What is query fan-out? How AI search splits your question

Query fan-out (also written query fan out) is the technique an AI search tool uses to take one question, split it into several related searches, run them, and combine what it finds into one written answer. Google uses the term for AI Overviews and AI Mode. OpenAI and Anthropic describe similar behaviour in ChatGPT and Claude without using the name. For a business it means your page can be cited because it answers one small part of a bigger question well, and that a ranking for a single phrase tells you less than it used to.

What query fan-out means

Google's documentation for site owners is the clearest starting point. It says both AI Overviews and AI Mode "may use a 'query fan-out' technique" and describes that technique as "issuing multiple related searches across subtopics and data sources" to develop a response.

Picture a researcher handed one broad question. Rather than typing the whole thing into a search box once, they write a list of narrower questions, look each one up and summarise what they found. That list is the fan-out. An AI system does the same in seconds, and the links under its answer come from those narrower searches as much as from the words you typed.

Google used the phrase when it introduced AI Mode as a Labs experiment in March 2025 and repeated it at I/O in May 2025 and at the Australian launch in October 2025.

How query fan-out works, step by step

The platforms publish only an outline. The steps below draw on Google's announcements, OpenAI's help centre, Anthropic's developer documentation and a pending Google patent application ("Search with stateful chat", published August 2024), which describes what a company might build rather than what is live.

1. Read the question and its context. OpenAI says ChatGPT may use your approximate location and, if memory is on, saved memories when it rewrites a search; the patent application adds time of day and earlier turns in the conversation.
2. Generate extra searches. The patent application describes a language model producing alternative and supplemental queries, rewrites of the original question and "drill down" queries. Google's public wording is simpler: AI Mode breaks your question into subtopics and runs many searches at once on your behalf.
3. Run them, often at the same time. Google says the related searches run concurrently and draw on web content plus sources such as the Knowledge Graph and shopping data for billions of products.
4. Review the results, and sometimes search again. OpenAI's example is a researcher asking about drugs that target CCR8 for cancer: ChatGPT might first send a search partner "CCR8 immunotherapy drug development 2025", then follow up with "CHS-114 conference 2025". Anthropic says Claude can search several times within a single request.
5. Pick supporting pages and write one answer. Google says that while a response is being generated its models identify more supporting pages, which lets it show a broader, more varied set of links than a classic web search.

Which AI search tools use it

Product names change quickly, so this list reflects what each company had published as of late 2026.

• Google AI Overviews and AI Mode. Named explicitly by Google. AI Mode began rolling out to everyone in the US on 20 May 2025 and reached Australia in English on 8 October 2025, the day after Google announced an expansion taking it to more than 200 countries and territories in total. AI Overviews appear only when Google's systems judge them additive, so they often do not trigger.
• Google Deep Search. Announced for AI Mode in May 2025 as the same fan-out technique "taken to the next level", able to issue hundreds of searches and produce a fully cited report in minutes.
• ChatGPT search. OpenAI does not use the words query fan-out. Its help centre says that when ChatGPT uses search partners it usually rewrites the question into one or more narrower searches and may send more specific follow-ups after reviewing the results.
• ChatGPT deep research. Launched in February 2025 as an agent that finds, analyses and combines hundreds of online sources into a report, changing direction as it goes.
• Perplexity Deep Research. Perplexity says this mode runs dozens of searches across hundreds of sources and refines its research plan as it learns.
• Claude. Anthropic's developer documentation says Claude decides when to search based on the prompt, that searching can repeat several times in one request, and that it answers without searching when a question draws on stable knowledge such as established facts.

Query fan-out versus classic keyword search

In classic search, one query produces one results page and one ranked list. You could track your position for a phrase and know roughly what a searcher saw.

With fan-out, the typed question is only the starting point. The searches that decide which pages get used are the hidden ones the system generates. Your page can be cited because it answers one of those narrower searches well, even if it would never rank for the original wording, and ranking first for the typed phrase does not guarantee a place in the answer.

• Questions are longer. Google says early AI Mode testers asked queries nearly three times the length of traditional searches. A longer question carries more angles, and each angle can become its own search.
• Answers are personal. OpenAI's example is "good restaurants near me" becoming "top restaurants San Francisco", or "good vegan restaurants San Francisco" if you have told ChatGPT you are vegan, so two people typing the same words can see different sources.
• Some prompts never search. Anthropic lists cases where Claude answers from what it already knows, such as established facts, and OpenAI says ChatGPT may search the web automatically when a question "would benefit from current information". If your topic is stable background knowledge, an assistant may never go to the web for it, and an answer built without a search cannot cite your page.

The kinds of searches an AI fans out to

Google, OpenAI, Perplexity and Anthropic do not publish the fan-out queries for your topic. The examples below are illustrations, built from the query types the primary sources describe.

Suppose someone in Australia asks an assistant: "What is the best accounting software for a small trades business, and is it worth paying for?" A fan-out might generate searches like these:

• Rewrites of the question: best accounting software small business Australia.
• Comparisons: Xero vs MYOB vs QuickBooks for tradies.
• Specific features: accounting software BAS lodgement; job costing for trades.
• Price and value: Xero pricing Australia; cheapest accounting software with payroll.
• Trust: accounting software reviews small business Australia.
• Next steps: how to switch accounting software; set up accounting software as a sole trader.

Each type has a basis in the sources: the patent application names rewritten versions, supplemental queries and "drill down" queries; OpenAI's examples show location being added and a more specific follow-up sent after the first results; and Google says AI Mode suits "comparing detailed options", which makes comparison searches a safe assumption for any buying question.

Why query fan-out matters for your business

Visibility in an AI answer is decided sub-question by sub-question. The answer is assembled from pages that satisfied the hidden searches, so a site that covers a topic in depth, including the follow-up questions a buyer would ask, has more chances to be picked than a site with one page tuned to one phrase. Google frames this as an opening for more types of sites to appear.

Facts about your business need to be findable too, because a fan-out may search for your opening hours, service area, prices or whether you handle a particular job. Google specifically tells site owners to keep Merchant Center and Business Profile information up to date for AI features.

Personalisation cuts the other way. Because location and memory change the generated searches, a rank-tracking report for one phrase describes one version of one results page, and says little about what a buyer in Parramatta with a history of asking about commercial fit-outs will see.

There is plenty of room to improve: in our audit of 719 B2B companies the average AI visibility score was 50 out of 100 and only 11% scored 70 or above (full breakdown in State of B2B AI Visibility 2026). Making a business findable, understandable and quotable in these answers is what AI SEO services cover, and it builds on ordinary SEO.

How to optimise for query fan-out

Google says no special optimisation is required. Its bar for appearing as a link in AI Overviews or AI Mode is the same as for ordinary results: the page has to be indexed and allowed to show a snippet. In Google's words there are "no additional technical requirements" and "specific optimization isn't required", though the usual SEO fundamentals still apply. What changes is the emphasis.

1. Let the crawlers in. Google asks site owners to make sure robots.txt, their CDN and their hosting all let its crawler reach the page. OpenAI says that to be eligible for ChatGPT search a site must allow OAI-SearchBot and its host or CDN must accept traffic from OpenAI's published searchbot IP addresses.
2. Map the questions behind each core topic. For every service or product, list the follow-up questions a buyer asks (cost, comparison, process, timeframes, risks, local availability) and check your site has a clear, self-contained answer to each. A spreadsheet is enough.
3. Write answers that stand alone. A fan-out search arrives looking for one specific thing. Put the direct answer near the top of the relevant section, keep it in text rather than only in an image or video, make sure structured data matches the visible text, and link the page to its neighbours so the cluster is easy to find.
4. Choose depth over duplication. Google's guidance is "unique, non-commodity content" for people who are asking longer, more specific questions and follow-ups. One thorough page with well-labelled sections can answer many related sub-questions.
5. Keep business facts current. Update Business Profile and Merchant Center, and make sure prices, service areas and hours on your site match what appears elsewhere.
6. Check your preview controls. nosnippet, data-nosnippet, max-snippet and noindex limit what Google shows from a page, including in AI features, and Google warns that tighter settings limit how your content can be featured. A max-snippet value set years ago can cap you today.

For the fuller method, the GEO Playbook covers how a generative engine picks its sources, pages that get quoted, crawler access and a prompt tracker with a 90-day plan.

Common mistakes and myths

• Myth: you need a separate AI setup. Google requires no special file, schema or page type for AI features; the ordinary eligibility rules apply.
• Mistake: a thin page for every imagined sub-query. Fifty guessed queries answered by fifty 200-word pages is commodity content, the opposite of what Google says it wants. Group sub-questions into substantial pages.
• Mistake: blocking the bots that do the searching. A CDN bot-protection rule or a robots.txt line written to stop scrapers can also stop Googlebot or OAI-SearchBot. Check access before you blame the content.
• Mistake: judging success only by clicks. Google says clicks from pages with AI Overviews are higher quality and suggests watching conversions and engagement as well, not clicks alone.
• Myth: you can see or control the fan-out queries. You cannot control the fan-out, and you can only partly see it. Google, OpenAI, Perplexity and Anthropic do not give site owners a list of the fan-out queries behind their answers, and the examples in their documentation are illustrations. Microsoft's Bing Webmaster Tools shows a sample of the grounding queries its AI used when it cited your pages in Copilot and Bing's AI summaries. Third-party tools either record the searches from prompts they run themselves or model them, so either way you see a sample, not your buyers' actual searches. Nor can anyone guarantee a citation: OpenAI says placement in ChatGPT search results "is not guaranteed", and each assistant picks its own sources.

How to measure your visibility in fan-out answers

Google. Traffic from AI Overviews and AI Mode is counted in Search Console's Performance report under the Web search type. Since June 2026 there has also been a separate generative AI performance report, available to all websites worldwide as of 31 August 2026, showing impressions in AI Overviews and AI Mode by page, country, device and date. Queries are not a dimension, so it will not show which fan-out searches surfaced a page, and Search Labs experiments are excluded. If the report is missing, the usual cause is too few impressions.

Microsoft. Bing Webmaster Tools has an AI Performance report, in public preview since February 2026, that shows citations in Microsoft Copilot and Bing's AI summaries, which pages were cited, and a sample of the grounding queries its AI used to retrieve them.

Your analytics. Google's own suggestion for valuing AI traffic is to look at conversions and time on site rather than clicks alone. Segment visits from AI assistants so they can be counted, and check the tracking before you trust a trend. In October 2026 a widely seen ChatGPT referral drop in GA4 was traced largely to inline links that had lost their utm_source tag, so those visits were landing in Direct.

ChatGPT, Perplexity and Claude. As far as their public documentation shows, none of these gives site owners an impressions report like Search Console's. The practical substitute is a fixed set of real customer questions, asked on a schedule, recording whether you are cited, named or recommended and who is cited instead. That is what tracking AI citations and share of voice means in practice, and it is the baseline we set before any AI SEO work starts.

How to get started

A small business can do the first pass in a week without new tools.

1. Pick your three most valuable services or products.
2. For each, write down ten questions a buyer asks before and after they contact you, including price, comparison, process and "is it worth it" questions.
3. Search your own site for each question and mark the ones with no clear answer, a buried answer, or an answer only in a PDF or image.
4. Fix the gaps on existing pages first: a direct answer at the top of the relevant section, and a short FAQ where several questions cluster.
5. Confirm Googlebot and OAI-SearchBot are allowed in robots.txt and at your CDN, and that your Business Profile details match your site.
6. Ask the same thirty questions in Google AI Mode, ChatGPT and one other assistant. Record who is cited. Repeat monthly.

If the acronyms are new, our guide to AI SEO explains the wider field, and our guide to generative engine optimisation explains where GEO fits with SEO and AEO.

Frequently asked questions

Is query fan-out the same as query expansion?

They are related. Query expansion adds synonyms and related terms to a single search to improve one results list. Fan-out generates separate searches, runs each one and merges what they return into a generated answer with its own links.

Does ChatGPT use query fan-out?

OpenAI does not use the term. Its documentation describes the same pattern: rewriting a prompt into one or more targeted queries, sending more specific follow-ups after reviewing the first results, and adjusting queries for location and memory.

How many searches does a fan-out run?

It varies, and the platforms publish only ranges. Anthropic's developer documentation says simple factual questions typically use one to three searches and comparative or multi-entity research can use ten or more. Research modes are the same idea at a bigger scale: dozens of searches for Perplexity's Deep Research, hundreds for Google's Deep Search. Treat any precise average you see quoted as an outside estimate.

Can I keep my content out of Google's AI answers?

Yes. Google's Search generative AI control in Search Console (Settings > Search generative AI), rolled out to all websites worldwide as of 31 August 2026, lets you exclude your site's links and content from AI Overviews, AI Mode and generative AI features in Discover. Excluded sites get no traffic or impressions from those features, and Google says the setting does not affect ranking elsewhere in Search. Short of that, Google says robots.txt rules for Googlebot are the crawl control for Search, and that nosnippet, data-nosnippet, max-snippet and noindex limit what is displayed, including in AI features. Google-Extended is a separate control for AI training and grounding in some of Google's other systems.

The practical next step is the question list: thirty real buyer questions, checked against your site this week, then tracked monthly with the playbook's prompt tracker.