In many AI systems, query fan-out happens behind the scenes: people asking a question usually see only the finished answer, not the subqueries the system ran to produce it. To “fan out” means to spread out from a single point. Google uses the term for its AI features AI Overviews and AI Mode.

Query fan-out is part of the retrieval step behind an AI answer. When an AI system bases its answer on content it retrieves at the moment a question is asked, this is called retrieval-augmented generation (RAG). To retrieve anything, the system needs queries, and it writes them itself. The prompt—the input a person actually enters—is only the starting point.

How query fan-out works

Google describes the subqueries as concurrent, related queries that the large language model behind the AI system generates to get more information on a question. Its example is the question “how to fix a lawn that’s full of weeds,” which might lead to subqueries such as “best herbicides for lawns,” “remove weeds without chemicals,” and “how to prevent weeds in lawn.”

For each subquery, the system retrieves matching content—at Google from its own web index and from other data sources as well. The language model then combines the results into one answer. According to Google, its models find additional supporting web pages along the way, so an answer can show a broad and diverse set of links. Fan-out isn’t limited to questions in text form: in April 2025, Google extended AI Mode to questions about photos. According to Google, AI Mode can understand the entire scene in an image, including which objects it shows and how they relate to one another. It then issues multiple queries about the image as a whole and about the individual objects.

A close relative: query rewriting

Query fan-out is closely related to query rewriting. In query rewriting, the AI system doesn’t pass a question on word for word but turns it into one or more targeted queries. A research paper on the technique, presented at the EMNLP conference in 2023, attributes the need for it to a gap between how a question is worded and the knowledge needed to answer it. The line between the two is fluid: rewriting is about a better-worded query, fan-out about several aspects retrieved in parallel.

Providers describe their approaches differently:

  • ChatGPT: According to OpenAI, when ChatGPT search draws on other providers such as Microsoft, ChatGPT typically rewrites the question into one or more targeted queries and, after reviewing the initial results, may send additional, more specific ones.
  • Gemini: In its documentation for developers, Google states that a Gemini model with access to Google Search generates and runs one or more queries itself when needed.
  • Claude: According to Anthropic’s documentation for developers, Claude can search the web several times in a row for a single answer.
  • Perplexity: According to Perplexity’s Help Center, Pro Search, its feature for more in-depth answers, runs multiple searches across the web and is particularly good at breaking down ambiguous or multifaceted questions.
  • Microsoft Copilot: For work accounts, Microsoft explains that Microsoft Copilot derives short queries of a few words from a question and sends them to Bing. In Microsoft’s example, a request to summarize a company’s financial information and business strategy becomes one query on the strategy and one on the financials.

What query fan-out means for your website

For visibility in AI answers, query fan-out has an important consequence: a page can be retrieved and cited as a source for a subquestion the person asking never raised. A broad question about a suitable provider, for example, can lead to subqueries about prices, use cases, or comparisons. Whether a page is retrieved can therefore depend on whether it answers one of those subquestions, not only on whether it matches the original question.

It makes sense, then, to use the relevant pages to answer the related questions your customers actually ask: comparisons, prices, use cases, and typical follow-up questions, each in clear, self-contained sections. How well a single section works on its own is a matter of citability. That covering such questions increases the chance of being retrieved is plausible but not proven. What’s more, there is no way to predict from the outside which subqueries an AI system will run for a given question.

Google, for its part, advises focusing on what your users want and not overdoing it. It might seem tempting to create separate content for every possible variation of a question, for example for questions other people have already asked or for the subqueries fan-out produces. According to Google, doing so primarily to manipulate generative AI responses in Google Search violates its scaled content abuse spam policy. A high number of pages doesn’t make a website higher quality or more relevant either, Google adds, and a site doesn’t need to cover every possible way a question might be phrased. For you, that means the yardstick is your customers’ questions and needs, answered on the relevant pages, not as many variations as possible.

Query fan-out also matters when measuring AI visibility: because the AI system writes the subqueries itself, they can differ for the same question—in ChatGPT, according to OpenAI, depending on factors such as approximate location or saved memories. This is one reason for answer variability in AI systems, which is why a single test question only shows a snapshot.