Query fan-out is the step where an AI engine splits one question into several related searches, runs them at once and builds its answer from the pages they return. To optimise for query fan-out, cover each sub-topic a buyer's question implies with a specific page, then track a plain-language prompt for each one.
Key takeaways
- Google says Google Search Central, 2026 that AI Overviews and AI Mode may use query fan-out, so one question can pull pages that never ranked for the original wording.
- You cannot see most sub-queries directly. Decompose the buyer question yourself, then check the result against what engines actually cite.
- Cover sub-topics on pages that deserve to exist. Google warns that creating pages per variation to manipulate responses breaches its scaled content abuse policy.
- Track one prompt per sub-topic, per engine, on a fixed schedule. A named mention in the answer is the test, not a ranking.
- Sub-queries are getting more specific: Peec AI measured ChatGPT fan-outs growing from about 6 to about 12 words. Specific attributes on your pages matter more each month.
What is query fan-out in AI search?
When you type a question into an AI search feature, the engine rarely sends your exact words to a search index. It breaks the question into narrower searches, one per piece of information it needs, fetches results for all of them and then writes one answer. Each of those narrower searches is a fan-out query, or sub-query.
Google describes it for its own products. In its Google Search Central AI features documentation, as checked on 2026-11-11, both AI Overviews and AI Mode "may use" a query fan-out technique that issues multiple related searches across subtopics and data sources. Google's Google blog, May 2025 put it more plainly: AI Mode is "breaking down your question into subtopics and issuing a multitude of queries simultaneously".
Google's own optimisation guide for generative AI features gives an example. For the question about fixing a lawn full of weeds, fan-out queries might include best herbicides for lawns, removing weeds without chemicals and preventing weeds in a lawn. One question, three searches, three potentially different sets of pages.
Query fan-out: a set of concurrent, related queries generated by the model to fetch additional relevant search results. Paraphrased from Google Search Central's AI optimisation guide.
Query fan-out is not unique to Google. ChatGPT does it too, though less often than many articles imply, as the numbers below show.
How often do AI engines actually fan out?
Ahrefs analysed 66.6 million free, logged-out ChatGPT responses and their 15 million fan-out queries from August and September. Its Ahrefs study, October 2026 found that 14.4% of responses contained fan-out queries, with about 1.6 fan-out queries on average in those responses. It also found that 6.8% of fan-out queries used a site: operator to search a single website.
Those figures are internally consistent, which makes a useful check. 66.6 million responses at 14.4% is roughly 9.6 million responses with fan-outs. At 1.6 queries each, that is about 15.3 million queries, matching the 15 million reported. The same arithmetic puts site: queries at roughly 1 million. That is my calculation from Ahrefs' published rates, not a figure Ahrefs states.
What it means in practice: most ChatGPT answers in that sample used no fan-out at all, and when they did, there were only one or two queries. The sub-queries that exist are the ones deciding which pages get retrieved. The site: share is a hint that engines sometimes go straight to a named domain, so a clear, crawlable site structure matters when a brand is already known.
The shape of the queries is changing too. Peec AI analysed over 20 million ChatGPT fan-outs between October 2025 and January 2026. Its Peec AI research, February 2026 reports that average fan-out length rose from about 6 words to about 12, while the number of fan-outs per prompt stayed roughly constant. Peec's reading is that ChatGPT is making each query more precise rather than issuing more of them.
Both datasets come from one engine and a limited period. Treat them as direction, not as a law that applies to Gemini, Perplexity or Claude.
Why fan-out changes what you optimise for
A classic SEO brief starts with a keyword. A fan-out brief starts with a decision the buyer is making. The buyer's question is the parent. The engine's sub-queries are the evidence it wants. Your page has to answer one of those sub-queries well enough to be retrieved, and then be clear enough to be quoted.
Three consequences follow.
- Ranking for the head term is not enough. A page can rank for the parent question and still be absent from the answer if it says nothing on the comparison, price or integration sub-query.
- The unit of competition is the sub-topic. You may beat a rival on category pages and lose on pricing pages, and the answer will name the rival.
- Longer, more specific queries favour concrete attributes. If a sub-query reads like "project management software for 20-person agencies with client portal", a page that states team size, features and constraints in plain words has an edge over a generic one.
For the wider picture of how this fits with the rest of your programme, see AEO vs SEO vs GEO and the seven-step GEO strategy.
The fan-out worksheet: turn one buyer question into sub-queries
Most explainers stop at a definition and a lawn example. This is the worksheet to run instead. It takes about an hour for one buyer question.
Step 1: write the parent question the way a buyer asks it
Use a real decision, with constraints. "Best project management tool for a 20-person marketing agency that bills by the hour" is a good parent. "Project management software" is not, because it carries no constraints for the engine to split on.
Step 2: split it along eight sub-query types
Every buyer question implies some mix of these. The examples in the table are illustrative, written by hand to show the method. They are not observed engine output.
| Sub-query type | Illustrative sub-query | Page that answers it | Prompt to track |
|---|---|---|---|
| Category definition | what is agency project management software | Category explainer | What is the best way to manage client work in a 20-person agency? |
| Shortlist | best project management tools for agencies | Your own roundup or comparison, plus third-party lists | Best project management tools for marketing agencies |
| Constraint fit | project management tool with time tracking and invoicing | Feature page naming the constraint | Project management tool with time tracking for hourly billing |
| Size or segment | project management for teams of 20 | Use-case page for the segment | Project management for a 20-person agency |
| Price | agency project management pricing per user | Public pricing page | How much does agency project management software cost? |
| Integrations | project management integrates with Slack and Google Drive | Integrations page | Which project management tools integrate with Slack? |
| Alternatives and versus | X vs Y for agencies | Alternatives and comparison pages | Alternatives to a named incumbent for agencies |
| Proof and freshness | agency project management reviews 2026 | Dated reviews, docs, changelog | Is the tool good for agencies? What do reviews say? |
Step 3: mark what you cannot guess
You will miss sub-queries. Use observation to fill gaps. Goodie's query fan-out guide, as checked on 2026-11-11, lists three ways to see them: Chrome DevTools for ChatGPT, the Gemini grounding API and the AI Mode thoughts view in Google AI Studio. Ahrefs says its Brand Radar lets you study fan-out queries across ChatGPT and Perplexity. Whichever you use, collect many answers and look for patterns. A single fan-out tells you very little.
Step 4: map each sub-query to a page, or to a gap
For each row, write the URL that answers it. If none does, record a gap. If a page exists but buries the answer in paragraph six, record a rewrite. Use three statuses: covered, weak, missing.
Step 5: check the page can be read
A page that exists but cannot be fetched or parsed never enters the answer. Run it through the free Noma AI SEO checker, which shows whether ChatGPT, Claude, Perplexity and Google AI Mode can read your page, with its SEO checks and top fixes. Pair it with the free AI crawler checker to confirm that GPTBot, OAI-SearchBot, ClaudeBot and PerplexityBot are not blocked in robots.txt.
Step 6: turn each row into a tracked prompt
Write the prompt in buyer language, not in the engine's likely query syntax. You are testing what a person would ask, not reproducing the hidden query. Keep wording fixed so changes over time mean something. The prompt scoring sheet and Chapter 3 of the AEO guide cover how to select and prune.
How do I find the sub-queries AI engines run for my topic?
Use these methods in order of effort, and expect the first to be the most useful for most teams.
- Decompose by hand. Use the eight types above. It takes minutes and gets you most of the structure.
- Read the engine's own sources. Ask the parent question in each engine and list every cited URL. Pages cited are evidence of which sub-queries retrieved them. Noma's article on how AI engines choose citations covers this.
- Inspect fan-outs directly. Use the developer-side methods Goodie lists, or an industry-wide tool such as Ahrefs Brand Radar. Record the date and engine version with each pull, since Peec's data shows behaviour shifts within months.
- Use the People also ask box and related searches. They are a proxy, not the engine's actual list. Label them as such in your worksheet.
- Check the long tail with the Noma free AI prompt finder. It lists questions buyers ask about a category, tagged by intent, which helps you spot sub-topics you did not list.
A fair limit: none of these methods gives you a complete, stable list. Engines change their behaviour, and the same prompt can produce different sub-queries on different days. That is why the worksheet tracks outcomes, meaning who is named, rather than treating any list of sub-queries as final.
How do I optimise for query fan-out without spamming?
Google's guide says it is tempting to build separate content for every variation of how people might search, including fan-out queries, and that doing so primarily to manipulate rankings or generative AI responses violates its scaled content abuse spam policy. Take that as the boundary.
The workable rule: one page per decision-relevant sub-topic, not one page per phrasing. "Pricing per user" and "how much does it cost" belong on one pricing page. "Integrates with Slack" and "integrates with Google Drive" belong on one integrations page with a section each.
Then make each section quotable:
- Put the answer in the first sentence under a question-style heading.
- Name your product in full and state the fact: who it suits, what it costs, what it connects to.
- Date anything that changes, such as prices and features.
- Use tables for comparisons, so the engine can lift a row.
- Add the constraint words buyers use: team size, industry, budget, region.
Fan-out also rewards pages you do not own. Shortlist and comparison sub-queries often retrieve third-party lists, so earned placement matters. See how to get listed on third-party lists.
How do I check which sub-queries name my brand?
This is where the worksheet pays off. Each row has a prompt. Run each prompt on every engine you care about, several times, and record three things: whether your brand is named, which competitors are named instead and which URLs are cited. Then read the grid by sub-query type.
| Pattern in the grid | Likely meaning | Fix |
|---|---|---|
| Named on shortlist, absent on price | Pricing is hidden or gated | Publish a dated public pricing page |
| Named on category, absent on constraint fit | No page states the constraint | Add a use-case or feature section with the buyer's words |
| Absent everywhere, cited pages are third-party lists | Retrieval comes from roundups | Earn placement, publish a fair alternatives page |
| Named on one engine only | Engines retrieve different sources | Compare cited URLs per engine |
| Cited but not named | Page retrieved, brand not framed in the answer | State the brand and the claim in one sentence |
Do it by hand first. A manual baseline in ChatGPT takes an afternoon for 20 prompts. The cost is repetition: you need to run each prompt more than once because answers vary, as AI citation stability explains. The AI Visibility Measurement Standard sets out numerators, denominators and margin of error.
Where Noma fits
Noma by Nurdd is an AEO and AI visibility platform. It asks 13 AI engines, including ChatGPT, Gemini, Perplexity and Claude, the buyer questions you give it every day, and shows where your brand is named, which competitors are named instead and which pages the answers cite. Put each worksheet row in as a tracked prompt and Noma shows which sub-topics name you. It does not reveal the engines' hidden sub-queries.
See it on your own brand: book a demo or start the 3-day free trial. A card is taken at signup so billing can start on day 3; cancel before then and nothing is charged. Plans and prompt limits are on the pricing page.
Common mistakes with query fan-out
- Treating a tool's sub-query list as the truth. It is a sample from one engine on one day.
- Writing one page per phrase. It dilutes your site and risks Google's spam policy.
- Tracking only the parent question. You then cannot tell which sub-topic is costing you the mention.
- Ignoring pages you do not own. Many sub-queries resolve to reviews, lists and community threads.
- Changing prompt wording between runs. You lose the ability to compare.
- Assuming every answer fans out. In Ahrefs' ChatGPT sample, most responses did not.
How we researched this
We reviewed Google Search Central documentation and Google's product announcement on AI Mode, published research datasets from Ahrefs and Peec AI, and the top-ranking vendor guides on query fan-out, including Goodie's. We took no data from Noma customers and ran no new experiments. The sub-query examples in the worksheet are illustrative, written to show the method. Facts, figures and links were checked on 2026-11-11.
What to do next
Pick your most valuable buyer question, fill in the eight-row worksheet and mark each row covered, weak or missing. Fix the missing rows first, then track every row as a prompt. To see which of them already name your brand across 13 AI engines, book a demo or start the free trial.
See what AI engines say about your brand
Free · no account · or start a 3-day trial
Questions people ask
What is query fan-out in AI search and how do I optimise for it?
Query fan-out is when an AI engine splits one question into several related searches, runs them together and writes one answer from the results. To optimise, break your buyer question into sub-topics such as price, integrations and comparisons, and cover each on a clear, crawlable page. Then track a prompt for each sub-topic to see whether you are named.
How do I find the sub-queries AI engines run for my topic?
Start by decomposing the question by hand into types such as shortlist, price, constraints and alternatives. Then check observed data: developer tools for ChatGPT and Gemini, or a tool such as Ahrefs Brand Radar, and list the URLs each engine cites. No method gives a complete, stable list, so verify with tracked prompts.
Do all AI answers use query fan-out?
No. In an Ahrefs study of 66.6 million free ChatGPT responses, 14.4% contained fan-out queries, averaging about 1.6 queries each. Google says AI Overviews and AI Mode may use the technique, not that they always do.
Should I create a page for every fan-out query?
No. Google's guidance warns that creating content for each variation mainly to manipulate responses violates its scaled content abuse policy. Build one page per decision-relevant sub-topic and give each question its own section.
Are fan-out queries the same as keywords?
Not exactly. They are searches the model writes itself, and Peec AI found ChatGPT's averaged about 12 words by January 2026. They read more like specific phrases than short keywords, so pages should state concrete attributes in plain language.
Can Noma show the hidden sub-queries?
No. Noma by Nurdd tracks the prompts you set across 13 AI engines and shows whether your brand is named, who is named instead and which sources are cited. Use it to check which sub-topics name you, not to read the engines' internal searches.