How do ChatGPT, Gemini and Perplexity choose their sources?

An AI search engine chooses its sources in two steps. First it looks for candidates, in its own index or through an external search engine, often with several searches at once. Then the model uses the pages that answer the question or a sub-question most clearly, and brands that are mentioned in many places come back more often.

This article is for business owners who want to be cited, not for people looking for a reference list for a paper. The question "why does ChatGPT name my competitor and not me?" is part of the work of an SEO agency today.

First find, then choose

Finding is technical. ChatGPT rewrites your question into one or more searches for its search providers, and after the first results it can send more precise searches ^[1]. For AI Overviews and AI Mode, Google calls this "query fan-out": several related searches on subtopics at the same time ^[2].

With ChatGPT you only qualify if you allow OAI-SearchBot, including in your host or CDN ^[1]. If you block it, you appear at most as a navigation link ^[3].

Choosing is about content and trust. OpenAI ranks on several factors meant to lead users to relevant, reliable information, with no guaranteed position ^[1]; for its AI features, Google relies on the same ranking and quality systems as for its regular results ^[4]. No provider discloses more than that, so anyone who sells you a secret algorithm knows more than the makers themselves.

Where each engine gets its candidates

So a robots.txt that only thinks of Googlebot, or a firewall that stops unknown bots, can keep you out of ChatGPT, Perplexity or Claude without you noticing.

Which engine matters most? In Belgium in August 2026, 78.17% of the traffic that chatbots send to websites came from ChatGPT, 11.13% from Gemini and 5.39% from Copilot ^[10], ^[11]. That is traffic, not users, but it shows where to look first.

What a Google position still tells you

A good Google position makes you a likely candidate, but AI looks more widely.

• AI Overviews. According to Ahrefs, 76% of the sources were in the Google top ten in July 2025, and only 37.9% in a measurement from March 2026. Ahrefs attributes this partly to a new model since January 2026 and to query fan-out ^[12].
• ChatGPT, Gemini and Copilot. Of the links these three assistants cited, only 12% were in the Google top ten for the same question; for Perplexity it was almost one in three ^[13].
• A Belgian test. Online Marketing Monkey asked ChatGPT search 100 questions. For 78% of the commercial questions, at least one cited site was in the Google top ten; for informational questions, 36% ^[14].

Because of query fan-out, what surrounds your main question counts too: price, pros and cons, alternatives. A page that answers such a sub-question clearly can be cited while it sits on page three for the main question.

In that Belgian test, 93% of the sources for commercial questions were the brand's own site, or a site mentioned in the answer. For informational questions, American niche blogs (27%), Wikipedia (23%) and sites from the Netherlands (20%) weighed heavily, and Belgian sites got 13% ^[14]. It was, however, a hundred questions at a single moment, with no mention of date, model or location.

What makes a page likely to be chosen

Sources, quotations and statistics. Researchers from Princeton and elsewhere measured that citing sources, adding quotations and adding statistics gave 30 to 40% more visibility in generated answers; keyword stuffing yielded little or nothing ^[15]. That was, however, a test set-up with models from 2023.

Mentions elsewhere. Across 75,000 brands, the number of mentions on the web correlated most strongly with visibility in AI Overviews (correlation 0.664), much more strongly than the number of backlinks (0.218) ^[16]. Correlation does not prove causation.

Cover sub-questions, without tricks. According to Google, creating pages for every variant of a question, mainly to steer AI answers, falls under its spam policies ^[4]. One page that answers the sub-questions is something else:

1. The answer at the top, in the first two sentences.
2. One question per page, with the sub-questions a customer asks along with it.
3. Figures with a source and date, and a visible date of the last update. OpenAI itself advises users to check when a source was updated ^[1].
4. Your town and region in the text. For local questions, ChatGPT uses an approximate location based on the IP address ^[1].

Why your competitor does get mentioned

The full step-by-step plan is in how to get found in ChatGPT. Whether an llms.txt file or extra structured data helps is covered in llms.txt and structured data.

How to measure which sources AI uses in your sector

1. Draw up 20 to 30 questions your customers really ask, in Dutch and French if you sell in both languages.
2. Ask each question several times in each engine. AI Overviews and AI Mode can use different models and therefore show different links ^[2]; measure them separately.
3. Note the cited domains. The five or ten that keep coming back are the places where you want to be mentioned.
4. Repeat monthly with the same list.
5. Track the traffic. ChatGPT adds utm_source=chatgpt.com to links to your site ^[17]; Cédric explains how to read UTM codes in UTM codes explained. For Copilot, Bing Webmaster Tools has offered a report in public preview since February 2026 ^[18].

Rather not start on your own? The free scan makes a baseline measurement of your visibility in ChatGPT, Gemini, Perplexity and Google AI Overviews.

You cannot buy a place among the sources. ChatGPT does show ads in Belgium, but they appear below the answer and advertisers cannot steer the answers ^[19]. How that works is covered in ChatGPT Ads in Belgium.