Query fan-out: how AI Mode builds answers and how to be included

Type a long question into Google's AI Mode and it does not run one search. Google says AI Mode and AI Overviews may use a technique called query fan-out: the system breaks your question into subtopics, runs several related searches at the same time and writes one answer from what it finds. That is why a page can be cited for a question it does not rank for, and why the page in position one can be left out.

This guide is for small-business owners, content leads and in-house marketers who keep hearing "query fan-out". By the end you will be able to explain it accurately, tell Google's guidance from the myths, map the sub-questions behind any customer question with free tools, decide whether a sub-question needs its own page, and check whether your changes worked.

What we say about how AI Mode works comes from Google's own documentation, linked where used. Where we go further, we label it as inference. Facts were checked in October 2026.

What query fan-out is, in Google's own words

Query fan-out is a technique in which an AI search system turns one question into a set of related searches, runs them at the same time and builds its answer from the combined results. Google uses the term for both AI Overviews and AI Mode.

Google's guide to optimising for generative AI features, published in May 2026, defines fan-out queries as a set of concurrent, related queries generated by the model to request more information and fetch additional relevant search results. Its example: someone asks how to fix a lawn that is full of weeds, and the fan-out queries might include "best herbicides for lawns", "remove weeds without chemicals" and "how to prevent weeds in lawn".

Google's AI features and your website page says AI Overviews and AI Mode may issue multiple related searches across subtopics and data sources to develop a response. While the answer is being generated, the models identify more supporting pages, so Google can show a wider and more diverse set of links than a classic results page would.

How far it goes

When Google announced AI Mode updates at I/O 2025, it said AI Mode breaks a question into subtopics and issues a multitude of queries simultaneously. Deep Search, the research option in AI Mode, takes the same technique further: Google says it can issue hundreds of searches and produce a fully cited report in minutes.

At I/O 2026 in May, Google said AI Mode had passed one billion monthly users and brought AI Overviews and AI Mode together into one AI Search experience, so a follow-up asked from an AI Overview continues in AI Mode. Google also reports that the average AI Mode query is about three times the length of a traditional search. Longer questions have more parts, and each part is a candidate for its own search.

What Google has not said

Google has not published how many searches a typical question triggers, which data sources it uses for each one, or how it chooses which retrieved pages to cite. Articles that describe those details rely on patent filings or outside tests. Patents show what Google has explored, not what runs in Search today, so treat any precise mechanism, including our own reading of the studies below, as inference.

Who can be included

The entry rules are ordinary. Google says a page must be indexed and eligible to show in Search with a snippet, with no additional technical requirements for AI Overviews or AI Mode. The May 2026 guide adds that your site must not be opted out of Search generative AI features in Search Console. If you are unsure what your robots.txt, CDN or that control allow today, our guide to blocking AI crawlers without leaving AI search covers each switch, and the technical SEO audit checklist covers indexing problems.

How other AI search tools split questions: a documentation check

Google is the only one of the four companies we checked that names "query fan-out" in the documentation we read, but Microsoft, OpenAI and Perplexity all describe answers built from several searches. Only Microsoft gives site owners a partial view of the searches its AI ran.

Method: in October 2026 we read each company's own documentation and help pages for its AI search products and recorded two things: whether it describes splitting one question into several searches, and whether a site owner can see those searches. These pages change often, so repeat the check before relying on it.

What the comparison tells you

1. The behaviour is common, so the method carries over. Mapping sub-questions helps with every engine, not just Google.
2. Nobody shows you Google's actual fan-out queries. A tool that lists "the" fan-out queries for your topic is making its own predictions.
3. Bing's grounding queries are the best first-party window. They are sampled, but they are real data about what an AI searched for when it used your pages.

Why fan-out changes which pages get cited

Because the AI searches for several sub-questions, the pool of pages it can cite is wider than the top results for the question you track. Studies since 2025 find that many cited pages do not rank in the top 10 for the original query, and that ranking for related questions goes with more citations.

The clearest sign is in Ahrefs' data. In an updated study reported by Search Engine Journal in early 2026, 38% of pages cited in AI Overviews also ranked in the top 10 for the same query, down from 76% in Ahrefs' July 2025 version. Ahrefs says better citation parsing explains part of the drop, and it points to query fan-out as a likely cause: pages that do well across related sub-queries can be cited without ranking for the original question. Google has not confirmed that explanation.

The same question can also surface different pages on different surfaces. Using September 2025 US data, Ahrefs found that AI Mode and AI Overviews cited the same URLs only 13% of the time, though their answers were 86% similar in meaning. That predates the May 2026 merger, but the lesson holds: the answer can stay the same while the sources change.

The evidence, side by side

Method: we collected the published studies that test fan-out or citation overlap directly and recorded the engine, sample, finding and main caveat for each. Two of them point in different directions, which is the useful part.

What we take from it

This is our reading of the studies together, not a finding any one of them states.

• Ranking for the head term is no longer enough. Pages that also rank for neighbouring questions get more chances to be retrieved, which fits both the Surfer correlation and Ahrefs' 38%.
• Covering every sub-question is not the lever. In the AirOps data, the coverage effect shrank to very little once the match between the page and the query was held constant. Clear, closely matched answers and good rankings mattered more.
• Every check is a snapshot. With 13% URL overlap between two Google surfaces and brand lists that rarely repeat, judge trends across many runs, never one answer.

For the wider picture of how AI answers choose sources across engines, including crawler access and reputation, see our guide to getting cited in AI search.

What fan-out does not mean: myths Google has addressed

Most fan-out tactics being sold today are covered by Google's May 2026 guide, and several are on its list of things you can ignore. Google's short version is that optimising for generative AI search is "still SEO", because its AI features are rooted in its core ranking and quality systems.

What the sceptics get right

An evidence review by Passionfruit, an SEO agency, accepts that the mechanism is real but argues the tactics built on it are overclaimed. It leans on the AirOps finding that fan-out coverage was nearly irrelevant to citation rates, with retrieval rank and heading-to-query similarity predicting citations better. We agree with the direction: fan-out widens the pool of candidate pages, but it rewards what good SEO always did, answering the real question and the questions around it clearly.

Map the sub-questions behind a topic with free tools

You cannot see Google's real fan-out queries, but you can predict them well enough to plan a page. Start from the full question a customer asks, collect the related questions Google and other engines already show, then sort and trim them into a list of distinct sub-questions.

1. Write the customer's question in full. Use the words from sales calls, enquiry emails, reviews and chat logs, not a two-word keyword. AI Mode questions are long, so your starting point should be too.
2. Ask AI Mode and note what it covers. Run the question three times in a signed-out or private window and write down the subtopics, headings and cited pages. Subtopics that appear in every run are the ones to cover first.
3. Collect Google's own related questions. Note People also ask, related searches and autocomplete suggestions for the question and each of its parts. In Search Console, filter the Performance report with a custom regex such as ^(\S+\s+){6,}\S+$ to list queries of seven words or more that already bring you impressions. Google says AI Overviews and AI Mode are reported in the Web search type of that report.
4. Add Bing's grounding queries if you have them. If Copilot already cites your site, export the grounding queries from Bing Webmaster Tools. They are the nearest thing to real sub-queries any platform shows.
5. Sort by intent. Paste up to 50 questions into the free AI Prompt Intent Classifier to separate learning questions from comparing and buying ones. Intent decides where a sub-question goes, so this step matters more than search volume.
6. Merge rewordings and trim. "Modular kitchen price" and "cost of a modular kitchen" are one sub-question. Remove anything outside what you sell or know. Our rule of thumb is 8 to 12 distinct sub-questions per main question.
7. Check demand where it matters. For the sub-questions that might deserve their own page, check search demand with the methods in our keyword research guide.

Third-party fan-out generators can speed up step 2, and Search Engine Land's fan-out guide recommends them for surfacing related questions, paraphrases and follow-ups. Treat their output like your own list: predictions from another model, good enough to plan sections and pages, but not a list of searches Google will run.

A worked fan-out map: one question, twelve sub-questions

Here is the method applied to one example. The business, the question and the sub-questions are an illustration: they are our prediction of a plausible fan-out, not queries captured from Google.

Example: a modular kitchen firm in Bengaluru wants to be cited for the question "How much does a modular kitchen cost for a 2BHK flat in Bengaluru, and is it worth it over a carpenter-made one?" It is the kind of long, two-part question AI Mode users ask. Working through the steps above gives twelve sub-questions in four groups.

What the map shows

• Half belong on one page. Six share the main question's learning intent and need short answers, so they become sections of one guide.
• One deserves its own page. "Modular vs carpenter-made" is a comparison people search for on its own, so it gets a separate page, linked both ways.
• Three are buying questions. Choosing a firm, reviews and payment options are answered by the service page, the Google Business Profile and reviews. Our guide to local SEO for AI search covers how those sources feed AI answers.
• Two need no new section. Warranty is already covered on an existing page, so the guide links to it. Resale value is a fair question, but the firm cannot back up an answer, so a section would only add guesswork.

Notice what the map does not contain: a page per sub-question. Twelve thin pages would compete with each other and are the pattern Google's guide warns against.

One page or many? A decision rule

Keep a sub-question on the main page when it shares the main question's intent and can be answered in a paragraph, list or table. Give it its own page only when it has a different intent, people search for it on its own, and you have enough first-hand material to fill a page.

Why not one giant page?

Covering everything on one page sounds safe, but the evidence does not favour it. In the April 2026 AirOps study of ChatGPT citations, pages covering 26–50% of the fan-out sub-queries were cited more than pages covering all of them, and pages over 5,000 words did worse than pages under 500. That is one engine and one dataset, so treat it as a caution against bloat, not a word limit.

The opposite mistake is fragmentation. Search Engine Land's guide argues that each article in a topic hub should be self-contained, because a page covering one narrow angle is less likely to hold up across the whole fan-out. That is a practitioner view, not Google's, but it matches the rule: split by intent, not by wording. If you already have overlapping pages, our post on service pages versus blog posts explains how to stop them competing.

Write passages that answer sub-questions clearly

Give each sub-question a heading in the words customers use, answer it in the first one or two sentences, then add the evidence. A passage that makes sense when quoted on its own is easier for any AI system to use and easier for any reader to trust.

1. Name the sub-question in the heading where it reads naturally: "How long does installation take?" rather than "Timelines".
2. Answer first. Put the number, the yes or no, or the recommendation in the opening sentence.
3. Make paragraphs stand alone. Repeat the subject instead of "it" or "this", and include the unit, place or condition, so a quoted paragraph still makes sense.
4. Use the right format. Tables for comparisons, numbered lists for processes, short paragraphs for explanations.
5. Add what only you know. Your own price ranges, timelines, photos and the mistakes customers make. Google's guide calls this non-commodity content, and it is the reason to cite you rather than the next page.
6. Keep the page readable. Do not chop it into dozens of tiny sections; Google says chunking is not needed.

Before and after: an example passage

Example, for the kitchen firm above. Weak: "Many factors affect timelines and every project is different, so it is best to speak to us." Strong: "A 2BHK modular kitchen takes [X] working days to install in Bengaluru once the design is signed off. Factory production takes [Y] of those days and fitting takes [Z]." The brackets are placeholders for your real figures. The strong version answers the sub-question, names the conditions and can be quoted on its own.

To check a draft, run it through the free AI Answer Snippet Extractor, which scores every paragraph out of 100 as a quotable snippet. Rewrite the lowest-scoring paragraphs under your sub-question headings first; they usually open with background instead of the answer. Titles and internal links still need the basics in our on-page SEO checklist.

What not to do

• Do not publish near-duplicate pages for rewordings of the same sub-question.
• Do not invent figures to make a passage more quotable. A wrong number in an AI answer gets repeated back to your customers.
• Do not write sections for sub-questions you cannot answer well. Leave them out, as the example firm did with resale value.

Check coverage against competitors and AI answers

Once the page is drafted, compare it with the pages AI answers already cite and with your competitors' sites. You are looking for sub-questions, entities and whole topics that they answer and you do not.

1. Re-run your question in AI Mode. Use the same signed-out setup three times and record each cited page. For each cited page, note which of your sub-questions it answers.
2. Compare entities. Put your page and a cited competitor page into the free Entity Coverage Checker. It lists the entities each page mentions. A material, fitting or standard you never name often points to a sub-question you missed.
3. Find whole missing topics. The Competitor Content Gap Finder reads your sitemap and your competitors' sitemaps and lists topics they cover that you do not. Run each result through the decision rule before writing anything.
4. Log it in a coverage grid. One row per sub-question, as in the template below. Review it whenever you update the page.

The rows above are an illustration for the example firm. One limit: a competitor being cited does not prove its coverage is the reason. Rankings, reputation and wording match all play a part, so match their completeness where it helps readers, not their pages.

Measure whether it worked

Track three things before and after each change: AI impressions by page in Search Console, citations and grounding queries in Bing Webmaster Tools, and a repeated sample of AI Mode answers for your main questions. None of them shows Google's fan-out queries, so read them together and look for trends.

Search Console

Google's Generative AI performance report shows how often links to your site appeared in AI Overviews and AI Mode, grouped by page, country, device and date (checked October 2026). It has no query dimension, so compare each updated page over equal periods before and after the change. Our guide to the Search Console AI performance report covers setup and limits. If the report is missing, Google says the usual reasons are too few AI impressions or an opt-out.

Bing Webmaster Tools

Bing's AI Performance report shows how often your pages are cited in Copilot and Bing's AI summaries, and the grounding queries behind those citations. The data is sampled and shows no individual prompts. New grounding queries that match the sub-questions you added suggest the new sections are being used.

A repeated sample of AI answers

Single checks mislead. In its 2026 study, SparkToro found less than a 1-in-100 chance that ChatGPT or Google's AI would give the same list of brands twice when asked the same question 100 times. Run each main question several times on a fixed schedule, record whether your page is cited, and track the share of runs where you appear.

Query fan-out does not change what earns a citation so much as how many chances you get. A page that answers the real question and the questions around it clearly, with facts only you can supply, is retrieved for more of those searches. A page split into a dozen thin variants is not.

If you would like help mapping sub-questions and building pages this way, our content SEO and AI search optimisation services cover it, and you can get in touch to talk it through.