Query Fan-Out: The Hidden Step That Decides If AI Mentions You
Answer engines split one question into many searches. New research puts numbers on which retrieval signals predict whether your brand gets named, and our own measurement shows the same pattern.
Nobody searches the way AI searches. A buyer types one question; the engine turns it into a handful of its own queries, runs them at once, and writes a single answer out of whatever comes back. That intermediate step, query fan-out, is where your brand is included or dropped, and it happens before a single word of the answer is generated. New research published in September 2026 puts numbers on it: two observable signals in that retrieval path separate the brands that get named from the ones that do not. Brands showing neither were named in under 4% of answers. Brands showing both, in over 90%. We checked it against a run of our own, and the pattern held there too: the brands AI named most were the ones whose own sites it kept reading.
Key takeaways
- Query fan-out means the engine answers several questions it wrote itself, not the one the user typed. Optimizing for the literal prompt misses the queries that actually run.
- Across 34,960 unbranded observations, brands with no presence in the visible retrieval path were named in 2.8% of GPT answers and 3.8% of Gemini answers.
- When the brand's own domain turned up among the retrieved sources, those rates rose to 49.0% and 58.4%.
- When the engine also searched for the brand by name in its own sub-queries, they reached 91.4% and 100%.
- Visibility compounds. A brand that was not mentioned last run and has no exposure this run is named 1.6–1.9% of the time; a brand that was mentioned and still has exposure is named around 80%.
- In our own 24-answer vector-database run, the four brands named in 83–88% of answers owned half of the ten most-cited sources; the two named in none had no site in that list.
- The study is observational and comes from one vendor's customer data, so read it as a strong association, not a proven causal chain.
What query fan-out actually is
Google introduced the term when it announced AI Mode in May 2025: "AI Mode uses our query fan-out technique, breaking down your question into subtopics and issuing a multitude of queries simultaneously on your behalf." Its AI features guide now defines fan-out queries as "a set of concurrent, related queries generated by the model to request more information and fetch additional relevant search results to address the user's query."
The guide's own example: someone asks how to fix a lawn full of weeds, and the engine goes looking for best herbicides for lawns, removing weeds without chemicals, and how to prevent weeds in the first place. The user never typed any of those.
This is why "what keyword should this page target?" is the wrong question in AI search. The engine is not matching your page against the prompt. It is matching your page against questions it invented on the way to answering the prompt — and it will happily answer the user with a brand it found while researching a sub-question you never thought to write about.
Two signals separate the named from the invisible
A paper published on arXiv on 19 September 2026 by Benjamin Tannenbaum analyzed 34,960 prompt-engine observations spanning 2,854 monitored prompts across 75 anonymized projects, with repeated GPT and Gemini runs between June and September 2026. Every observation was unbranded: any run where the brand's name appeared in the user's prompt was thrown out, so this measures unprompted visibility only.
Two things were recorded about each run. Own-domain exposure means at least one stored source URL belonged to the brand's own domain. Branded fan-out means the brand's name occurred in at least one of the sub-queries the engine generated. Crossing them gives four states, and the mention rates in each are not close together.
| Retrieval condition | GPT | Gemini |
|---|---|---|
| Neither signal present | 2.8% | 3.8% |
| Own domain cited only | 49.0% | 58.4% |
| Branded fan-out only | 64.4% | 84.8% |
| Both present | 91.4% | 100% |
Read the top row first, because it is the one most brands live in. With nothing in the visible retrieval path, you are named in roughly one answer in thirty. Everything else in an AEO program is an attempt to leave that row.
The effect survives the obvious objection that some prompts are simply easier than others. Holding the prompt and engine fixed and comparing repeated runs that differed only in whether the brand's domain was retrieved, the stratified common odds ratios for own-domain exposure were 15.3 on GPT and 29.7 on Gemini. This is not an artifact of some brands having friendlier questions; the same prompt behaves differently depending on what the engine happened to pull in.
Treat Gemini's 100% as "nearly always" rather than a law of nature. The direction of the finding matters more than the last decimal.
We saw the same pattern in our own run
Numbers from someone else's dataset are worth checking against your own. Our benchmarks page publishes a recorded Recall run in one category, vector databases: six unbranded buyer questions, each asked four times, 24 answers in all, recorded on 17 September 2026. Here is who got named, next to whether the brand's own site was among the ten sources cited most across those answers.
| Brand | Named in (95% interval) | Own site among top-10 cited sources |
|---|---|---|
| Pinecone | 88% (69–96%) | pinecone.io, 19 citations |
| Qdrant | 88% (69–96%) | qdrant.tech, 26 citations |
| Weaviate | 83% (64–93%) | weaviate.io and docs.weaviate.io, 28 citations |
| Milvus | 83% (64–93%) | milvus.io, 6 citations |
| Elasticsearch | 21% (9–40%) | not in the top ten |
| Chroma | 17% (7–36%) | not in the top ten |
| pgvector | 0% (0–14%) | not in the top ten |
| Redis | 0% (0–14%) | not in the top ten |
The four brands the engine named almost every time are the four whose own pages it kept pulling in. Every brand below them had no page in that list. It is the paper's own-domain row, visible in a single category.
Two cautions, because they matter as much as the table:
- This is 24 answers, not 35,000. The intervals are wide, and the top four overlap, so this run cannot rank Pinecone, Qdrant, Weaviate and Milvus against each other. What it does show is the drop from 83% to 21%, which is outside both intervals.
- We record which sources were cited, not every sub-query behind them, and we list only the ten most-cited sources, so "not in the top ten" does not mean never retrieved. Like the paper, this is an association, not proof that citations cause mentions.
The zeros carry their own lesson. pgvector ships inside Postgres, and the questions asked for vendors and platforms; an extension is neither. That is the branded fan-out problem from the other side: the engine never went looking for pgvector because it does not think of it as an answer to that kind of question. No page on pgvector's site fixes that. Changing how the category talks about it does.
The second signal is the one you cannot write your way into
Own-domain exposure is familiar territory: it is what AI citations measure, and it responds to the usual work of publishing genuinely useful, extractable pages.
Branded fan-out is different in kind. It means the engine decided, unprompted, to go and search for you while researching the topic. Nothing on your website causes that directly. It reflects what the model already believes about your category before it reads a single retrieved page — the association it built from training data and from how often your name travels alongside the topic elsewhere on the web.
That distinction sets where effort goes. Retrieval is earned with content. Being searched for by name is earned off-site: reviews, roundups, forum threads, documentation, press, and the accumulated weight of other people describing your product in your category's language. It is the difference between being findable and being known, and it is why share of voice is worth tracking separately from citations.
Notice also that branded fan-out on its own outperforms own-domain citation on its own, on both engines. A brand the model thinks of but does not retrieve still gets named more often than a brand it retrieves but does not think of.
Visibility compounds, in both directions
The study's most uncomfortable finding is about momentum. Looking at consecutive runs of the same prompt:
- Not mentioned last time, no exposure this time: named in 1.6% of next runs on GPT, 1.9% on Gemini.
- Mentioned last time, exposure this time: named in 80.5% and 83.7%.
Prior visibility carried independent predictive weight. A model using only a brand's history reached an AUC of 0.937 on GPT and 0.917 on Gemini; live retrieval signals alone reached 0.880 and 0.840; together, 0.963 and 0.942. History was the stronger of the two.
Absence, in other words, is sticky. This is the same pattern visible in citation volatility from the other direction: engines keep returning to sources and names they have already settled on, and the churn happens in the tail. The practical reading is that a brand starting from zero should expect a slow start followed by acceleration, and should judge progress on trend rather than on whether this week's check found a mention.
What the study does not show
Worth being straight about the limits, because they are real:
- It is observational, not causal. The author says so explicitly: the model is "predictive and observational, not a causal description." Getting your domain retrieved and getting mentioned may both follow from the same underlying relevance. Nobody ran the experiment where a page is injected into retrieval to see what happens.
- Own-domain citation is an incomplete measure of exposure. The paper flags that third-party pages may carry equal or greater influence without ever appearing in the stored source list. The 2.8% row is not "the engine saw nothing"; it is "the engine saw nothing we could observe."
- The data comes from one vendor's customers. The observations are drawn from monitored projects on the Aiso platform, whose founder wrote the paper. That is a self-selected set of brands already investing in AI visibility, and the vendor has a commercial stake in the conclusion.
- Two engines, one season. GPT and Gemini over four months. Perplexity, Claude, and Copilot are not in it, and engine behavior moves fast.
None of that sinks the finding. An effect that survives holding the prompt and the engine fixed across repeated runs is hard to explain away. But it argues for treating these numbers as the shape of the thing rather than as constants to plan a budget against.
What to do with this
Four things follow, and none of them are new tactics so much as a reordering of old ones.
- Cover the cluster, not the query. If the engine is issuing eight sub-queries to answer one prompt, a single page aimed at the prompt is competing for one of eight slots. Write the definition, the comparison, the pricing question, and the objection as real pages. Google is explicit that this should not mean writing fragments for machines: keep it normal content, organized with normal headings, written for readers.
- Work the off-site side deliberately. Branded fan-out is bought with presence in the places your category is discussed, not with more pages on your own domain. Track where competitors are named and you are not; that is the map. Our guide to AI competitor analysis covers the method.
- Expect a lag, and measure for it. If prior visibility predicts current visibility, the first weeks of any push will look like failure. Set the review window in months and watch the direction.
- Instrument the retrieval path, not just the answer. Mention rate alone cannot tell you why you were left out. Recording what the engine retrieved, and where available what it searched for, is what turns a missed answer into a fixable gap.
How to audit your own fan-out coverage
- Start from a real buyer prompt, phrased as a customer would phrase it, with your brand name left out.
- List the sub-questions it implies — five to ten related queries covering definitions, comparisons, pricing, use cases, and objections.
- Check what the engine actually retrieved, running the prompt with web search on and recording every cited source, plus the sub-queries where the engine exposes them.
- Mark which sub-questions you cover with a page that answers them directly, and which of those pages showed up in the retrieved set.
- Close the gaps with real pages written for readers, and refresh the ones that already exist.
- Track it on a schedule, because a single run is a snapshot and retrieval shifts day to day.
Recall does this part automatically: it runs your prompt set on the answer engines you choose, among them ChatGPT, Gemini and Perplexity, on a schedule; records every source cited at the moment it is cited; and reports each rate with its sample size and interval, as in the table above. That shows you which of the four rows you are living in for each prompt, and whether a change has actually moved you out of it. See a recorded run before you decide.
Check your own category. Pick ten unbranded prompts your buyers ask, run them repeatedly, and look at what the engines retrieved on the runs where you were not named. The gaps repeat.
For the fundamentals underneath this, start with answer engine optimization, then see how to track your brand in AI search.
Frequently asked questions
What is query fan-out?
Query fan-out is the step where an AI answer engine breaks your question into subtopics and issues several related searches at once, then writes one answer from everything it retrieves. Google describes it as 'a set of concurrent, related queries generated by the model to request more information and fetch additional relevant search results to address the user's query.'
Does query fan-out mean I should target more keywords?
Not exactly. Fan-out rewards covering a topic completely rather than building one page per keyword, because the engine is searching for the sub-questions around your topic, not the phrase the user typed. Depth across a cluster of related questions beats narrow pages aimed at single queries.
Does getting cited mean I get mentioned in the answer?
Usually, but not always, and the gap is large. In a September 2026 study of 34,960 unbranded observations, brands whose own domain appeared among the retrieved sources were named in 49.0% of GPT answers and 58.4% of Gemini answers, against 2.8% and 3.8% when neither signal was present.
What is branded fan-out?
Branded fan-out is when the engine puts your brand name into one of the sub-queries it generates for itself. It signals the model already associates your brand with the topic before it reads any sources, which is why it tracks with much higher mention rates than retrieval alone.
How do I influence query fan-out?
You cannot edit the sub-queries, so you work on the two things they feed on: being retrievable for the sub-questions around your topic, through genuine coverage and extractable content, and being known well enough in your category that the engine searches for you by name, which comes from third-party mentions rather than your own pages.