Query fan-out

Query fan out is the step where an AI engine breaks one prompt into 5 to 15 sub-queries. It runs each one against its sources, retrieves passages, and builds a single answer. Google filed a patent application for it. A page is cited by how often it wins across the sub-query set, not by its rank on the head term.

How query fan-out works

Start with one prompt. The engine does not answer it whole. It splits the prompt into 5 to 15 smaller sub-queries. That split is query fan out. It is spelled two ways: query fan out and query fan-out name the same mechanic. Each sub-query targets one intent or one entity. The system retrieves passages for all of them at once, then synthesizes one answer. It cites the sources it used, 2 to 7 on average, per Profound's tracking across the LLMs it watches.

Google owns the name. It calls the step query fan-out in its AI Mode documentation. It walked through the mechanic on the I/O 2025 stage. Elizabeth Reid, Google's Head of Search, showed it on May 20. Search breaks a question into subtopics and fires many queries at once. Google also filed a patent for it: application US20240289407A1, "Search with stateful chat." That filing is public and specific. It is a primary source for how the split works.

Behind the split sits a filter. Retrieval pulls candidate passages. A scoring pass ranks them for relevance, factual density, and source trust. Synthesis rewrites the survivors, and a final pass orders the credits. You earn a citation by surviving the filter, then by being trusted enough to name.

Retrieval-augmented generation , or RAG, is the plumbing under this. The engine grounds its answer in the passages it pulls, then cites them. A clean semantic triple, one subject, one verb, one object, reads as a single fact. It survives that grounding step. So entity SEO and citation share reward tight blocks over long ones. A liftable page also escapes the zero-click trap, because your brand rides inside the answer.

The diagram is plain. One query sits on the left. It fans out into 5 to 15 sub-queries in the middle. Each sub-query pulls its own set of sources. On the right, the engine counts how often each source shows up across the set. Frequency drives the citation. Appear in many sub-queries and you get named. Appear in one and you tend to get dropped.

Why it changes what you optimize

Keyword logic optimizes one page for one query. That logic breaks under fan-out. Your prompt no longer maps to a single results page. It maps to a spread of sub-queries, and each one is a separate contest. You can rank first for the head term and still lose the generated answer. A competitor that appears in seven of twelve sub-queries beats you when you appear in two.

Watch a real prompt. A buyer asks an engine for the best AI visibility tool for a 10-person agency. The engine fans it out. The sub-queries look like pricing comparisons, tool-X-versus-tool-Y, agency reviews, integration questions, and "is this tracking worth it". Rank first for the head keyword and you can still miss most of those. Cover them and you get cited again and again inside the same answer.

Under fan-out logic, the target moves. You stop chasing one position. You start answering the member questions of your topic inside the page. Each self-contained block competes in its own sub-query. That block has a name and a spec: the answer capsule, 40 to 60 words, directly under a question-shaped heading. Write one under every H2 and each heading enters a different sub-query. The spread is wide. Profound's AEO playbook cites Nectiv's study of roughly 60,000 Google fan-out queries. The average prompt fans out into about 9 sub-queries. That is nine separate contests to enter.

Related terms

Query fan out lives in the GEO / AEO glossary of emerging terminology. It pairs with the answer capsule , the block you write to win each sub-query. It sets up AI share of voice , which counts how often you get named across the answers it produces. Sibling terms sit close by. GSVO tracks the same visibility, agentic commerce covers the AI shopping agents that act on these answers, and prompt volume counts the demand behind them. For the full mechanics, the retrieval and filtering stages behind the split, see the fan-out section of the generative engine optimization playbook.