I asked ChatGPT to recommend a good gaming monitor at a sane price — one question, one sentence. Behind the scenes it didn’t search for that sentence. It broke it into a pile of separate searches that ran almost in parallel, and built one answer out of the results.
That’s query fan-out, and if you work on AI visibility, it’s one of the most useful things to understand.
What is query fan-out, and how does it work?
Instead of running one search on what you typed, the system reads your intent and generates a set of sub-queries that run in parallel. Google describes the mechanism itself in its official AI Mode post: it uses a query fan-out technique that breaks the question into subtopics and runs many searches at once. Analyses of Google’s patents tie this to a mechanism called Scatter-Gather, where the Scatter step spreads the queries out and the Gather step collects them and assembles the answer.
The range depends on the question. A simple query might split into two or three searches; a typical one splits into a handful, up to a few dozen; and in Deep Search modes Google itself talks about dozens to hundreds of searches for a single question — all in roughly the time of one normal search.
You’ll also see this written as query fanout or query fan out, and described as search fan-out or query expansion. I stick with query fan-out because it’s the precise term, but they all describe the same thing.
Which search engine do the queries run on?
Easy to miss, but these searches don’t run on some magic AI index — they run on ordinary search engines. And ChatGPT isn’t limited to Bing: as of 2026, it’s reported that OpenAI also pulls Google results via SerpApi, a third-party service that scrapes the SERP, after Google refused to give OpenAI direct index access.
What that means for you is direct. AI answers sit on top of Google and Bing search results — and probably other sources that shift over time — and your organic ranking there still feeds your AI visibility. If the source is the SERP, your SEO still feeds the system. GEO is built on SEO; it doesn’t replace it.
AI picks a passage, not a whole page
AI retrieval works at the passage level, not the whole-page level. The reason is cost: the fan-out runs dozens of searches, each returning dozens of results, and there’s no way to pull and feed the model the full content of all those pages. That’s far too many tokens, and it doesn’t fit the model’s context window.
Roughly, content is broken into passages, and each sub-query pulls the passages closest to it in meaning — not the page that ranks highest for a keyword. So one focused paragraph from a small site can beat a giant guide, if it answers a specific sub-query directly. A passage that answers several sub-queries at once — say, both a model’s price and how it compares to another — is likely more useful than one that answers just one.
Why isn’t #1 on Google always cited in AI?
Here’s one of the most important gaps to understand. Even if you rank first and show up in the results the queries return, that doesn’t guarantee your passage gets picked for the final answer. The fan-out puts far more passages into the running than the model ends up citing, and the synthesis keeps only a few. Good organic ranking gets you a ticket to the contest; it doesn’t guarantee a citation. Being retrievable isn’t enough — your passage has to be the one that gets chosen.
What a real query fan-out looks like
To not leave it as theory, I took one prompt — “recommend a good gaming monitor for new games at a sane price” — and looked at what ChatGPT actually searched behind the scenes. One prompt produced a whole fan of searches:
best value 27 inch 1440p 180Hz IPS gaming monitor 2026 official specs
site:ksp.co.il gaming monitor 27 1440p 180Hz IPS
site:ivory.co.il gaming monitor 27 1440p 180Hz IPS
Gigabyte GS27QA official specifications
Dell Alienware AW2725DM official specifications
AOC Q27G4XF official specifications
site:zap.co.il GS27QA AW2725DM Q27G4XF
site:ivory.co.il "Q27G4XF" price
site:ksp.co.il "GS27QA" price
site:ivory.co.il "AW2725DM" priceA few things stood out. The fan came out multilingual, with specs and model names in English alongside local searches; the model went straight to sites it already knows in the category, using site:; and it pulled candidate models from its own knowledge, then ran real-time price checks on them. In practice it worked in two stages — first broad research to find candidates, then focused price searches.
Worth remembering what this test does not prove. It’s one run, one category, one moment; it doesn’t represent all of ChatGPT, and not every question produces a fan like this. What it does show is what a real fan looks like, and why it’s worth sampling many of them instead of trusting a single run.
Query fan-out is keyword research for the AI era
Keyword research used to be the foundation of all SEO work: check what people search, then build content that answers exactly that. Query fan-out is the parallel for the AI era.
Instead of guessing what the user will type, you see exactly what the AI searches when someone asks about your field, and which sub-queries, sites, and entities — the brands, products, and concepts the model recognizes in your space — it assembles into the answer. That’s your target list, straight from the system.
An entity is a brand, product, person or concept that search engines and AI models recognize as one distinct thing, with a name, attributes and relationships to other entities.
And notice the reversal. Today we search in natural language, in full prompts the way we talk, and the AI is the one breaking them into short keyword searches with operators like site: — exactly the way we used to search, before the AI did it for us. The AI inherited the old craft of searching, and keyword research just moved to the other side.
How to check your site’s fan-out
The most useful thing a fan-out gives you is the list of queries themselves. They tell you exactly what the AI searches when someone asks about your field, and from them you derive what to cover.
But there’s a catch worth knowing: these queries aren’t deterministic. The same prompt, run twice, can return different fans. There’s model randomness, the effect of personalization, of time, and of a model version that keeps updating — and one sample simply isn’t enough to draw conclusions from.
So the work is to monitor and sample, a lot. Run the prompts that matter to you again and again, collect the queries from each run, and look at the aggregate: which sub-queries repeat, which sites get site:, and which entities keep surfacing.
What shows up consistently is what’s stable enough to build on. Even if a single run is noisy, the aggregate is stable, and it reveals the sub-intents, the entities, and the sources the AI leans on in your field. It’s a work map based on what AI actually does, not on a guess.
You don’t have to do all this by hand. If you have a subscription to an AI-visibility tool, it monitors your important prompts over time and collects the query fans automatically. And if you don’t, a lot of SEOs just install a Chrome extension that records the query fan-out while they run the AI themselves, reading the real searches the model runs in the background. More manual, but the same data without a subscription.
I actually built a Chrome extension for ChatGPT that does exactly this. Open it next to ChatGPT, run your prompts as usual, and it records the query fan-out, the sources, and the citations in real time, and keeps everything local on your device. Free, and available on the Chrome Web Store.
From there, the check itself, on the whole fan and not a single query:
- Break down and sample: run the core prompts several times and read the searches the model actually ran, not your guess of them.
- Check coverage at the passage level: for every repeating sub-query, do you have a passage that answers it directly? Is it retrievable? And if you’re not cited, who is?
- Mark the gaps: where you have no passage, where the passage is weak, and where a competitor holds the answer.
- Write retrievable passages: a question-shaped heading, a direct short answer, a consistent entity, so an AI system can pull the passage without parsing the whole page.
- Strengthen the sources the model already knows: if you’re one of the sites it reaches for with site:, make sure your product, price, and spec pages are accessible, structured, and consistent.
And what not to do: don’t write forty pages around every synonym, don’t draw conclusions from a single run, and don’t confuse showing up in the results with being cited. The query fan-out is a signal of direction, not a shopping list for content.
Key points
query fan-out isn’t a buzzword — it’s how AI Mode and ChatGPT already work today: one prompt, many parallel searches, one synthesized answer. For you it means you’re competing on dozens of sub-queries, that visibility is decided at the passage level, and that it all still runs on Google and Bing results, so your SEO didn’t go anywhere. It just got another layer.
The queries themselves are the most useful signal you have. They change from run to run, so you need to sample them over time, but the aggregate shows exactly what to focus on. Real AI visibility starts with solid technical SEO, an entity the model recognizes, and passages that answer the questions behind the question directly.
