Is my website enough to be visible in AI?
Published on · updated on
No, except when the user names your company. Your website is necessary to be visible in AI: it is what they read when asked for your prices, your documentation, or your services, and it must be open to their bots. But on a market question that does not name you, AIs mostly choose third-party sources: media, administrations, publishers they already know. A website, even an excellent one, does not create by itself the reputation that drives an AI to search for you.
For what type of question? It all depends on the question. On a question that names your company, a clear and accessible website is often enough. On a market question, it is almost never enough. On a local search, other sources come into play, such as the business listing on Gemini.
What the website does, and what nothing else can replace
First, the website is a condition for access. Google indicates that, to be shown in its generative AI features, a page must be indexed and able to appear in Google Search with a snippet. OpenAI indicates that a website must allow its OAI-SearchBot bot to pass in order to appear in ChatGPT search responses. A closed or poorly indexed website is absent, regardless of its quality.
The website is then the natural source for questions that name you. On questions about troubleshooting professional software that named a brand, that brand's documentation or support accounted for 58% of ChatGPT's citations, 30% on Claude, and 36% on Gemini. When a customer asks an AI how to use your product or what your service costs, it is your website that it reads, and a clear page makes all the difference.
What the website cannot do alone
On a question that does not name anyone, the AI decides for itself where to search. On ChatGPT, 72 to 91% of search queries carry the name of a specific publisher according to our corpora: a company whose name it does not know is not searched for. On our troubleshooting questions asked without naming a brand, the administration accounted for 45% of ChatGPT's citations. The website of a company unknown to the model rarely lands in these responses.
The Muck Rack study published in May 2026, covering more than 25 million links cited by ChatGPT, Claude, and Gemini across seventeen sectors, attributes 13.7% of citations to content from the company targeted by the question: 86% go elsewhere, to the press, encyclopedias, institutions, or other companies' content. The brand's website is one source among others, and rarely the main one.
External studies point in the same direction. In 2025, researchers from the University of Toronto measured a systematic bias of AI engines in favor of independent media, at the expense of brand websites, on recommendation requests. And even when the website is found, it must be chosen: ChatGPT cites only about one in ten pages of what its search returns.
A limit that is often forgotten: the website's language
A website in French only answers a portion of the searches launched for questions that were nevertheless asked in French. On our questions from September 15, 2026, 95% of ChatGPT's responses included at least one search in English, 75% on Gemini, and 32% on Claude. The pages that answer those searches are written in English.
The visibility of a website in AI therefore also depends on the language in which they search, and not just the one in which your customers write. An English version of essential pages broadens the searches the website can answer, without guaranteeing that it will be cited.
What the AIs did with this question
When we asked this question to the three AIs, ChatGPT understood it as a question of technical access: its searches targeted the documentation of AI bots, such as GPTBot, OAI-SearchBot, Google-Extended, or ClaudeBot, and it mostly cited documentation from Google and OpenAI. Claude and Gemini mostly cited agency and SEO consulting websites, and Claude explicitly looked for web presence beyond the website: reviews, forums, social networks.
Both readings are correct and complement each other. The website must be accessible, that is ChatGPT's answer; it is not enough, that is the track followed by Claude. A page that answers this question must address both, otherwise it only answers half of the searches that AIs launch.
The weight of website construction in ChatGPT's choice
In ChatGPT, we measured what distinguishes the pages it cites from those it reads and then discards. For each page read, we know two things: its construction score, calculated by our test based on what the page contains, and its position in ChatGPT's search results. Both do better than chance at identifying the cited page, but not to the same extent: of the advantage they have over chance, construction accounts for about one-third, and the position in the results for about two-thirds. The position therefore weighs approximately twice as much. The result is consistent across our various corpora, between 26% and 36% for construction.
At the top of the results, construction barely differentiates the pages anymore. In the first third of the list, where seven out of ten ChatGPT citations are located, the construction score does barely better than chance, and on the first three results, it is no longer distinguishable from it. It matters mostly further down: a poorly constructed page, landing far down the results, is almost never rescued.
This third is an order of magnitude, not a limit. Well-constructed pages are also more often well-positioned, and what the two indicators have in common can be attributed to one or the other: by sharing it equally, construction weighs about 23%; by attributing it entirely to the site, 44%. The measurement only applies to ChatGPT, for pages it has already read, and for our construction score. In Claude, which mostly reads well-constructed pages, it yields nothing conclusive.
What happens before reading also matters. Nearly half of the pages cited by ChatGPT came from sites it had itself named in its search, and these sites were cited twice as often as the others. Some of these names were already in the question asked. Being known to the AI, and then being found by its search, does not depend solely on the site's construction.
What this changes for a business
Make your site the best answer to questions that name you: service pages, pricing, documentation, contact details, open to AI bots and written in clear paragraphs, in the languages they search. This is the part of the work that depends entirely on you.
For market questions, the site is not enough. Your business must be mentioned, for good reasons, by the sources that AIs consult in your field, and useful pages must answer your customers' questions where they go looking. Google points out that non-authentic mentions do not help: a useful mention is a deserved mention.
What we do not know
Our measurements by type of question cover professional software troubleshooting and market questions, in French and in English, via the models' API. We did not measure local searches, nor the effect of website improvement on citations.
Sources
- Google Search Central — "Optimizing your website for generative AI features on Google Search", .
- OpenAI — OpenAI crawlers, .
- Muck Rack — "What is AI Reading?" (Generative Pulse), p. 5, .
- Chen, Wang, Chen, Koudas (University of Toronto) — "Generative Engine Optimization: How to Dominate AI Search", .
- FaireDuBruit measurements, from September 4 to 16, 2026 — see "How we measured".