What Is Query Fan-Out?

What does query fan-out mean in AI search?

Query fan-out is the generation of several related searches from a single user question, run concurrently so the model gathers more results before it answers. Google uses the term for AI Overviews and AI Mode and defines it in its Search Central documentation.

Google defines the term in two of its developer documents, and the two definitions describe the same mechanism from different angles.

"A set of concurrent, related queries generated by the model to request more information and fetch additional relevant search results to address the user's query."

AI Overviews and AI Mode "may use a 'query fan-out' technique — issuing multiple related searches across subtopics and data sources — to develop a response."

The first definition names what fan-out produces: queries written by the model, not by the user, and issued concurrently. The second names what they cover: subtopics and data sources. Read together, they describe a model that treats the user’s question as a brief and writes its own research plan before it searches.

A sub-query is one of those model-written searches. It is usually narrower than the question and aimed at one aspect of it — a product type, a method, a comparison, a precaution. Each sub-query returns its own results, which is why a page can appear in an AI Mode answer for a question whose words it never uses.

Google also states the effect. The same AI features document says the models “identify more supporting web pages, allowing us to display a wider and more diverse set of helpful links … than with a classic web search.” Fan-out is the reason that list is wider: more searches reach more pages.

Query fan-out in AI search is unrelated to fan-out in software messaging or digital electronics.

What does a fan-out look like for a real question?

Google's example turns "how to fix a lawn that's full of weeds" into 3 sub-queries: "best herbicides for lawns", "remove weeds without chemicals" and "how to prevent weeds in lawn". Each sub-query targets one subtopic the original question implies but does not state.

Laid out as a tree, Google’s published example shows how far the sub-queries travel from the original wording:

• "how to fix a lawn that's full of weeds" — the user's question
  • "best herbicides for lawns" — a product choice
  • "remove weeds without chemicals" — an alternative method
  • "how to prevent weeds in lawn" — the problem after the fix

None of the three repeats the question. The first turns “fix” into a purchase decision, the second anticipates a user who wants to avoid chemicals, and the third moves past the immediate problem to prevention. A complete answer to the original question draws on all three, so the pages retrieved for each can all be cited.

Two practical points follow. A page about lawn herbicides can be cited in an answer to a question about fixing a weedy lawn without ever ranking for that question. And a page that only restates the question — “how to fix a lawn full of weeds” in its title and little else — competes for the parent query only, while the answer is built from the children.

How does query fan-out work, step by step?

Query fan-out works in 5 steps: the model breaks the question into subtopics, writes a search for each, runs them in parallel, merges and reranks the results, and writes one answer from the merged set. The merged set holds more pages than one search returns.

1. Decompose the questionThe model reads the prompt and identifies the subtopics a full answer needs.
2. Write the sub-queriesIt generates one search per subtopic, in its own words.
3. Run the searches concurrentlyAll sub-queries go to the search index at once, each returning its own result set.
4. Merge and rerank the resultsThe result sets are combined into one list and scored against the question.
5. Synthesize the answerThe model writes a single response from the top passages and links the sources it used.

Google’s public statements confirm steps 1 to 3. Announcing AI Mode on 20 May 2025, Elizabeth Reid, Google’s Head of Search, wrote that it “uses our query fan-out technique, breaking down your question into subtopics and issuing a multitude of queries simultaneously.” The Gemini API documentation for grounding with Google Search describes the same sequence from the developer’s side: the model “analyzes the prompt and determines if a Google Search can improve the answer”, then, if needed, “automatically generates one or multiple search queries and executes them”, and finally “processes the search results, synthesizes the information, and formulates a response.”

The Gemini API wording adds a condition the AI Mode announcement leaves out. Fan-out is not guaranteed. The model decides whether a search helps at all, and the number of queries varies from one to many. A simple factual question may produce a single search; a comparison or a planning question produces several.

The scale can go much higher. The same Google post introduced Deep Search, a research mode inside AI Mode, which “can issue hundreds of searches, reason across disparate pieces of information, and create an expert-level fully-cited report in just minutes.” That is the same technique applied with a larger budget of searches.

How are the sub-query results merged?

Google does not publish how AI Mode merges the result sets, so the merge step is the least documented part of the mechanism. Research systems show one standard approach. RAG-Fusion, a 2024 retrieval method, fuses ranked lists with reciprocal rank fusion: each document earns a score from its rank in every list that contains it, higher ranks earn more, and the scores are added. A document that appears near the top of several sub-query lists outranks one that tops a single list.

Whatever method an engine uses, the merge rewards the same property. A passage relevant to several sub-queries has several chances to be retrieved and more evidence of relevance when the lists are combined. After the merge, a reranker scores the surviving passages against the original question, and only the highest-scoring ones reach the model.

Which AI search engines use query fan-out?

Google documents query fan-out in AI Mode, AI Overviews and Deep Search. AI Mode issues a multitude of simultaneous queries, and Deep Search issues hundreds. Gemini grounding through the Gemini API also generates one or more searches per prompt.

The table shows that fan-out is documented for Google’s products and for OpenAI’s developer tool, while Perplexity does not describe its query generation publicly. The OpenAI row describes the web search tool in OpenAI’s API, where each search is logged as a web_search_call action — search, open_page or find_in_page. It is evidence that OpenAI models run multi-search retrieval, not a description of how ChatGPT search works internally.

For site owners, the Google rows matter most, because AI Mode was launched with fan-out as its stated method. How that translates into page planning for one engine — mapping the sub-queries AI Mode is likely to write for a market — is the subject of AI Mode SEO.

Where does query fan-out come from in retrieval research?

Query fan-out descends from multi-query retrieval research such as RAG-Fusion (2024), which generates several queries from one prompt and fuses their ranked results with reciprocal rank fusion. The same paper reports answers drifting off-topic when generated queries lose relevance.

The parent idea is retrieval-augmented generation (RAG), introduced by Lewis and colleagues at NeurIPS 2020: a language model answers from passages that a retriever fetched for the question. Classic RAG runs one retrieval per question. Fan-out changes the retrieval half, running many.

RAG-Fusion, described by Zackary Rackauckas in arXiv 2402.03367 in January 2024, is a direct research analogue. It works by “generating multiple queries, reranking them with reciprocal scores and fusing the documents and scores.” The paper applied it to product questions at Infineon Technologies, and manual evaluations found its answers accurate and comprehensive because the generated queries looked at the original question from several perspectives.

It also named the risk. Answers “can occasionally diverge from the topic” when the generated queries are not relevant enough to the original question. The risk is structural: every sub-query the model writes is a guess about what the user needs, and a poor guess retrieves pages about a neighbouring question. Scoring the merged results against the original question, rather than against each sub-query alone, is the standard guard against that drift.

Neither paper describes Google’s production system, and Google has not published the model or method behind its fan-out. The research shows why the approach exists — one query rarely retrieves everything a complex answer needs — and where it breaks. For the full pipeline around this step, from query to citation, see how AI search works.

The mechanism runs inside the engine; the part a publisher controls is whether a page answers the sub-queries it produces.

How do you cover fan-out sub-queries on a page?

Cover fan-out sub-queries by answering each subtopic of the main question in its own clearly headed passage on the same page. Google states pages do not need to be cut into tiny pieces, so complete sections serve better than fragments.

The checklist targets the merge step. A section that fully answers “remove weeds without chemicals” can be retrieved for that sub-query, and a page with sections for all three lawn sub-queries has three chances to enter the merged list for the parent question.

Google’s AI optimization guide sets a limit on how far to take this. It states: “There’s no requirement to break your content into tiny pieces for AI to better understand it.” The guidance is about headings and complete passages, not about splitting an article into dozens of thin pages or one-sentence blocks. The same guide advises focusing on non-commodity content — material that says something the other retrieved pages do not — which is what gives a merged answer a reason to cite one page over another.

Whether answer-first sections are lifted more often than story-first ones is still an open question. Grow AI Search’s Experiment 04, which compares answer-first and story-first passages in AI Overviews and AI Mode, is planned; no results exist yet.

The same coverage principle sits inside the wider method for optimizing pages for AI search, which adds crawler access, index eligibility and measurement to the page-level work.

Can you see the sub-queries an engine generated?

Yes, partially: the Gemini API and OpenAI's web search tool both return the search queries the model executed. For Google Search itself, Search Console's Generative AI performance report shows a page's impressions in AI Overviews and AI Mode by page, country, device and date.

The developer APIs are the only direct window. A grounded Gemini API response includes the search queries the model ran, and an OpenAI web search call logs each search action. Running a market’s key questions through these APIs shows the kind of sub-queries a model writes for them. The queries come from the API models, not from AI Mode or ChatGPT search, so treat them as a sample of how models decompose a question rather than a copy of any engine’s live fan-out.

Google Search does not expose fan-out queries. The Generative AI performance report in Search Console, launched on 3 June 2026 and available to all sites from 31 August 2026, reports impressions in AI Overviews and AI Mode grouped by page, country, device and date. Its help page lists no query data, so it shows which pages were linked, not which sub-queries retrieved them. AI Mode and AI Overviews data also counts within the standard Performance report under the Web search type.

A workable routine combines the two sources. Run the ten or so questions that matter most to the business through a grounded API call each quarter, record the queries each one produces, and compare them with the headings on the pages meant to answer those questions. Subtopics that appear in the model’s queries but not on the page are the gaps the coverage checklist is meant to close. Then watch the Search Console report to see whether the pages that gained sections start to collect impressions in AI Overviews and AI Mode.