Query Fan-Out: How AI Search Splits One Question Into Many

Query fan-out is the step where an AI search engine takes your one question, rewrites it into several narrower searches, runs them (often at the same time), and writes one answer from everything those searches return. Google documents it for AI Mode and AI Overviews, and ChatGPT, Perplexity and Gemini do their own versions. It is why a page can be cited for a prompt it never ranked for, and why a page ranking first for the head term can be left out.

This guide explains the mechanism using what Google, OpenAI and Perplexity have actually published, shows what we saw when we captured real fan-out queries from the Gemini API on 18 October 2026, and gives you a repeatable method to map the sub-queries for any topic and turn them into one well-built page. Every number below is labelled with who measured it and how.

What is query fan-out in AI search?

Query fan-out is a retrieval technique: a language model expands one prompt into several related search queries, retrieves results for each, and grounds a single answer in the combined results. Google's own glossary defines it as "a set of concurrent, related queries generated by the model to request more information" in its Guide to Optimizing for Generative AI Features.

Google's example is useful because it is so ordinary. For the question "how to fix a lawn that's full of weeds", the fan-out queries might include "best herbicides for lawns", "remove weeds without chemicals" and "how to prevent weeds in lawn". None of those is what the user typed. Each one retrieves a different set of pages, and the final answer can cite passages from all three sets.

Two things follow. First, the unit of retrieval is the sub-query, so a page competes for the hidden searches, not only for the visible prompt. Second, fan-out sits on top of normal search. Google is explicit that AI Overviews and AI Mode are rooted in its core ranking systems and that a page "must be indexed and eligible to be shown in Google Search with a snippet" to appear as a supporting link (AI features and your website). Fan-out widens the net. It does not replace the index. If you are new to the wider discipline, our explainer on generative engine optimization covers the context.

How does Google AI Mode use query fan-out?

AI Mode uses a Gemini model to decide which subtopics a question needs, runs many Google searches for those subtopics at once, and assembles one answer with links. That is the official description, and it has been consistent since launch:

• March 2025, launch. Google said AI Mode "uses a 'query fan-out' technique, issuing multiple related searches concurrently across subtopics", running on a custom version of Gemini 2.0 (Google, 5 March 2025).
• May 2025, Google I/O. Head of Search Elizabeth Reid described AI Mode "breaking down your question into subtopics" and issuing "a multitude of queries simultaneously", which lets Search "dive deeper into the web than a traditional search". The same post introduced Deep Search, which "can issue hundreds of searches" to build a cited report (Google, 20 May 2025).
• July 2025, more detail. Robby Stein, VP of Product for Search, explained that for a Nashville trip the system "may think of a bunch of questions like great restaurants, great bars, things to do if you have kids" and then "start Googling basically". He added that AI Mode taps Google's real-time systems such as Google Finance and the Shopping Graph (Search Engine Journal).
• November 2025, Gemini 3. Google said query fan-out was "getting a major upgrade": Search can perform more searches and "find new content that it may have previously missed" (Google, 18 November 2025).

Google patents point the same way, with a caveat. The application "Search with stateful chat" (US20240289407A1) describes "generating one or more synthetic queries using the LLM output" and selecting result documents, and the granted "Thematic search" patent (US12158907B1) describes generating sub-themes and refining a first query with a second. A patent shows what Google has invented, not what is running in production, so treat these as plausible descriptions rather than a spec.

Do Google AI Overviews use query fan-out too?

Yes, sometimes. Google's documentation says both AI Overviews and AI Mode "may use a 'query fan-out' technique, issuing multiple related searches across subtopics and data sources". The word "may" matters: AI Overviews are built for speed on short queries, and the November 2025 post says harder questions in both features are routed to the larger model while simpler ones use faster models. In practice, expect lighter fan-out on a quick AI Overview and heavier fan-out in AI Mode and Deep Search. Google does not publish counts for either.

How do ChatGPT and Perplexity fan out a prompt?

ChatGPT rewrites your prompt into one or more search queries, sends them to its search providers, and may follow up with narrower queries after reading the first results. OpenAI's help article on ChatGPT search says it "typically rewrites your query into one or more targeted queries", gives a biotech example where a first query is followed by "additional, more specific queries", and notes that with Memory on, the rewrite can include saved details (a vegan user in San Francisco asking for restaurants becomes "good vegan restaurants San Francisco"). Personalization means two users can trigger different fan-outs for the same prompt.

Perplexity's Pro Search works through a question step by step and can initiate "follow-up searches that build on previous results". That is sequential fan-out: search, read, search again.

Third-party measurements fill in what the vendors do not publish:

• Volume per prompt. Peec AI analysed 5 million fan-outs from April 2026 and counted 2.1 per prompt on ChatGPT, 1.4 on Perplexity and 6.8 on Grok (Peec AI, May 2026). Profound ran 10,000 prompts for 14 days and saw roughly 1.4 to 2 searches per run on ChatGPT, Perplexity and Copilot.
• Longer, more specific queries. Across 20 million ChatGPT fan-outs in five countries, average length roughly doubled from about 6 words in October 2025 to about 12 in January 2026, while the number per prompt stayed between 2.3 and 2.8 (Peec AI, February 2026). ChatGPT is not searching more often; it is searching more precisely.
• Little repetition. In Profound's study, 91% of ChatGPT's query strings were unique across repeated runs, against 14% on Perplexity, which stays close to the user's wording.
• Not every prompt searches. DataForSEO found that 47.5% of 100,000 ChatGPT prompts triggered fan-out searches (DataForSEO, May 2026). The rest were answered from the model's own knowledge.

Why does ChatGPT search in English for non-English prompts?

Nobody outside OpenAI knows for sure, but the pattern is well measured. Peec found that in about 78% of non-English cases ChatGPT added English-language searches, and 43% of all fan-out steps for non-English prompts ran in English, from 94% for Turkish to 66% for Spanish (Peec AI). Their examples include German software prompts returning no German companies. OpenAI does not document how it picks the language. The practical implication for a brand outside the English-speaking world is concrete: your local-language pages are only part of the retrieval pool, so you also need to be described accurately in the English-language sources that rank for your category.

How many fan-out queries does one prompt generate?

Anywhere from zero to hundreds, and the published averages disagree because they measure different things. Here is how to read them:

• Seer Interactive, Gemini 3 API: 10.7 per prompt (range 3 to 28) across 501 prompts, up 78% from 6.01 on Gemini 2.5 (Seer, November 2025). Seer forced grounding on every prompt, so every prompt searched.
• Seer Interactive, repeated runs: 8.52 per prompt across 100 prompts run 13 times (Seer, January 2026).
• Peec AI and Profound, consumer engines: 1.4 to 2.1 per prompt on ChatGPT, Perplexity and Copilot, measured from tracked prompts.
• Deep research modes: Google says Deep Search can issue hundreds; Ahrefs reported one ChatGPT Deep Research task running 420 searches (Ahrefs, March 2026), a single anecdote rather than an average.

So the honest answer is: a normal chat answer runs a handful of searches, AI Mode runs more, and research modes run far more. What matters more than the count is the shape. The same Seer study found 95% of Gemini 3 fan-outs had zero monthly search volume, and in the repeated-run study only 8 of 11,029 unique fan-outs appeared in all 13 runs. Fan-out queries are long, specific and mostly invented on the spot, which is why a keyword tool will never show you most of them.

What we saw when we captured real fan-out queries

To see the mechanism first-hand rather than repeat other people's summaries, we ran our own small capture on 18 October 2026. We sent prompts to the Gemini API with the Google Search grounding tool switched on (not forced; the model decides whether to search) and logged every search query the model executed, as returned in the API response. We used three models: gemini-2.5-flash, gemini-3.1-pro-preview and gemini-3.8-flash.

The commercial prompt was: "Which project management tool is best for a 15-person remote marketing agency, and how much will it cost us per year?" We ran it 13 times (5, 3 and 5 runs per model). We also ran Google's own lawn-weeds example 6 times.

Three findings stood out:

1. Newer models skipped discovery. Gemini 2.5 Flash searched for "best project management tools for remote marketing agency" and then priced up to eight vendors. The two Gemini 3.x models mostly went straight to queries like "clickup pricing plans annual" for three or four vendors they already had in mind. In 7 of 8 Gemini 3.x runs, no search asked which tool was best. The shortlist came from the model, and search was used to verify prices.
2. Pricing was the one constant. All 13 runs searched for pricing, 9 of 13 added "annual", and 4 of 13 added a year (2025 or 2026). None of the 13 runs searched for reviews, integrations, client or guest access, or time tracking, even though those matter to an agency. Those needs were answered, if at all, from the model's own knowledge.
3. Some prompts do not search at all. For the lawn-weeds question, Gemini 3.8 Flash answered from its own knowledge in 3 of 3 runs. Gemini 2.5 Flash searched in 2 of 3 runs, with 5 and 6 queries such as "natural weed control methods for lawns" and "chemical weed killers for lawns", which closely mirror Google's documented example.

Limitations: this is 19 API calls on one day. The Gemini API with grounding is not the same system as Google AI Mode, which adds its own ranking, Shopping Graph and personalization layers. Treat it as a look at how Gemini models plan searches, not as a measurement of AI Mode. Even so, the lesson for a vendor is practical: if the model already holds a shortlist, you need to be on it (which comes from being described across the web, not from one page), and the query that does get run, such as "[your brand] pricing per user annual", needs to land on a crawlable, current pricing page.

Why is ranking first for the head term not enough for AI citations?

Because AI Mode retrieves for the sub-queries, not only for the prompt, and the pages that win the sub-queries are often not the pages that rank for the head term. seoClarity checked 1,000 transactional US queries in September 2025 and found that only 19% of 12,011 AI Mode citations came from the organic top 20, and only 28% of queries had at least one top-10 URL cited (seoClarity, October 2025). iPullRank's Mike King made the same observation: cited pages often did not rank in the top 100 for the original query, because they were retrieved by synthetic sub-queries (iPullRank).

AI Overviews are different, and it is worth being precise. Ahrefs found that 76.1% of AI Overview citations in its July 2025 sample of 1 million Overviews came from the top 10 (Ahrefs), and when the authors tested whether fan-out explained the rest, the data did not support it. So the summary is: for AI Overviews, ranking well is still the main lever; for AI Mode and chat engines, ranking for the head term gets you into one of several retrieval pools, and the other pools are decided by how well you answer each sub-question.

What does a fan-out map for one prompt look like?

A fan-out map lists the sub-questions a prompt implies, the type of each, and where your page answers it. Here is the map we built for the project management prompt above, marked against what the Gemini models actually searched in our 13 runs. It has 12 sub-queries, which is in line with the 8 to 11 that published studies typically report for a rich prompt.

Read the zeros carefully. They do not mean "skip these". They mean the model answered those needs from what it already knew, so the way to influence them is to be accurately described across the web (reviews, comparisons, documentation), not to wait for a search that may never run. The rows with high counts are where one page can win retrieval directly.

How can I find the fan-out queries for my topic?

Combine a manual map (to cover what models think about) with a sample of observed queries (to see what they actually search). This is the method we use:

1. Write the prompt as a buyer would ask it. Full sentence, with the constraints a real person adds (team size, budget, location, deadline). Short keywords produce short, unrepresentative fan-outs.
2. Break it into typed sub-questions. Use the types in the map above: discovery, constraint, verification, calculation, recency, comparison, entity, implicit need, validation. Aim for 8 to 12.
3. Add the questions real people ask. Pull People Also Ask and related searches for the head term and two or three sub-queries. These are not fan-outs, but they are the same questions in human words. AI Rank Lab's keyword planner returns PAA and related searches alongside volume, which speeds this step up.
4. Capture observed fan-outs. Run the prompt 5 or more times through a source that returns real queries: the Gemini API with Google Search grounding (the response lists the queries it executed, as documented in Google's grounding docs), or SE Ranking's free ChatGPT Fan-Out Query Extractor. Some practitioners also read ChatGPT's network responses in browser DevTools; that relies on undocumented fields OpenAI can change, and we did not verify it for this article.
5. Group by theme, not string. Collapse "clickup pricing plans annual" and "ClickUp pricing per user per month annual" into one theme: ClickUp annual pricing. Seer's repeated-run data shows exact strings rarely repeat, so count how often a theme appears across runs.
6. Mark each theme against your page. Answered well, answered thinly, or missing. The missing high-frequency themes are your outline.
7. Re-run quarterly. Models change. Gemini 3 fanned out 78% more than Gemini 2.5 in Seer's data, and in ours the newer models searched differently again.

Which tools show fan-out queries, and are they observed or predicted?

The most important question to ask about any fan-out tool is whether it shows queries an engine actually ran (observed) or queries an LLM thinks the engine might run (predicted). Both are useful, but only observed queries are evidence.

Disclosure: AI Rank Lab is our product. We list it first because this is our site, and we've kept the same criteria for every tool below.

We have not listed prices because most of these are bundled into wider plans that change often; check each vendor's pricing page on the day you buy (the SE Ranking tool was free when we checked on 18 October 2026).

How should I structure a page so each passage answers one sub-query?

Give each high-frequency sub-question its own clearly headed section that opens with a direct, self-contained answer, then supports it. Fan-out retrieval matches passages against many specific queries, so a section that answers one question cleanly in its first two sentences is easy to retrieve and quote. This is good writing for readers too, and Google is clear that you do not need anything beyond that: "There's no requirement to break your content into tiny pieces for AI to better understand it", and its systems "can understand synonyms", so there is no need to match fan-out wording word for word.

Use this outline template for one buyer question:

1. Opening answer (2-3 sentences). The direct answer to the parent prompt, quotable on its own.
2. Key takeaways box. Four or five lines a reader (or a model) can lift.
3. H2: the discovery sub-query ("Which X is best for Y?"), with a shortlist and the criteria behind it.
4. H2 or H3 per constraint (team size, budget, region, use case). Each opens with the answer for that constraint.
5. A table for verification sub-queries: prices, limits, specs, with the date you checked them.
6. A comparison table for the "A vs B" sub-queries.
7. H3s for implicit needs the buyer did not say but will care about.
8. An ordered list for the procedure, if the question implies doing something.
9. FAQ entries for the remaining short sub-questions.
10. Sources and methodology inline, so each claim carries its own evidence.

Our AEO checklist covers the technical side (indexability, snippet eligibility, page speed) that decides whether any of these passages can be retrieved at all. AI Rank Lab's Content Suite builds an outline like this from a prompt and competitor research, and its SEO, AEO and GEO audit checks the page structure once it is live.

Should I create a separate page for every fan-out query?

No. Google's guide says plainly that while "it might be tempting to create separate content for every possible variation of how people might search", doing so to manipulate rankings or AI responses violates its scaled content abuse policy, and that "a high quantity of pages doesn't make a website higher quality". Fan-out also works against the tactic: the strings change on almost every run, so a page per string chases a target that will not be there next time. One complete page per real buyer question, with a section per sub-question, is both the compliant and the effective approach. Separate pages make sense only when a sub-question is a genuinely different job (for example, a full pricing page or a dedicated integration guide).

Common query fan-out mistakes

• Building one page per sub-query. Scaled content abuse in Google's terms, and pointless given how unstable fan-out strings are.
• Optimizing for exact fan-out strings. Only 8 of 11,029 Gemini 3 fan-outs repeated across all 13 of Seer's runs, and 91% of ChatGPT strings were unique in Profound's study. Target themes.
• Trusting keyword volume to prioritise. 95% to 99% of fan-outs have zero recorded volume in Seer's data. Prioritise by how often a theme appears across runs and how close it is to a buying decision.
• Treating predicted fan-outs as evidence. A tool that generates likely sub-queries is a brainstorm, not a measurement. Validate with observed captures.
• Ignoring the model's prior. In our capture, newer models searched only for brands they already knew. Being described accurately on third-party reviews, comparisons and documentation is part of fan-out optimization.
• Letting the pricing page go stale or block crawlers. Verification queries such as "[brand] pricing annual" ran in every one of our 13 runs. If that page is outdated, hidden behind a form or rendered only by JavaScript, the model verifies against someone else's page.
• Forgetting the English pool. For non-English markets, Peec's data shows a large share of ChatGPT's searches run in English.
• Adding special AI markup or llms.txt for Google. Google says you do not need new machine-readable files, AI text files or special schema.org markup to appear in its AI features.

How do I measure whether fan-out optimization is working?

Measure citations and impressions for the parent prompt over repeated runs, not single snapshots. Two sources together give a fair picture:

• Google Search Console's Generative AI performance report. Launched on 3 June 2026 and rolled out to all sites by 31 August 2026, it shows impressions, pages, countries, devices and dates for AI Overviews, AI Mode and Discover AI features (Google Search Central). It does not list fan-out queries, so use it to see which pages gain AI impressions after you restructure them.
• Repeated prompt tracking across engines. Run the parent prompt on a schedule and record which domains and URLs are cited. Expect noise: across 968 repeated runs of 42 prompt-engine pairs tracked in AI Rank Lab (February to October 2026), the median overlap of cited domains between consecutive runs was 0.41 (Jaccard), and a median of just 3.5% of cited domains appeared in every run. With that much run-to-run variation, judge changes over weeks of runs, not one check. Our guide to the AI Overview checker explains how to set this up for Google, and the 30-day ChatGPT citation playbook covers the same loop for ChatGPT.

Be honest about attribution. No published controlled study isolates fan-out coverage as the cause of citation gains, and the Ahrefs AI Overview data did not support the fan-out explanation for out-of-top-10 citations. Fan-out mapping is a planning method that makes complete pages more likely; it is not a switch.

The short version

Query fan-out turns one prompt into several precise, changing searches. Google documents it, OpenAI documents it, and the measured counts depend on the engine, the model and the method. Map the sub-questions by type, check a sample of observed queries, group them into themes, and answer each theme in a clearly headed section of one complete page, with dated facts on the pages that verification queries land on. Then measure over repeated runs.

If you want to see which pages ChatGPT, Claude, Gemini, Perplexity, AI Overviews and AI Mode cite for your prompts, and turn the gaps into outlines, you can track your prompts in AI Rank Lab.