Key Takeaways
- Fan-out: one question becomes several sub-queries, each searched independently before the answer is written.
- Citation split: 37.9% of AI Overview citations rank in the top 10 for the original query.
- Direction: that overlap sat near 76% in July 2025, so direct SERP selection is shrinking fast.
- YouTube: 18.2% of citations from pages outside Google’s top 100 were YouTube URLs.
- Coverage: cluster breadth wins more citation slots than one strong head-term position.
- Opacity: engines do not expose their sub-queries, so coverage has to be simulated.
- Passage design: every section must answer one sub-query without needing the rest of the page.
- Measurement: count the angles you cover, not the position you hold.
What query fan-out is
Query fan-out is a retrieval technique where an AI search engine generates several related queries from one user question. It runs them concurrently, then builds its answer from the pooled results.
Your page no longer competes for the question that was typed. It competes for the questions the engine asked on its behalf.
Google documents this behaviour directly. For a question about fixing a lawn full of weeds, the system may also search for the best herbicides, removing weeds without chemicals, and preventing weeds altogether (Google Search Central, 2026).
The user asked one question. The retrieval layer asked four.
Why AI engines split a single query
Splitting exists because conversational questions carry several information needs at once. Long compound strings degrade retrieval quality badly.
Take a question like: which project management tool suits a remote team of ten on a fifty euro monthly budget. It contains three separate constraints:
- Product category: the use case behind the request.
- Team size: the working model it needs to fit.
- Price ceiling: the budget the answer has to respect.
No single indexed page reliably satisfies all three. Narrow queries return cleaner results, so the engine runs several and assembles partial answers instead.
How fan-out changes which pages get cited
Fan-out decouples citation from ranking. A page can be cited without ranking for the question the user typed, because it ranked for one of the sub-queries instead.
Ahrefs analysed 863,000 keyword SERPs and 4 million AI Overview URLs in March 2026. Only 37.9% of cited URLs appeared in the first ten blocks for the same query. Roughly a third did not rank in the top 100 at all (Ahrefs, 2026).
Their earlier study put that overlap near 76% in July 2025. The gap is the fan-out.
Fan-out behaviour by engine
The mechanic is shared across engines. The expansion behaviour and the source pool are not, and that difference should shape where you invest.
Google AI Overviews fan out into video and long-tail sub-queries
Google has confirmed fan-out as part of how AI Overviews and AI Mode retrieve sources. The system does not just expand your query into synonyms, it expands into adjacent intents the original question implies.
For example, a search for “best running shoes for flat feet” can fan out into sub-queries like stability shoe reviews, arch support recommendations, and podiatrist-approved brands, each pulling from a different set of pages.
- YouTube share: 18.2% of cited pages that fell outside Google’s top 100 were YouTube URLs (Ahrefs, 2026), and YouTube mentions are the strongest single correlating factor with AI Overview visibility in Ahrefs’ broader brand study.
- What to do: pair every commercial or how-to article with a short video covering the same core question, since video wins sub-query slots that text pages are already crowded out of.
- What to check: whether your own top pages rank for the sub-queries, not just the head term, using Search Console’s query report.
ChatGPT search expands only when it decides to search
ChatGPT runs an internal classifier before deciding whether a query needs live web data at all. Below a certain confidence threshold, it answers from training data and never touches the web, meaning your newest content is invisible to that response no matter how well it is structured.
When it does search, the prompt is broken into several parallel web queries, and reasoning-heavy modes expand further still. A question like “what’s the best CRM for a 10-person sales team” might fan out into pricing comparisons, integration checks, and G2 review summaries as separate searches.
- What to do: keep pricing, integration, and comparison content current, since these are the sub-query categories ChatGPT reaches for most often on commercial questions.
- What to check: ask ChatGPT your target question directly and watch whether it shows a web-search indicator. No indicator means it answered from training data, and content freshness will not move that response.
Perplexity shows its expansions in the open
Perplexity displays the searches it runs alongside the sources it drew from, directly in the response. That makes it the only major engine where you can observe fan-out without a third-party tool.
- What to do: run your three highest-value prompts on Perplexity and record every sub-query it lists, then check which ones your site already has a page for.
- What to check: repeat the same prompt on different days. The sub-queries shift, so a single run is a sample, not a fixed map.
Claude expands the least of the four
Claude answers most questions without searching at all, and expands into web queries only when the request explicitly needs current information. When it does search, the expansion is narrower than Google’s or ChatGPT’s.
- What matters instead: consistent brand mentions across the sources Claude can reach, since narrow expansion means fewer chances for a new sub-query to surface your page for the first time.
- What to do: prioritise getting cited accurately on third-party sites and documentation over building a wide internal cluster aimed specifically at Claude.
My guide to ranking in Claude covers this in more depth.
How to find the sub-queries behind a prompt
Engines do not publish their expansions. The practical goal is approximation, not exact replication. Recurring themes across many simulations matter more than mirroring one run.
- Perplexity observation: run your target prompt on Perplexity and read the queries it shows.
- Qforia: a free fan-out simulator, available at qforia.streamlit.app.
- Screaming Frog: paired with the Gemini API, useful for extracting expansions at scale. Download it at screamingfrog.co.uk.
- SERP features: People Also Ask boxes and related searches are an older, cruder version of the same expansion logic.
- Support and sales logs: the constraints real buyers add to questions are the constraints engines fan out on.
How to build content for sub-query coverage
Coverage is an architecture problem before it is a writing problem.
1. Map the cluster before the head term
List the sub-questions a prompt would generate. Check which existing page owns each one.
Unowned sub-questions are your brief. Publishing another page aimed at the head term adds nothing if three of its sub-queries have no home.
2. Write self-contained sections
Each section should answer its heading in the opening lines and stand alone if lifted out.
If a passage needs the two sections above it to make sense, it cannot be extracted. Extraction is the unit of citation.
3. Cover adjacent formats
A short tutorial video, a support doc, or a comparison table can win sub-query citations your article will not.
Format diversity widens the number of retrieval slots you are eligible for on a single topic.
Why cluster ranking beats head-term ranking
The eligibility gap is easiest to see side by side.
| Approach | What it earns you |
|---|---|
| One ranking page | A single citation chance, tied to whether the engine’s expansion happens to include your exact query. Strong position, narrow eligibility. |
| Topic cluster | One eligible page per sub-query. Ten pages covering ten angles compete in ten separate retrievals for the same user question. |
| Mixed formats | Slots that text cannot reach, particularly video on informational and how-to expansions where written pages are already crowded. |
Head-term ranking still helps. It competes for the direct citation and the featured snippet. It is no longer the whole game.
How to measure fan-out coverage
Coverage is measured as a percentage of sub-queries owned, not as a position.
Build the sub-query list for your three most valuable prompts. Map each one to a URL on your site. The unmapped rows are your content plan, and the ratio is your baseline.
1. Track it through Search Console first
Google Search Console shows which queries already surface impressions for a page, including long-tail phrasing you did not target directly. A sudden cluster of impressions around a phrase you never wrote toward is often a sub-query the engine is already fanning out to.
2. Use a keyword tool to widen the net
Tools like Ubersuggest surface related and question-based keywords around your head term, which is a fast way to draft an initial sub-query list before refining it against live prompt testing. Treat the output as a starting hypothesis, since these tools model traditional search demand rather than AI fan-out specifically.
3. Re-run prompts on a fixed schedule
Re-run the same prompts monthly and log how often you appear across repeated runs. Frequency across many runs is stable enough to trend. A single screenshot of one response is noise, because expansions and citations shift between runs.
Frequently asked questions
Does fan-out apply to every query?
No. Simple navigational or factual questions are often answered without expansion, and some are answered without any live search. Fan-out is most active on complex, multi-constraint, conversational questions, which is exactly where commercial research queries sit.
Can you see the sub-queries an engine runs?
Not directly for Google. Perplexity displays the searches it performs, and simulation tools like Qforia approximate expansions. Treat all of these as samples: the point is spotting recurring themes across many runs, not reproducing one exact set.
How many sub-queries does one prompt generate?
It varies by engine, question complexity, and reasoning mode. Simple prompts may trigger a handful, while deeper reasoning modes expand considerably further. Plan for breadth rather than a fixed number, since the count changes between runs of the same prompt.
Does fan-out make head-term rankings worthless?
No. Ranking for the head term still competes for the direct citation and the featured snippet, and it remains the page buyers land on when they verify what an AI told them. Fan-out adds eligibility elsewhere, it does not remove eligibility here.
How long does cluster coverage take to pay off?
Longer than a single page refresh and shorter than classic link building. Each new page covering an unowned sub-query becomes eligible immediately once indexed, so returns arrive incrementally as coverage widens rather than at one visible threshold.
Want your content mapped against real sub-queries?
I’ll audit your top pages against the prompts your buyers actually run and show you exactly which sub-queries you’re missing.
Book a free strategy call