old is cited 1.2 to 1.3 times more often than an older page. Updating a page when the information changes is useful; changing its date without changing its substance is not, and Google warns against the practice. The year in the title, often recommended, did not help ChatGPT on any of our corpora.
What you read about GEO that does not hold up
The most quoted figure, “up to 40% more visibility”, comes from the paper that launched the term in 2024: it was measured in a simulated engine where the sources are already supplied to the model. A benchmark presented at NeurIPS in 2025, C-SEO Bench, concludes that most conversational optimization methods are largely ineffective, often harmful. Schema.org markup, the question-and-answer format, and the llms.txt file have no proven effect on citations.
Before following a piece of GEO advice, it is worth asking which AI it applies to, which type of question, what was measured, and whether the measurement was repeated. Advice that answers none of these questions is an opinion.
What this changes for a business
Doing generative engine optimization (GEO) seriously is a loop: measure, write, distribute, measure again. The first two steps happen on your site; the third requires being published elsewhere, regularly, on the topics of your market.
This is FaireDuBruit’s method: every week, one article on your site and three on our specialist media, in your language and in English, written from your facts with the rules measured on this page. Our media do not publish independent opinions. We measure AI reads on these media, not yet the effect of these publications on their citations, and no citation is guaranteed.
What we do not know
Our measurements cover the API of GPT-5, Claude Sonnet 4.6, and Gemini 3.5 Flash with web search, not their apps or Google’s AI Overviews. They describe what distinguishes cited pages from discarded ones; we have not tested any optimization recipe on the same page, and the effect of a publication on citations is not measured.
Sources
- Aggarwal et al. — “GEO: Generative Engine Optimization”, KDD 2024, .
- Puerto et al. — “C-SEO Bench: Does Conversational SEO Work?”, NeurIPS 2025, .
- Anthropic — documentation for Claude's web search tool, .
- Muck Rack — “What is AI Reading?” (Generative Pulse), p. 5, .
- Google Search Central — “Optimizing your website for generative AI features on Google Search”, .
- FaireDuBruit measurements, from September 4 to 16, 2026 — see “How we measured”.