Query Fan-Out: How Google's AI Turns One Question Into a Dozen Searches

One question typed, a dozen searches run

Classic search matched your words against an index and handed back ten links. Google's AI features add a step: the model works out what else it needs to know, searches for that too, and only then writes a reply. Google's AI optimization guide defines query fan-out as "a set of concurrent, related queries generated by the model" to retrieve more search results.

The consequence for a business is easy to state and easy to miss. The pages that support an AI answer are not only the pages that rank for what the customer typed. They are also the pages that rank for the questions the model asked on the customer's behalf, which nobody at your company has ever seen. Google's Search Central page on AI features says these extra searches let its systems find more supporting pages while the response is generated, and show a wider and more varied set of links than a classic search.

So a small site with a precise answer to a narrow sub-question can earn a link beside the national brands. The GEO hub covers every engine; this page stays on the mechanism.

Google's own account, release by release

Google has described fan-out in five places since early 2025. Read in order, the technique grows with each model, and the advice to site owners gets more explicit.

Notice the word Google keeps using: may. Fan-out is something the systems can do, not a stage every search passes through. AI Overviews often do not appear at all, and AI Mode falls back to plain web links when it lacks confidence. Nothing Google has published says a two-word question fans out the way a comparison does.

The weedy lawn, and a roof after hail

Google's guide uses a plain example. Someone asks how to fix a lawn full of weeds. The model might also search for the best lawn herbicides, for ways to remove weeds without chemicals and for how to prevent weeds. Three extra searches, and three different kinds of winning page: a product comparison, an organic method, a seasonal routine. A lawn company whose site says only 'weed control services' has one shot in four.

Here is the same pattern on a local buying question. Google does not publish the sub-queries it runs, so this is our illustration of what a careful buyer, and a model acting for one, would also want answered.

A typical roofing site answers one row, the trust row, through its Google profile. The other five go to whoever bothered to answer them: a manufacturer, an insurer's blog, a national cost site. The local roofer gets named, if at all, for the one row where it had something to say. GEO for roofing companies covers the trade in full.

What Google has confirmed, and what it has left blank

Fan-out is now a favorite word in GEO sales decks, so separate what Google stated from what others inferred.

Then there is the folklore. Some tools and prompt templates promise to show 'the fan-out queries' for a topic. They are guesses, sometimes useful ones. Google's AI optimization guide says no third-party tool has access to its internal ranking or AI systems. Use a simulated list to brainstorm, never as a log of what Google did. Our page on AI SEO myths tests more claims like it.

What one outside study saw when it forced the searches

The clearest public measurement we found comes from Seer Interactive, a marketing agency. In November 2025 it ran 501 prompts through the Gemini 3 API with Google Search grounding forced on, and recorded the searches the model issued. It counted 10.7 searches per prompt on average, ranging from 3 to 28, up from 6.01 with Gemini 2.5. The searches averaged 6.7 words, 95% had no measurable global search volume, 21.3% contained a year and 26.4% named a brand.

• Keyword tools miss most of them. A list sorted by volume drops exactly these queries.
• The searches are specific. Six or seven words is an attribute or a comparison, not a head term.
• Dates show up often. A page that says when its prices were last checked can match a query with a year in it.
• Brand names appear in a quarter. A contractor that never names the equipment brands it installs cannot answer a sub-search that does.

Coverage beats a longer keyword list

If the model writes its own searches, chasing phrasings is a losing game. Google's guide says as much: AI systems understand synonyms and general meaning, so there is no need to fill gaps in long-tail keywords or cover every way of saying a thing. It also says one page can cover several topics and there is no ideal length. What matters is whether the facts behind the sub-questions are on the page at all.

In our practice, the same facets recur for nearly every local service: cost, timing, risk, alternatives, local conditions, proof and the current year. A page that answers each under its own heading covers most plausible fan-outs without guessing their wording. How to phrase those passages so they survive extraction is the subject of content AI can quote; how a site earns trust across a subject is covered in topical authority.

You cannot rank for a search you will never see. You can answer the question it was asking.

The tempting shortcut is forty thin pages, one per imagined sub-query. Google calls that scaled content abuse when the aim is to manipulate rankings or AI responses. Build a separate page only when a person would want it as a page. Keyword clustering is how we decide which questions share a page and which earn their own.

Mapping the fan-out for one service

This is the procedure we run for clients, and an owner can run it with a notepad in an afternoon.

1. 1

   Write the opening question

   Phrase it the way a customer says it on the phone, with the town: 'is it worth repairing a 15-year-old AC unit in Schertz'.

2. 2

   List the facets

   Cost, timing, risk, alternatives, local conditions, credentials, and the brands or models involved. Each facet is a likely sub-search.

3. 3

   Collect real phrasings

   Pull questions from the page's Search Console queries, sales calls, reviews and the People Also Ask boxes.

4. 4

   Read the AI answers

   Ask the question in AI Mode and on a search that shows an AI Overview. Open every cited link and note which facet each one answers.

5. 5

   Grade your own site

   Mark each facet answered, weak or missing. Weak usually means the fact is buried in a paragraph about something else.

6. 6

   Fix in place before adding pages

   Add a heading, a direct answer and the number. Create a new page only where the facet deserves one for human readers.

7. 7

   Check again in a month

   Models and competitors change, and so do the cited links. Rerun the list on a schedule.

With dozens of services this gets tedious. On sites built with Athena CMS, the Prometheus agent finds customer questions the site does not answer and drafts the page, for a person to approve.

Sub-searches that land on your listings, not your pages

Fan-out reaches past web pages. Google's launch post named the Knowledge Graph, real-time sources and shopping data among AI Mode's sources, and its AI optimization guide says AI responses can include product listings and local business information, with Merchant Center feeds and Business Profiles helping products and services appear. So a sub-search for hours, ratings, price or stock may be answered from your profile or feed without touching the site.

Measuring searches you will never see

Google gives site owners results, not the searches behind them. Clicks from AI Overviews and AI Mode count in the Web search report in Search Console. Since August 31, 2026 every site also has a Generative AI performance report, but it shows impressions only, broken down by page, country, device and date, with no query column. A page that collects AI impressions without ranking for the head term is probably supporting someone's sub-question. That is our inference, not a Google statement.

For the engines outside Google, put a real customer question through our free AI visibility check. It asks ChatGPT and the others through their APIs and shows each answer and who was named; API answers can differ a little from the consumer apps. Repeating a fixed set of questions over time is prompt tracking, and how AI chooses businesses explains what happens after retrieval.

At DataCram, fan-out mapping is part of the GEO work in our monthly SEO programs, with reports that show the answers. Programs start at $625 a month and most local businesses land between $625 and $2,250, with no long contract. Scope is set after a free GEO audit; GEO pricing explains what moves the number.