Type a question into ChatGPT, Claude, Perplexity, or Google AI Mode and it looks like you ran one search. You didn’t. Behind the screen, the system quietly turned your single question into a batch of related searches, ran them in parallel, and assembled the answer from whatever came back.
Google has a name for this. Query fan-out. It’s one of the more important shifts in how search works right now, and it’s also one of the more over-explained, so this note sticks to what’s actually documented, shows you how to see it happening for yourself, and is honest about what it does and doesn’t change.
One Query, Many Searches
The cleanest way to understand fan-out is to look at how query handling has evolved.
Search started one-to-one. One query matched against documents, and the pages that best matched that exact query came back. Type “Sydney plumber” and you got pages about Sydney plumbers.
Then it became many-to-one. Google learned that “Sydney plumber,” “plumber in Sydney,” and “plumbing service Sydney” are the same intent, and resolved them to the same set of results. Different phrasings, one answer set.
Fan-out flips that to one-to-many. One query becomes many searches. Ask AI Mode “how do I start a podcast” and it doesn’t just search that phrase. It breaks the question into sub-queries covering equipment, hosting, branding, guest sourcing, promotion, and more, runs them at once, and synthesizes one answer from the combined results.
That’s the shift. You’re no longer competing for one query. You’re competing across the whole spread of sub-queries your topic generates.
Where This Is Documented
Google named the technique publicly. At Google I/O in May 2025, Head of Search Elizabeth Reid described AI Mode breaking a question into subtopics and issuing “a multitude of queries simultaneously” on the user’s behalf. That’s Google’s own language, not a third-party guess.
Two Google patents describe the machinery. The first, “Thematic Search” (US12158907B1, filed December 2024), describes a system that organizes search results into themes and sub-themes, generating narrower queries from an initial one. The patent’s own example is a search that gets broken into themes like neighborhoods, cost of living, and things to do. It uses a large language model to summarize each theme.
The second, “Search with Stateful Chat” (US20240289407A1), describes the query generation itself: an LLM receives structured instructions and produces alternate queries that emphasize different intents, such as comparative queries or exploratory queries.
Google gets credit for naming and documenting the technique, but fan-out is not a Google-only phenomenon. ChatGPT, Claude, Perplexity, Copilot, and the other AI answer tools all work this way: they take your prompt, expand it into multiple searches, gather results, and build a single answer from them. Google’s patents are the most detailed public description of the mechanism, but the behavior shows up across all of these systems. If you care about being cited in AI answers, and at this point that means more of your audience than Google AI Mode alone, fan-out is the process deciding whether your content gets pulled in.
Two honest caveats, because they matter. Google does not confirm that any given patent is running in production, and Google’s patents typically describe several possible implementations, not one fixed method. So the patents tell you how Google has thought about this and what’s plausible, not a guaranteed blueprint of the live system. As always, treat patents as strong evidence, not gospel.
The Types of Sub-Queries
Across the patents and the analysis of real fan-outs, the sub-queries tend to fall into recognizable buckets. You don’t need to memorize these, but they’re useful for understanding what the system reaches for:
- Related queries: adjacent subtopics that add context
- Implicit queries: unstated concerns the system predicts you have
- Comparative queries: side-by-side evaluations of options
- Recency queries: time-sensitive versions that pull current information
- Reformulations: the same intent phrased differently
- Contextual queries: angles personalized to location, history, or behavior
- Next-step queries: what people typically do after the initial search
The throughline is that the system isn’t just rephrasing your query. It’s anticipating the fuller set of things you probably want to know, including things you didn’t ask.
How Deep It Goes
The volume varies with how complex and how ambiguous the query is. Analyses from groups like Seer Interactive and Nectiv put the average around nine to eleven sub-queries per prompt, with most prompts triggering somewhere between five and eleven. A meaningful share go higher, past twenty in some cases.
The extreme end is deep research modes. ChatGPT’s Deep Research has been observed running into the hundreds of searches for a single request. One documented example of a simple shopping query spun up over 400.
The pattern is consistent: the more dimensions a query has that need resolving, the deeper the fan-out goes. A vague query forces the system to either ask you a clarifying question or go gather context on its own, and gathering context means more searches.
You Can Watch It Happen
Here’s the part most coverage skips. You don’t have to take anyone’s word for what fan-out looks like. For ChatGPT and Claude, you can extract the actual sub-queries the model ran.
I built a bookmarklet that does this for you and have shared in the past. Open a ChatGPT or Claude conversation where the model searched the web, click the bookmarklet, and it pulls out the search queries the model actually ran, the reasoning it worked through, and the sources it touched. The latest version separates sources the model likely accessed during the search from sources it actually cited in the answer, which is a distinction worth having, since what a model looked at and what it used are not the same thing. The updated bookmarklets are here.
This is worth being precise about, because it’s a real distinction. The bookmarklet reads the actual queries from the conversation, so for ChatGPT and Claude you’re seeing ground truth, the real searches the model ran. Other tools work differently. They use a model like Gemini to simulate the fan-out Google AI Mode would likely generate, because Google doesn’t expose its internal queries the way the chat tools do. Both are useful. One shows you what actually happened in a ChatGPT or Claude session, the other predicts what Google would probably do. Don’t confuse a simulation for the real thing, and don’t expect the real thing to be identical run to run, because fan-out is probabilistic and the exact wording changes each time.
What you do with this is straightforward. Run a few real prompts in your niche, pull the actual sub-queries, and look at what the model went hunting for. That tells you the questions your content needs to answer to be useful across the whole fan-out, not just for the headline query.
What This Actually Changes
A lot of fan-out content jumps from “here’s how it works” to “so optimize for fan-out” with advice that’s vague at best.
What changes is the unit of competition. You’re no longer trying to rank one page for one query. Your content is being evaluated, often passage by passage, against a spread of related sub-queries. A page that nails the main query but ignores the obvious follow-ups is now competing at a disadvantage, because the system is actively searching those follow-ups and pulling the best answer for each from wherever it finds it, including from pages other than the one that best covers the main topic.
What doesn’t change is the fundamentals. Fan-out rewards content that genuinely and thoroughly covers a topic and its real sub-questions. That’s the same thing topical authority and comprehensive coverage always aimed at. If you’ve been building content that answers the full set of questions a reader actually has, fan-out works in your favor. If you’ve been publishing thin pages built around a single keyword, fan-out raises the cost of that.
So the practical takeaway is not a new playbook. It’s a sharper reason to do the thing good SEOs already knew they should do: understand the full set of questions behind a topic and cover them well. The difference now is that you can see those questions directly, by pulling the real fan-out queries and reading them.
And the payoff isn’t limited to Google. The same comprehensive coverage that helps you across a fan-out is what gets you cited in ChatGPT, Claude, Perplexity, and Copilot answers too, because they all run this same expand-and-gather process. The prize is being the source these systems pull from, wherever your audience is asking. Watch a few real fan-outs in your own niche and you’ll have a better sense of what your content is actually up against than any amount of theory will give you.
