What Are Fan-Out Queries?
Fan-out queries are the hidden sub-queries an AI search engine generates from a single question, runs in parallel, and then synthesises into one answer. Google calls the mechanism query fan-out and uses it in AI Mode and AI Overviews. It means your page is no longer competing for one keyword. It is competing for a set of related questions you never see.
Someone asks Google AI Mode a question. Google does not go looking for the page that best matches those words. It breaks the question into subtopics, fires off a batch of separate searches at once, gathers what comes back, and writes an answer from the pieces.
Those separate searches are fan-out queries. Nobody typed them. The model wrote them.
How does query fan-out work?
Query fan-out expands one query into many, retrieves sources for all of them at the same time, then builds a single answer from the results. Google introduced the technique alongside AI Mode in March 2025 and described it publicly that May (Google, 2025).
The sequence runs roughly like this:
- The query gets assessed. The system works out intent, complexity and whether fan-out is needed at all. "Capital of Spain" does not need it. "How do I pick a construction ERP" does.
- Sub-queries get generated. A custom version of Gemini writes them. Google's patent application describes this step as prompted expansion, where the model is instructed to produce queries emphasising different kinds of intent (Google patent application US20240289407A1).
- Everything runs in parallel. Against the live web, and against Google's own graphs including the Knowledge Graph and Shopping Graph.
- Sources get filtered, then synthesised. Retrieval is not citation. Most of what comes back never makes the answer.
The last step is the one people underestimate. One study of 548,534 pages found ChatGPT cites only around 15% of the pages it retrieves (AirOps, 2026). Getting pulled in is the easy part, and the gap between being fetched and being quoted is the whole game. That distinction sits at the heart of how retrieval-augmented generation works.
What kinds of sub-query does it generate?
The model is prompted to cover different angles rather than rephrase the same thing eight times. The usual categories:
- Equivalent: the same question in different words
- Broader: the concept one level up
- Narrower: a more specific angle
- Comparative: X versus Y
- Related: semantically adjacent topics
- Implicit: the thing the person did not ask but probably wants
- Parallel: sibling topics at the same level
That implicit category is where most B2B SaaS companies get caught out. Someone asks which tool is best for their situation. The model quietly also asks about pricing, integrations, implementation time and who it suits. If you have no page addressing those, you are absent from the sub-queries that decide the recommendation.
Why does content from page three end up in AI answers?
Because it was never ranking for the query you were watching. It was ranking for one of the hidden sub-queries.
This is the single most useful thing to understand about fan-out. A page sitting at position 25 for your head term can be position 2 for a narrow sub-query the model generated, get retrieved, and get cited. Meanwhile the page you spent six weeks getting to number one for the head term never gets pulled in, because the answer was assembled from sub-queries it does not cover.
Ranking and citation have come apart. They used to be the same thing.
What does fan-out change about SEO?
It moves the unit of competition from the keyword to the topic.
In classic search, visibility was binary. You were on page one for a term or you were not. With fan-out, visibility is spread across a set of sub-queries you cannot see, and you can be retrieved for a dozen of them while being cited for none.
Three practical consequences for B2B SaaS:
- Topic coverage beats keyword targeting. A single well-optimised page cannot answer a fan-out set. A cluster can, which is why cluster architecture now does more work than on-page tuning in B2B SaaS SEO.
- The boring pages matter more than they used to. Pricing, integrations, comparison and use-case pages are exactly what implicit and comparative sub-queries look for. Team 4 sees this constantly on audits: strong rankings, thin coverage of the adjacent questions, no citations.
- Ranking reports stop telling you the truth. Position tracking for your head terms will not show you which sub-queries you are missing, which is why tracking AI search visibility has become a separate job from rank tracking rather than a version of it.
How do you optimise for fan-out queries?
You cannot optimise for a query you cannot see, so the work is coverage rather than targeting.
Map the fan-out before you write. Free and paid tools now simulate it, including Qforia from iPullRank, Locomotive's AI Visibility and Coverage Analysis, and WordLift's fan-out simulator. Run your head terms through one and you get 20 to 30 likely sub-queries to plan against.
Build the cluster, not the page. Cover the comparative, pricing, integration and use-case angles as their own pages, each answering one sub-question properly.
Answer directly, near the top. Sub-queries are questions. A page that buries its answer in paragraph nine is harder to lift a passage from.
Make your entities unambiguous. The model needs to know what your product is and what it relates to. Entity optimisation is what makes you retrievable across adjacent sub-queries rather than just your own brand name.
Do not confuse coverage with volume. Filling gaps with commodity content gets you retrieved and not cited. The pages that get quoted carry a number, a result or a position nobody else has.
Fan-out is also why AI visibility work sits alongside SEO rather than replacing it. Retrieval still runs through the index, so the fundamentals hold. What changes is that you now need to be right about a set of questions instead of one, which is the whole basis of how Team 4 approaches generative engine optimisation.
Frequently asked questions
Is query fan-out only a Google thing?
No. Google names it explicitly in AI Mode and AI Overviews, but ChatGPT, Perplexity and Copilot use the same basic pattern: expand the query, retrieve in parallel, synthesise. The implementations differ. ChatGPT runs sub-queries against Bing, while Google uses its own index and graphs.
How many fan-out queries does a single search trigger?
It varies with complexity. Simple factual questions may trigger almost none, while multi-part or comparative questions can generate dozens. Google has publicly referred to issuing large numbers of searches at once for a single question in AI Mode.
Can I see the actual fan-out queries Google used?
Not directly. Google does not publish them. Simulation tools reconstruct a likely set using the same prompting logic, which is close enough for planning, but treat the output as a model of the fan-out rather than a record of it.
Does fan-out mean keyword research is finished?
No, it changes what you do with it. You still need to know what your buyers search for. The difference is that a keyword is now the starting point for a set of sub-questions rather than a target to build one page against.
About Team 4
Team 4 is a B2B SaaS marketing agency in London. We build Inbound Engines for seed to Series B software companies: SEO, GEO, content, paid media and CRO run as one system and measured against pipeline rather than traffic. Mapping the fan-out before building a content plan is now a standard part of how we scope organic work.
