The short version
- Many AI engines do not answer purely from memory. They run their own background searches first, then build the answer from what those searches return. That process is called query fan-out.
- The queries an engine chooses to run are a direct signal of how it frames your category, and every one of them is a topic you could be ranking for, or getting cited on.
- Coverage of these queries is not consistent across engines. Some disclose the real text, some disclose only a count, some disclose nothing, and a page scraped off a search results feed can never disclose any.
- Be skeptical of any "fan-out" number that looks suspiciously large. If it is not clearly labeled as provider-disclosed, assume it was reconstructed.
Ask ChatGPT, Gemini, Perplexity or Google AI a question that depends on anything current, a comparison, a price, an opinion the model was not trained on, and most of the time it will not answer purely out of memory. It runs a search first. Often more than one. It rewrites your question into several search-engine queries of its own choosing, sends them out, reads what comes back, and only then writes the answer you see.
That process has a name: query fan-out, sometimes called query expansion. One prompt fans out into several searches. It is not a rumor or a reverse-engineered theory, it is a documented part of how retrieval-augmented answers get built, and for a marketer it is one of the most useful, most misunderstood signals in GEO.
Why the fan-out queries matter more than the answer
Most GEO work starts from the prompt you think a buyer would type, and checks whether the answer mentions you. That is useful, but it treats the model as a black box. The fan-out queries open the box a crack. They are the model's own reformulation of your category, written by the model, for the purpose of actually finding an answer, not written by a marketer guessing at buyer intent.
That makes them a different, sharper kind of keyword research. A traditional keyword tool tells you what people typed into a search box. A disclosed fan-out query tells you what an AI model decided it needed to know to answer a question about your category: which comparisons it reaches for, which modifiers it adds ("for agencies", "pricing 2026", "vs [competitor]", "reviews"), and which related questions it treats as part of the same intent. Every one of those is a topic you can choose to own, ignore, or lose to whoever already owns it.
The honesty problem
Here is where it gets messy, and where a marketer needs to be careful about what they are looking at. Not every AI engine tells you what it searched for, and the ones that do vary in how much they tell you. There are three real disclosure levels, and a fourth situation that is not disclosure at all.
What "fan-out data" can actually mean
Read any fan-out report against this before you trust a number
Illustrative categories, not a per-tool ranking. The point is the shape of the problem: three genuinely different levels of evidence, plus one situation where no evidence can exist.
The practical trap is that a lot of "fan-out" reporting quietly blends these together. A tool can show you the real, disclosed queries for one engine sitting in the same list as invented, plausible-sounding queries for an engine that discloses nothing, with no visual distinction between them. The reconstructed ones are often written to look thorough: dozens of variations, long tails, comparisons you would believe a model might run. They might even be reasonable guesses. They are still not observed, and presenting them as if they were is the actual dishonesty here, not the guessing itself.
One well-documented example, and why it is the exception
Google's Gemini API is a clear, public example of a provider that discloses the real thing: when grounding with search is enabled, the response metadata can include the actual list of search queries the model ran to answer. That is exactly the kind of data worth building a workflow around, because you can verify it is real by reading the same API response yourself.
Other engines are not there yet, or expose less. Some surface a count of searches performed without the text. Some, depending on the mode and the query, run no background search at all, in which case there is correctly nothing to disclose. And Google AI Overviews and AI Mode results, when you are looking at them the way a searcher does, off the results page, are not an API call at all. There is no field for a fan-out query to sit in. Anyone showing you fan-out numbers for an Overview is showing you an estimate dressed up as an observation.
None of this is fixed forever. Providers change what they expose. The honest position is to track what each engine actually discloses today, not to memorize a permanent list, and to say plainly when a number is not available rather than filling the gap with a guess.
What a suspiciously large number is telling you
Once you know the disclosure levels, a specific red flag becomes obvious: a "fan-out" count that is large for an engine you know only discloses a count, or nothing, was not observed. It was generated. A dashboard showing forty fan-out queries for a single prompt, on an engine that has never returned query text to anyone, is not more thorough than a dashboard showing four real ones. It is less trustworthy, because the four are verifiable and the forty are not.
Ask the same three questions of any tool making this claim to you:
- Is this the actual text the provider returned, or an estimate? If the tool cannot answer clearly, assume estimate.
- Does the count match what the engine is known to disclose? A number well above what the provider's own API exposes is a reconstruction, however well it is dressed up.
- Is coverage labeled per engine, or is everything shown as one undifferentiated list? A single blended list is where real and invented queries go to hide together.
Turning real queries into work
Where the disclosed queries are real, they are pure content opportunity, and the read is straightforward:
- Find the queries about your category, not just your brand. The most useful ones are often not searches for you by name, they are the comparisons and modifiers the model reaches for to answer a broader question, which is exactly where a rival can quietly own the framing.
- Check which brand each query actually surfaces. If a disclosed query keeps returning a competitor's page, that is a specific, addressable target, not a vague "we should rank higher" goal.
- Treat repeated phrasing as the model's vocabulary for your category. If several disclosed queries independently reach for the same comparison or the same modifier, that phrasing is worth matching on the page, because it is what the model is already searching for.
- Track it before you write for it. A query worth owning is worth watching over time, the same way you would track a keyword's rank, so you can tell whether the work actually changed anything.
This is also where the reporting-integrity habit matters most: when you roll disclosed queries up into a rate, like "we are named in 40% of the fan-out searches about this topic", that is a proportion, and it needs the sample size behind it (a handful of disclosed queries is not the same claim as a few hundred). Read it as a daily-refreshed signal you watch move over weeks, not a one-time scan you screenshot and move on from.
How llemmy handles this
llemmy's Fan-outs view shows the background searches AI engines disclosed while answering your tracked prompts, and it labels every engine honestly by what that engine actually exposes rather than presenting one blended number. Where an engine discloses the real query text, you see the real query text. Where an engine discloses only a count, you see a count and nothing invented to fill in the rest. Where nothing is disclosed, including for answers read off a search results page, the view says so instead of guessing. Each real query is one click from being added to rank tracking, so a fan-out finding turns into a tracked topic instead of a screenshot. Run a free GEO audit or start tracking free to see what your category's models are actually searching for.
FAQ
What is query fan-out in AI search?
Query fan-out (also called query expansion) is what happens when an AI engine, instead of answering purely from what it already knows, runs a set of its own background searches to gather current information before writing an answer. One user prompt fans out into several search-engine queries the model chooses on its own, then the answer gets built from what those searches return.
Do all AI engines show you their fan-out queries?
No. Coverage differs by engine and changes as providers update their APIs. Some expose the literal query text they searched, which is ground truth. Others report only a count of background searches without the text. Some disclose nothing about the process at all. Answers scraped off a search results page, like Google AI Overviews, are a different case: there is no API response to carry the data, so no fan-out queries are ever actually disclosed for them.
Why do some fan-out reports show suspiciously large numbers?
Because a tool can reconstruct or infer plausible-looking queries for an engine that never disclosed any, and present them next to the real ones without saying which is which. A blended list looks more impressive than an honest, shorter one. Treat a large fan-out count for an engine known to disclose only a count, or nothing, as a strong sign the underlying queries were invented rather than observed.
How should marketers use real fan-out queries?
Treat the disclosed queries as a direct read on how the model frames your category: the comparisons it reaches for, the modifiers it adds, the questions it treats as related. Track which of those queries mention your brand versus a competitor, and use the ones you are missing to guide what to track, what to write, and which specific claim to own on the page.
By the llemmy team, July 2026. Related reading: From crawl to citation: how an AI answer gets built, The AI Citation Gap, and How llemmy measures AI visibility (and what we don't claim).