Your page ranks well for a question but never shows up in the AI answer for that same question. When I see this, query fan-out is usually part of the reason. Query fan-out is how AI search splits one question into several related searches. It runs them at the same time and builds one answer from everything they return.
That means you’re not only competing for the query the person typed. You’re competing for the sub-questions the AI quietly asked for them. In this guide I explain how Google describes the process, what its latest research shows, how other AI tools do the same thing, and what you can change in your content and reporting.
What Is Query Fan-Out?
Query fan-out is a way of finding information. The AI doesn’t just run one search and sum up the top results. It writes a set of related searches, runs each one, and combines the results into an answer. I first looked at this closely when Google used it in its Web Guide experiment, which groups normal web results by topic.
Google’s own Search Central documentation defines a fan-out as a set of concurrent (running at the same time), related queries the model creates to ask for more information. It gives a simple example. Someone asks “how to fix a lawn that’s full of weeds”. The fan-out queries might include:
- best herbicides for lawns
- remove weeds without chemicals
- how to prevent weeds in lawn
Look at what happened there. The first question is about fixing a problem. The fan-out adds a product angle, another method and a prevention angle. A page that only answers the literal question may miss all three.
Why AI Search Works This Way
Chat-style questions are long. They often hold several needs at once. A single search rarely returns one page that covers all of them well.
Splitting the question lets the system pull the best source for each part. The answer can then cite a forum thread for one point, a maker’s page for another and a guide for a third. It’s also why AI answers tend to link to a wider range of sites than the classic top ten.
How Google Describes Query Fan-Out for AI Overviews and AI Mode
Google has been open about using this. When it launched AI Mode at Google I/O in May 2025, it said AI Mode uses query fan-out to break a question into subtopics and run many searches at once. It said the Deep Search option in AI Mode can go further, running hundreds of searches for one research-style question.
Google’s Search Central page AI features and your website puts it plainly. Both AI Overviews and AI Mode may use query fan-out, running several related searches across subtopics and data sources to build a response. While the response is being written, the models find more supporting web pages. That’s how these features can show a wider and more varied set of links than a classic results page.
The same page makes a few points that matter for site owners:
- The entry ticket is normal indexing. To appear as a supporting link, a page must be indexed and eligible to show in Google Search with a snippet. Google says there are no extra technical requirements.
- The two features can differ. AI Overviews and AI Mode may use different models and methods. So the links they show for the same question can vary.
- AI Overviews are selective. They only appear when Google’s systems decide they add something to classic Search.
For the fuller picture of how each feature picks sources, see my comparison of how AI Overviews and AI Mode pick their sources.
What Google Says Not to Do
Google’s guide to optimising for generative AI features also has a warning that many fan-out guides skip. Making separate pages for every possible way people might search, mainly to game rankings or AI answers, breaks Google’s scaled content abuse spam policy.
The guide adds that more pages don’t make a site higher quality. Google can judge relevance without an exact query match. It also says you don’t need to break content into tiny chunks for AI, write in a special AI style, or add special markup or files. So the answer to fan-out is better coverage on fewer, stronger pages. Not hundreds of thin ones. I fully agree with Google here.
Retrieve-for-Train: Google Research’s Work on Faster Fan-Out
On 15 September 2026, Google Research published a blog post on a framework called Retrieve-for-Train, or R4T. The paper behind it, by Google Research and University of Illinois Urbana-Champaign authors, was first posted to arXiv in March 2026.
The problem it tackles is cost and speed. Creating good fan-out queries with a large language model is slow. The model writes queries one token (a small piece of text) at a time and may call search again and again. R4T moves that heavy thinking into training, so the system that serves the answer can be much lighter.
How R4T Works, in Plain English
- Teach a large model what good fan-out looks like. A fan-out language model is trained with reinforcement learning (learning from scores). The rewards score the whole set of results, not each result alone.
- Collect its best work. The trained model runs on many queries. Its highest-scoring outputs become synthetic (machine-made) training data.
- Train a small, fast model to copy it. A light diffusion model (about 54 million parameters) learns to produce all of the search targets in one parallel pass.
The rewards balance three qualities, as reported in the blog post: groundedness (results must match real items), diversity (broad coverage without repeats) and alignment (staying close to what the person asked). The paper reports a steady 12 to 20 times speedup for the diffusion model over the step-by-step (autoregressive) approach.
What R4T Does and Does Not Tell You
Google has not said that R4T is running in Google Search, AI Overviews or AI Mode. The paper’s tests were on fashion outfit search and music playlist search, not web search. I treat it as a view of where Google’s thinking is going, not a description of the live system.
The useful lesson for content teams is in the rewards. Google’s researchers are clearly aiming for sets of results that are varied, fit together, and still stay on topic. If that thinking reaches Search, pages that cover a distinct, truly useful angle have a better chance of being picked than pages that repeat what ten others say.
How Other AI Search Tools Use Sub-Queries
Google named the technique, but it’s not unique to Google. Most AI search products now turn a prompt into several searches before they answer.
- Perplexity: Its help centre says Pro Search runs multiple searches across the web and is built to break down unclear or many-sided questions. It also shows users how it broke the question down. I find that makes it a handy research tool for you.
- ChatGPT search: ChatGPT turns a prompt into web searches before it answers. Its deep research mode runs many searches over several minutes for one task. The searches it ran are sometimes visible in the answer, so they’re worth checking.
- Deep research modes in general: Research-style modes, including Deep Search in AI Mode, run far more searches than a quick answer. So the number of sub-queries grows with how complex the question is.
The details differ by tool, and none of them publish their full logic. The common pattern is what matters. Your page gets pulled in for a sub-question, not always the main one. And the tool picks the clearest answer to that sub-question it can find. My guide to getting cited in AI Overviews, AI Mode and ChatGPT Search covers the platform-specific side.
What Query Fan-Out Means for Your Content
You can’t see Google’s fan-out queries for a given search. Google’s guide says no third-party tool has access to its internal ranking or AI systems. But you can make your content easy to find for the sub-questions a topic naturally raises.
Cover the Sub-Questions People Actually Have
Think about the question behind the question. Someone asking “is a heat pump worth it for an old house” will also want to know about running costs, insulation, grants, noise and how long it takes to install.
A good page answers the main question and the obvious follow-ups. That’s simply what a helpful expert would do. It also happens to match how fan-out works.
Here’s an example. A plumbing firm’s page on “how to fix a leaking tap” could include short sections on spotting the tap type, the tools needed, when a washer or a cartridge is the fix, and when to call a plumber. Each section is a likely fan-out target.
Use Clear, Self-Contained Sections
Google recommends organising pages with clear headings and sections. For fan-out, this helps because each sub-query may match one part of your page.
- Write headings that say what the section answers, such as “How much does a heat pump cost to run?” instead of “Costs”.
- Open each section with a direct answer in the first sentence or two, then add detail.
- Make sure a section still makes sense if someone reads only that part.
- Put specifics in text: prices, sizes, steps, conditions, dates.
This isn’t the same as chopping content into tiny pieces, which Google says you don’t need to do. It’s just good structure. Clear sections help human readers skim too, and in my view that’s the real test.
Build Topical Depth Across Your Site
Some fan-out queries will be too big to answer well inside one page. That’s where a set of related pages, linked sensibly, does better than one bloated article.
The line to watch is purpose. A separate page on “heat pump grants” is fine if it’s truly useful on its own. Twenty near-identical pages for every way of saying “heat pump cost” are exactly what Google’s scaled content abuse policy targets. My guide to planning topical authority shows how to map a subject without making thin pages.
Offer Something the Others Don’t
Fan-out pulls several sources into one answer. So the system has a reason to choose pages that add a different angle. Google’s guide stresses original, first-hand content that isn’t the same as everyone else’s.
Your own pricing data, real installation photos, local rules, a comparison you actually tested: these are harder to swap for another source than a generic summary. This is where I’d put most of my effort.
How to Research Likely Fan-Out Queries
You’re not trying to guess Google’s exact sub-queries. You’re trying to find the sub-questions a topic reliably raises, so your coverage has no obvious gaps.
- Start with the real question. Write the main question the way a customer would ask it, in full sentences, with their situation (“for an old house”, “on a budget”, “in winter”).
- Break it into facets. List the angles it raises: definitions, costs, comparisons, steps, risks, alternatives, prevention, local factors. The lawn example in Google’s guide shows this pattern: products, methods and prevention.
- Check Google’s own suggestions. Look at People Also Ask, related searches and autocomplete for the main query and for each facet. These show what people search next.
- Watch AI tools break it down. Ask the question in Perplexity Pro Search or ChatGPT search and note the searches they show. Run it several times, because the sub-queries change between runs. Look for patterns, not exact wording.
- Dig into your own data. Search Console queries for the page, sales and support questions, and site search terms often show the follow-ups customers really ask. I usually find the best ideas here.
- Map facets to content. For each facet, decide if it’s a section on the existing page, a separate page worth linking to, or out of scope. Then fill the gaps.
Be careful with tools that promise to show Google’s “real” fan-out queries. Google’s guide is clear that no third-party tool has access to its internal systems. Simulated fan-out can still help with brainstorming, as long as you treat it as a guess. My guide to keyword research for AI search goes deeper on turning keywords into question sets.
How to Measure the Impact of Query Fan-Out
No report shows which fan-out query triggered your citation. You have to measure the outcome instead.
Use Search Console’s Generative AI Reporting
Google says appearances in AI features are counted in the Search Console Performance report under the Web search type. Its AI optimisation guide also points to the Generative AI performance report to see how people find your content through these features. My walkthrough of Search Console’s generative AI reports explains how to read them.
Look for pages gaining impressions on a wide spread of long, specific queries. That pattern suggests a page is being matched to many sub-questions, not just its main term.
Track a Fixed Prompt Set
Pick 20 to 50 real questions from your research. Check them on a set schedule in AI Overviews, AI Mode, ChatGPT and Perplexity. Record whether your site is cited, which page, and which part of the answer it supports.
Answers change between runs. So check each prompt more than once and look at trends over weeks, not single results.
Connect It to Business Results
Citations only matter if they lead somewhere. Track AI referral visits and key actions in GA4. Compare pages you’ve expanded for sub-questions with pages you haven’t. If expanded pages earn more AI citations and steadier traffic, the work is paying off.
Common Mistakes With Query Fan-Out
- Mass-producing pages for every variation. Google names this as scaled content abuse.
- Burying answers. A great answer hidden in paragraph nine of a vague section is hard to match to a sub-query. This is the mistake I see most often.
- Chasing exact sub-query wording. Fan-out queries change between runs. Cover the need, not the phrase.
- Ignoring the basics. A page that isn’t indexed, or can’t show a snippet, can’t be a supporting link at all.
- Trusting “insider” tools. No outside tool sees Google’s internal fan-out.
Conclusion
Query fan-out means AI search answers the question behind the question. It runs several searches and stitches the best parts together. Google says AI Overviews and AI Mode may use it. Its September 2026 Retrieve-for-Train research shows it’s working to make fan-out faster and more varied. And tools like Perplexity and ChatGPT do something similar.
The right response is steady, not dramatic. Answer the real sub-questions, structure pages with clear sections, build depth without thin copies, and measure citations over time. My suggested next step: pick one important page, list its likely sub-questions using the steps above, and fill the gaps you find.
