How query fan-out works, and how to use it in your AI search strategy
When someone asks ChatGPT, Gemini or Google's AI Mode a question, the system doesn't run that question as a single search. It breaks the prompt into a set of related sub-queries, runs them in parallel, pulls passages from whatever those searches return and writes the answer from the passages it trusts most. That process is called query fan-out, and it's the reason ranking first for the literal query no longer guarantees your content shows up in the answer.
This guide explains how fan-out works, drawing on iPullRank's AI Search Manual and research from our team, and then walks through how to turn the data in queryfanout.com into decisions about what to publish, in which format and where. Every source is linked where it's used, and they're collected at the end if you want to go deeper.
What query fan-out is
Classic search treated a query as a fixed lens into an index: the words you typed were matched against documents, and the best matches were ranked. Generative search treats the query as a seed. The query fan-out chapter of the AI Search Manual describes how the system decomposes the prompt, rewrites it, predicts the follow-up questions you're likely to ask and routes each variant to a different source before reranking and filtering what comes back. Lazarina Stoy calls the result “the map of every related question” a system infers from a single query, in her piece on how AI search platforms expand queries with fan-out.
The practical consequence is that competition moves to the sub-query level. A prompt like “best half-marathon training plan for beginners” turns into searches about schedules, gear, injury prevention, nutrition, pacing and recovery. A page that only answers the head query competes for one of those branches, while a body of content that answers several branches can be cited several times in the same answer.
Not every prompt fans out. Simple factual questions, such as the capital of a country, are answered directly, while complex prompts with several criteria trigger the most expansion. Google's version of the mechanism is described in its Thematic Search patent (US12158907B1), which calls the sub-queries “themes,” and Google has said Deep Search in AI Mode can issue hundreds of searches for a single request. Prompts are also far longer than classic searches: Stoy's research with Similarweb data puts the average AI search prompt at roughly 70 to 80 words, against three or four on Google, which gives the system much more to expand from.
Where fan-out sits in an AI search system
Fan-out is one stage of a retrieval-augmented generation (RAG) pipeline, and the search architecture chapter of the manual tears down how each platform implements it. In Google's AI Overviews and AI Mode the flow runs through five stages. The system first interprets the query in several forms at once (the words themselves, a dense embedding, the entities involved and the task the user is trying to complete), then fans it out, retrieves from several sources, filters and deduplicates the results and finally has Gemini write the answer from the passages that survive.
Two details from that teardown matter for strategy. Retrieval is hybrid on most platforms, combining keyword matching (BM25) with vector search, so keyword optimization still matters for lexical recall even as semantic coverage becomes the bigger lever. And the platforms differ in where they look. ChatGPT keeps no web index of its own: it sends the queries it generates to Bing's API and fetches the pages it picks at runtime, which makes robots.txt access, load speed and server-rendered content conditions for being included at all.
Agentic RAG: when fan-out becomes a loop
The pipeline above still describes one pass: retrieve, then write. In Beyond RAG: Why Every AI Search Platform Is Now Agentic, Mike King argues that single-shot retrieval has given way to agentic architectures across the major AI search platforms, and that this changes what fan-out is. Naive RAG broke on compound questions that one embedding can't capture, couldn't recover from a bad first retrieval, had no way to send a rate lookup or a calculation to the right tool, and shipped answers without anything checking them.
An agentic system adds four things. It plans, decomposing the query into sub-queries and deciding which tool should handle each one before it retrieves anything, so the fan-out becomes a research plan rather than a one-off expansion. It uses tools: vector and keyword search, structured APIs, code execution, live browsing and MCP servers. It iterates, retrieving, reading and retrieving again across several hops. And it reflects, grading its draft for sufficiency, contradictions, freshness and source diversity before deciding whether to answer or go back for more. One user query can trigger somewhere between five and twenty sub-retrievals.
Two stages in that loop deserve the most attention. The router is the component King calls the most underrated: if your topic has no tool surface the router can call, it gets skipped. The critic is where the most content is dropped, because it checks the draft for sufficiency, contradictions, freshness and source diversity and cuts whatever doesn't hold up. King maps Google patents onto each stage, from generating query variants (US11663201B2) and answering from custom corpora through generated API calls (US20240362093A1) to stateful chat memory (US20240289407A1), pairwise ranking of passages by an LLM (US20250124067A1) and generative summaries (US11769017B1), while noting that patents show which subsystems Google considers strategic rather than exactly how it implements them.
The loop is also mostly invisible. Deep research modes in ChatGPT, Gemini and Perplexity show parts of their plan, but no platform exposes every stage, and citation tracking only sees the passages that survived. That's why simulating the fan-out matters: it gives you the planner's side of the picture, which is the part you can act on.
The three stages of fan-out
The manual breaks fan-out itself into three stages. Each one filters what can end up in the answer, and each one creates a different opportunity for your content.
Expansion and latent intent
The system classifies the intent, fills in the details the user stated, such as distance and audience, and infers the ones they didn't, such as time frame, fitness level and goal. It projects the query into embedding space to find neighboring concepts and predicts follow-up questions. The output is a dozen or more sub-queries; the manual's running example produces roughly 15 to 20.
Routing
Each sub-query goes to the source and retrieval method most likely to answer it. In the manual's table, a plan favors coaching sites and long-form schedules, a checklist favors reviews and comparison tables, a routine favors video with transcripts, and a definition favors reference sources. Higher-priority sub-queries get more retrieval passes.
Selection for synthesis
The system ranks passages, not pages. Extractability is the first gate, followed by evidence density, clearly stated scope, authority and corroboration, freshness and safety. Hundreds of candidate chunks are narrowed to about a dozen units that make it into the answer.
That shape is why the manual argues that intent coverage replaces keyword coverage, that multimodal parity (text, tables, downloads and video with transcripts) is now table stakes, and that relevance engineering has to happen at the level of the passage as well as the page. Good content gets left out when its key facts are buried far down a long page, locked inside an interactive element or published in a format the system can't parse.
The sub-query types you'll see in queryfanout.com
Every simulated sub-query in queryfanout.com is labeled with one of six types, so you can see which kinds of expansion a model leans on for your prompt:
Stoy's fan-out article lists the wider set of variants these systems generate: equivalent phrasings, follow-ups, generalizations, more specific versions, canonical forms, translations, entailments and questions put back to the user to clarify intent.
Why the same prompt fans out differently for different people
Traditional search personalizes downstream: it reranks a fairly stable set of results using your history, location and device. AI search personalizes upstream, at the moment the sub-queries are generated. As Stoy explains in How AI Search Personalizes Fan-Out Queries Using Memory and Context, two people who type the same prompt can trigger different sub-queries, retrieve different documents and get different answers, and a brand the system judges irrelevant to someone's context may never enter the consideration set at all.
How much context a platform uses varies widely. Perplexity works mostly from the current session and memories you've saved, ChatGPT and Claude draw on saved memories and chat history, Gemini can use Gmail, Photos, YouTube and Search history when Personal Intelligence is turned on, and Copilot grounds its answers in the Microsoft 365 Graph. The deeper the inference goes, the less of it you can see.
We think this is the year hyperpersonalization comes back to the core results page, not just the generative surfaces. When that happens, visibility becomes a distribution across contexts rather than a single rank. Expansion can also drift: Stoy's fan-out article shows how adjacent concepts and signals carried over from earlier sessions can pull an answer away from what the user is asking right now. The response is to cover the range of contexts your audience actually brings, research which platforms and prompt styles they use, and describe your entities consistently everywhere they appear, on your site and off it.
How to use queryfanout.com data in your strategy
The workflow below adapts the process in Francine Monahan's Query Fan-Out in Practice, which takes one retirement-planning query all the way to an omnimedia content plan, to the data queryfanout.com gives you. It also follows the practical advice Mike King and Garrett Sussman walk through in Query Fan Out: What it is, How it Works, and What You Should Do About it on iPullRank's YouTube channel.
- 1RunBoth models, several passes, a real-search pass
- 2PrioritizeStable, shared and confirmed sub-queries
- 3Map topicsClusters become hubs; find the gaps
- 4Pick formatsQuery class and routing format
- 5Check citationsWho gets cited today, and where
- 6Build & distributeExtractable passages, omnimedia
- 7MeasureSub-query recall, citation stability
- Re-runFan-outs change when models update, so repeat on a schedule
Run the fan-out on both models
Start with the prompt your customer would actually type, not your head keyword. Monahan's example starts from a full question, “What's the best way to save for retirement?”, written the way her persona would ask it. Run Gemini and OpenAI side by side with several simulated passes and at least one real-search pass. Each pass is a separate generation, so how often a sub-query comes back tells you how stable it is, and the real-search pass shows what the model searched for when it answered with the web.
Separate the signal from the noise
A single pass mixes durable sub-queries with one-off variations. Sort them by two things: how often they come back across passes, and who agrees on them. Sub-queries both models generate repeatedly, and those confirmed by a real search, are the branches you most need to win. Sub-queries that appear once, from one model, are worth noting but rarely worth building for.
Confirmed by both models or a real search, but it doesn't come back every pass. Cover it inside a hub page and keep watching it.
In the tool: ✓ predicted by the simulation, low frequency badge
Comes back pass after pass and is confirmed by the other model or a real search. Give it a dedicated, extractable answer first.
In the tool: High frequency, in “Both models”, ✓ predicted
One model, once. Note it, and build for it only if it matches a context your audience really has.
In the tool: 1× badge, hidden by the 2× minimum by default
Stable for one model only. Prioritize it if that platform matters to your customers.
In the tool: High frequency, in “Only Gemini” or “Only OpenAI”
Map the topics and find your gaps
Once you know which sub-queries matter, group them. The manual recommends building intent-complete hubs rather than chasing individual phrases, and the clusters in a fan-out are a good first draft of those hubs. Check each cluster against the pages you already have. A cluster with no page that answers it cleanly is a branch of the answer you can't be cited for.
If you're working with a large library, Monahan's article also lays out a relevance engineering workflow for pruning: embed every page, compare it to topic centroids with cosine similarity, layer in recent clicks and publish dates, and decide what to keep, revise or retire.
Pick the format for each sub-query
Routing decides which sources and formats can answer a sub-query, so format matters as much as topic. In Monahan's analysis the formats that earned citations were listicles, tables, interactive tools and calculators, step-by-step checklists and video, and the manual adds that video without a transcript and PDFs that are hard to parse may be skipped entirely. Every sub-query in queryfanout.com carries two clues: a query class, which says what kind of answer it's after, and a routing format, which is the kind of page the model expects to answer it.
| Query class | The searcher wants | Build | Routing formats you'll see |
|---|---|---|---|
| Short fact | a number, date, name or price | The answer in the first lines, plus a spec or data table | product_spec_sheet · pricing_page · dataset |
| Definition | what something is | A concise definition block or glossary entry | glossary/definition · faq_page |
| Instruction | how to do something | Numbered steps, a checklist, a video with a transcript | how_to_steps · checklist · tutorial_video/transcript |
| Comparison | the best option, or X vs Y | A comparison table, buyer's guide or review roundup | comparison_table · buyers_guide · review_roundup |
| Reason | why something happens | An explanatory passage with evidence and sources | web_article · faq_page |
| Consequence | what happens as a result | Outcome passages, and a calculator where numbers matter | web_article · calculator/tool |
| Yes / no | a yes-or-no answer | The answer in the first sentence, then the conditions | faq_page · forum/qna |
| Other / navigational | a brand, product or place | Entity pages with consistent facts and structured data | product_spec_sheet · map/local_pack |
Treat this as a starting point rather than a rulebook. Check the routing format the model chose and the live results for each priority sub-query before you build.
See who gets cited today
For your priority sub-queries, look at what's winning now. Open the Google and AI Mode results for each one, and ask your original prompt in ChatGPT, Gemini and Perplexity. Monahan groups the cited sources into four types: established institutions (in her example, financial firms), official and educational sources, reputable media, and community sites like Reddit and Quora. Where one type dominates, you either need comparable authority or a narrower angle that the incumbents don't cover; where community sites appear, that's where your prospects are having the conversation.
Build and distribute the content
Write for selection. The manual's criteria translate into a simple discipline: one idea per passage, the answer before the explanation, specific claims backed by sources, a clearly stated scope and a visible date or version. Use headings, lists and tables so passages separate cleanly from the page around them, and use structured data to make your entities unambiguous.
Then publish beyond your own site. AI answers draw on the whole content ecosystem, so plan shared, owned and earned media together: video with transcripts for instructional sub-queries, useful participation in the communities that get cited, and coverage in the publications the models already trust. Monahan's plan pairs each priority sub-query with an asset, from a pillar guide and an FAQ with a comparison table to a calculator, a YouTube how-to and a LinkedIn article.
The agentic loop raises the bar further, and Beyond RAG lays out what changes. Because answers are assembled from many sub-retrievals, dense topical neighborhoods get cited again and again. Passages have to hold up when an LLM compares them head to head with a competitor's, which favors self-contained passages that name their entities up front and state their scope. Being the clearest page that connects two entities makes you the bridge a multi-hop search passes through. Critics reward content that acknowledges counterarguments and says when advice doesn't apply, and they filter out salesy copy. Calculators, APIs and structured endpoints, exposed through structured data or an MCP server, give the router a tool it can call. And freshness acts as a gate, so keep dateModified, version numbers and “as of” dates current.
Measure and re-run
Rankings for a single query can't tell you whether this is working. The manual proposes input metrics that can: sub-query recall (how many branches of the fan-out you appear in), atomic coverage (the share of your content that meets the extractability criteria), evidence density and citation stability across regenerations. Track citations and brand mentions alongside them, and re-run your priority prompts on a schedule, because fan-outs change when models update.
Citations alone undercount you. King's example in Beyond RAG is a query whose twelve sub-retrievals include your content four times but cite you once, so citation tracking misses about three quarters of your footprint; he estimates citation counts can under-report by a factor of three to ten. He proposes measuring sub-query coverage, the ratio of retrievals to citations, how often your passages survive the critic, how central your pages are as bridges between entities, and whether tools you publish get called. His longer-term answer is distillation: running a smaller, observable agent locally so you can see the stage at which your content drops out.
Adapted from the observable audit in Beyond RAG, with queryfanout.com filling in the plan:
- Pick five high-value prompts and run each one in queryfanout.com, plus the deep research modes of ChatGPT, Gemini and Perplexity.
- Log every sub-query: the simulated ones, the real searches and the plan steps the research modes show.
- Search each sub-query on its own and mark whether your content appears.
- Compare that sub-query coverage with how often you're cited in the final answers.
- Label each miss: no content, too broad, poor chunking, missing schema, no tool surface or out of date.
Watch
Two sessions from iPullRank that cover the same ground on video.
Query Fan Out - What it is, How it Works, and What You Should Do About it for AI Search Strategy
How one query expands into hidden sub-queries, how retrieval, routing and synthesis decide what gets cited, a walkthrough of fan-out tooling, and what to change in your content.
How to Optimize for AI Mode Using Query Fan-Outs and User Context
How AI Mode rewrites and expands queries, why user context and embeddings change the results, and how relevance engineering replaces keyword targeting.
Go deeper
The iPullRank research this guide is built on.
Run it on Gemini and OpenAI side by side, or search everyone's fan-outs in the open index.
Is your brand ready for AI search?
Take the AI Readiness quiz to see where you stand and what to prioritize first.