One question, many searches

What is query fan-out - and why does it break keyword thinking?

Before an AI engine answers you, it searches itself first. It rewrites your question into a set of internal queries you never see, then builds one answer from whichever pages win those searches.

28 Labs · August 2026

Query fan-out is the process AI answer engines run before they respond. Instead of searching for your exact words, the engine breaks your question into multiple internal sub-queries - different angles, comparisons, follow-up questions - runs each one against the web and its own indexes, and synthesizes a single answer from whichever pages win those searches. Google names this technique explicitly in its AI Mode and AI Overviews documentation. Microsoft shows a visible trace of the same idea in Bing Webmaster Tools, where it calls the sub-searches "grounding queries." Your question and the engine's search list are rarely the same words.

How fan-out actually works

A user doesn't type search-engine keywords into ChatGPT, Gemini, or AI Mode. They type a real question, in their own words, often with several sub-questions folded into one sentence. The engine's job is to answer that whole question well, and one web search usually can't do it.

So the model decomposes the question first. Google's own documentation for AI Overviews and AI Mode describes this directly: the system "may use a 'query fan-out' technique, issuing multiple related searches across subtopics and data sources" before composing a response. It pulls from the live web, Google's Knowledge Graph, and specialized indexes like Shopping, depending on what the question needs.

1
User asks one real question
2
Engine rewrites it into sub-queries
3
Each sub-query retrieves and ranks pages
4
Engine synthesizes one answer from the winners
The sub-queries in step 2 are invisible to you. You only ever see the finished answer in step 4.

A worked example

Take a real buyer question: "best mattress for back pain in Singapore." A person asking that has several implicit questions bundled together. An engine using fan-out might split it into searches close to:

Each of those runs as its own retrieval. A page that ranks nowhere for "best mattress for back pain" can still win the "mattress firmness back pain" sub-query and get pulled into the final answer. A page that would rank well for the original phrase but says nothing about firmness, price, or Singapore availability can lose all four sub-queries and never make it in. The visible query and the queries that actually decide the answer are different objects.

Where you can actually see it happen

Two vendors have put fan-out in front of site owners, in different forms. Google describes the mechanism in its AI features and your website documentation, but it doesn't expose the actual sub-query list for any given answer. You know the technique exists; you don't get the receipts.

Microsoft goes one step further. Its AI Performance report inside Bing Webmaster Tools shows sample "grounding queries" - the reformulated searches Copilot and Bing's AI answers generated internally and the pages that won them for your site. It's the first place any major engine has let site owners see a real, if partial, trace of fan-out output. We cover what that report shows and where it stops in our review of Bing Webmaster Tools' AI Performance report.

Outside of that, you're inferring fan-out from outcomes, not observing it directly.

Why it differs by engine, and why nobody outside sees the full set

Google, ChatGPT, Gemini, Claude, and Perplexity don't share a fan-out implementation. Each retrieves from different indexes, weighs different source types, and decomposes questions with different models. Reporting on Google's AI Mode has described sub-query counts running from a handful for simple prompts to twenty or more in deep-research modes, but Google hasn't published a fixed number, and the number isn't fixed. Anyone quoting you an exact count or an exact list of "the fan-out queries" for a given brand question is estimating, not measuring - none of these engines publish their full internal query lists, and the sets shift as models and retrieval systems get updated.

Why single-keyword rank tracking mismeasures AI visibility

SEO rank tracking was built for a world with one visible query per search. You tracked a keyword, you watched a position, you knew what you were fighting for. Fan-out breaks that model at the root, because the query the AI actually searches for is usually not the phrase the tracker is watching.

Ranking well for the visible question is no longer the game. Covering the underlying question space is. A brand can hold position one on Google for its category keyword and still be absent from an AI answer, because the answer was built from four sub-queries that page never addressed. Conversely, a page nobody would target for SEO can win the sub-query that actually gets cited.

What's actually measurable

Fan-out sets aren't fully observable from outside the engine, and they change as models update. That's a real limit, and we say so plainly rather than pretend otherwise. What's measurable, consistently, is the answer end: when you ask an engine a real buyer question, repeatedly, on a schedule, does the answer name your brand. That's a fact you can log and re-derive, regardless of which sub-queries produced it. We go deeper on this distinction, and where instrumented AI-visibility tools stop being able to help, in what AI visibility tools can and can't measure.

The honest readFan-out means your competitive set for any single buyer question is wider than any keyword tool shows you, and narrower engine visibility into that set exists than most vendors admit. Chase individual fan-out strings and you're chasing exhaust. Cover the question space a real buyer works through, and measure whether the answer names you at the end of it.

What we do about it

We don't try to reverse-engineer an engine's fan-out list - nobody outside the engine can, reliably. Instead we build a frozen set of real buyer questions per client, covering the question space a buyer actually works through (product, price, comparison, location, "who should I use"), and ask it across ChatGPT, Claude, Gemini, and Perplexity on a repeating schedule. Every count re-derives to a specific question, engine, and answer. That measures the thing fan-out actually determines - whether you get named - without pretending we can see the machinery that produced it.

28 Labs measures how AI engines answer real buyer questions - and what moves the counts. Every number we publish re-derives to specific questions and engines. try28labs.com