How AI Assistants Choose Their Sources: RAG, Query Fan-Out and Citation Criteria

Quick quiz

How should you get mentioned by ChatGPT and other AI assistants?

Question 1 of 425%

What do you want to know or change?

Just see if AI assistants talk about my businessFollow it over time against my competitorsBe mentioned more often, starting now

Understanding how AI assistants choose their sources is the starting point of any work on AI visibility. The good news is that the vendors document more than most people think. Google states in its documentation on AI features that AI Overviews and AI Mode may issue multiple related searches across subtopics, and that a page only needs to be indexed and eligible for a snippet to be shown as a supporting link. OpenAI, Perplexity, Anthropic and Microsoft publish similar rules for their own assistants. This guide gathers what is official, separates it from guesswork, and adds what we measured on 301,025 AI answers.

What happens between a question and a cited answer?

Almost every assistant with web access follows the same pattern, called retrieval-augmented generation (RAG): the model does not answer from memory alone, it first retrieves documents and then writes from them. The steps are the following.

1. Decide whether to search. A stable fact (a definition, a math rule) is answered from training data. A question about prices, products, recent events or a named company triggers a search.
2. Rewrite and fan out. The model turns the user's sentence into one or several shorter queries, each covering a facet of the question. This is query fan-out.
3. Retrieve candidates. Each query is sent to a search index, which returns a list of pages. Only pages already in that index can be retrieved.
4. Read and filter passages. The model (or a filtering step) reads the results and keeps the passages that answer a sub-question. Pages compete passage by passage, which is why passage ranking matters more than the overall page.
5. Write and cite. The answer is composed from the retained passages, with links to the pages they came from.

Two consequences follow. A page that is not in the index used by the assistant never enters the race, however good it is. And a page can rank first on Google for the user's exact question yet lose the citation, because the assistant searched for something else.

What do OpenAI, Google, Perplexity, Anthropic and Microsoft document?

None of the five publishes a ranking formula. Each one does publish rules on access, indexing and citation, and these rules are the only firm ground.

OpenAI (ChatGPT)

OpenAI's documentation of its crawlers separates three agents. OAI-SearchBot builds the search index: sites that block it are not shown in ChatGPT search answers. GPTBot collects training data, and blocking it has no effect on search. ChatGPT-User visits a page on behalf of a user and does not decide what appears in search. Our guide to ranking in ChatGPT details the rest.

Google (AI Overviews and AI Mode)

Google is the most explicit: AI features use the same index and the same quality systems as classic Search, they may run a fan-out of related searches, and they can show a wider set of supporting links than ten blue links. There is no AI-specific markup to add. In other words, Google SEO remains the entry ticket for Google's AI answers.

Perplexity

Perplexity's bot documentation states that PerplexityBot exists to surface and link websites in its results and is not used to train foundation models. Perplexity-User fetches pages in real time for a user's question. Allowing PerplexityBot in robots.txt is the condition to be in its index.

Anthropic (Claude)

Anthropic's web search documentation describes the loop precisely: Claude determines when to search based on the prompt, the search can repeat several times within one request, and the final answer always carries citations, each quoting up to 150 characters of the source. Claude searches for current prices, recent events and information about specific organizations or products, and answers stable knowledge without searching. Our guide on how to rank in Claude covers its crawlers and index.

Microsoft (Copilot)

Copilot is grounded in Bing. Since February 2026, Bing Webmaster Tools includes an AI Performance report that shows how often a site is cited in Copilot and the grounding queries that retrieved its pages. Microsoft's advice in the same announcement: keep content fresh (with IndexNow), structure it with clear headings, tables and FAQs, and support claims with data and sources. See our guide to ranking in Copilot, and the one on ranking in Grok for xAI's assistant, which documents much less.

Which criteria decide who gets cited?

Combining the official documentation with what can be observed in answers, six criteria explain most citations. The first two are documented requirements; the others are consistent patterns.

1. Access and indexing

The assistant's search bot must be allowed and the page must be in the index it queries: Google's for Gemini and AI Overviews, Bing's for Copilot, OpenAI's and Perplexity's own indexes, Brave for Claude. Check your rules with our guide to robots.txt for AI crawlers.

2. Match with the hidden sub-queries

Retrieval happens on the fan-out queries, not on the user's sentence. A page that answers "best CRM" in general can lose to a page that answers "CRM for a 5-person real estate agency, price per user", because that is what the assistant searched for.

3. Passages that can be quoted alone

Assistants cite passages. A section that opens with a direct answer, a definition, a table row or a numbered step can be lifted out of the page and still make sense. A vague introduction cannot.

4. Evidence and freshness

Claude searches precisely when information may have changed, and Microsoft recommends supporting claims with data and sources. Dated figures, named sources, prices "checked" on a given date and visible update dates make a page safer to cite.

5. Corroboration elsewhere

An assistant that finds the same fact on several independent sites can state it with confidence. This is entity consensus: reviews, comparisons, directories and press mentions that describe your brand the same way.

6. The type of source each model prefers

Each assistant has its own habits. Some lean on encyclopedic sources, others on video, forums or review platforms. The next section shows how large these gaps are.

What do our citation studies show?

Our platform citation study analyzed 301,025 answers from ChatGPT, Gemini, Perplexity, Claude and Grok, collected between August 2025 and August 2026 on a panel of buyer questions (79% asked in French). It counted 1,820,738 AI citations pointing to 131,349 distinct domains. Two findings matter for source selection.

First, assistants do not cite the same number of sources. Gemini names 11.1 sources per answer on average, Grok 7.1, ChatGPT 4.1, Claude 3.5 and Perplexity 3.3. On the last three, a question offers three or four slots, so being the second-best page often means not being cited at all.

Second, the preferred source types differ completely from one assistant to another. Wikipedia appears in 3.43% of ChatGPT answers and 0.03% of Claude answers; YouTube in 3.69% of Gemini answers and 0.04% of ChatGPT answers. Only 6 platforms are cited by all five models. Our ChatGPT citation study also shows that selection changes over time: ChatGPT moved from 5.8 to 4.1 sources per answer over the period, with a low of 1.7 in June 2026.

How do you become one of the sources AI assistants choose?

The selection process translates into a short list of actions, valid for Google and for the AIs alike.

• Open the doors: allow OAI-SearchBot, PerplexityBot, Claude-SearchBot, Claude-User, Googlebot and Bingbot; check that your CDN does not block them.
• Be indexed where it counts: Google Search Console and Bing Webmaster Tools, a clean sitemap, IndexNow for fast updates.
• Cover the sub-queries: one page per precise question, built from the searches assistants actually run.
• Write answer-first passages: a direct answer at the top of each section, tables for comparisons, one idea per section, as in our guide to optimizing content for AI search.
• Prove and date: sources, figures, authors, visible update dates.
• Earn corroboration: reviews and mentions on the platforms that the assistants of your buyers actually cite.

Answers vary from one run to the next, so source selection is best observed over weeks. Sorank's AI query fan-out tracker records the searches each assistant runs on your prompts and counts them over time, and AI mention tracking shows whether your brand is cited and which sources take the slot when it is not. AI assistants choose the pages they can reach, that match their hidden searches and that prove what they say: measure those three things on your own questions, and you know what to fix.