AI Mode decomposes a question into multiple parallel sub-searches before synthesising an answer — Google calls this query fan-out and documents eight sub-query types in US Patent 11663201B2, filed 2018 and granted May 2023. The number of sub-queries actually issued per real question — the "8 to 16" range repeated everywhere — has no traceable primary source and no published measurement. The mechanism has a paper trail. The magnitude does not.
- Eight sub-query types are documented — comparisons, specifications, related products, how-to steps, pricing, locations, reviews, definitions — in a granted Google patent and Google's own AI features documentation. Graded traceable.
- The patent was filed in 2018 and granted in May 2023, years before AI Mode shipped. This is not a hasty bolt-on for the generative era; it is an older retrieval approach adapted to a new interface.
- No published, reproducible count of sub-queries issued per question exists. A count of documented types is not a count of searches run, and conflating the two is where the circulating range comes from.
- The strategic claim everyone draws from this — that broader topical coverage is associated with more fan-out capture — is a hypothesis nobody has tested, ours included.
- The sub-queries are issued server-side and no public interface reports them. Everything published about counts is inference from outputs, not observation of the process.
What this page covers, and what it does not
This page is about AI Mode's retrieval mechanism: what query fan-out is, what Google has actually documented about it, and which of the numbers attached to it survive tracing. It is a mechanism page with an honest gap where the magnitude figures should be.
It is not a comparison of what the two Google search surfaces cite. The AI Mode versus AI Overviews source delta is the page that takes up that question — whether the two surfaces return materially different source sets for the same query, and by how much. Nor does it carry AI Overview measurement: Google AI Overview statistics holds those figures, and it deliberately holds numbers this page does not, because nothing equivalent has been published for AI Mode.
What is documented, and what is only estimated
documented sub-query types Google's AI features may generate when fanning out a query — a taxonomy, not a count of searches issued.
from patent filing to grant. The mechanism predates the current AI Mode product by years.
the count of sub-queries issued per real-world question. No independent measurement at scale exists publicly, and no primary source for the circulating range has been found.
- Patent filed2018
US11663201B2
Multi-query decomposition described years before any generative search product shipped.
- Patent grantedMay 2023
US11663201B2
Eight sub-query categories documented in the granted patent.
- Fan-out documented publiclyCurrent
Google AI features docs
Google states AI Overviews and AI Mode "may use a query fan-out technique".
- Sub-query count measuredUnscheduled
Open question
No published, reproducible count with a disclosed method exists.
The timeline separates two kinds of confidence, and that separation is the point of the page. The mechanism has a paper trail. The magnitude has none — which is exactly what the measurement standard exists to keep visible, and why this sits on the tactic evidence scoreboard rather than in a list of things to go and do.
What query fan-out actually is
Google's documentation states that both AI Overviews and AI Mode "may use a query fan-out technique, issuing multiple related searches across subtopics and data sources, to develop a response." So instead of treating your question as one search, the system breaks it into narrower searches, runs each, and assembles the answer from what comes back.
- 1 Comparisons Alternatives, "vs" framing
- 2 Specifications Technical detail, attributes
- 3 Related products Adjacent options
- 4 How-to steps Process, instructions
- 5 Pricing Cost, plans
- 6 Locations Where, availability
- 7 Reviews / opinion Sentiment, experience
- 8 Definitions What is X
That is a different retrieval philosophy from classic search, and it is why so much of the current state of AI search reads differently from ranking work. Being the single best answer to the literal question typed is no longer sufficient. The system may also run an implied comparison search, an implied pricing search and an implied "how does this work" search behind the scenes. Coverage of Google's own description characterises the system as potentially running "ten, twenty, or more" searches behind a single visible question — a qualitative phrase, not a measured count, and the distinction matters below.
Hypothesis Fan-out is best understood as an aggressive extension of classic query expansion rather than a new category. Search systems have long expanded queries with synonyms and diversified result sets so an ambiguous query returns several interpretations; both address the same problem, that a user's typed words underdetermine what they want. What appears new is scope and automation: fan-out generates whole sub-questions the user never typed. That framing is ours, a reading of the published description rather than a claim about the implementation, and its payoff is a prior worth holding — expansion favours documents covering a topic rather than a phrase.
Google's patent documents eight sub-query types behind AI Mode's query fan-out. The specific number of sub-queries issued per real question — the '8-16' figure everyone repeats — has never been measured and published.
Share on XHow does AI Mode differ from AI Overviews?
| Dimension | AI Mode | AI Overviews |
|---|---|---|
| Surface | A dedicated, conversational search mode | A component embedded in classic results |
| Turn structure | Multi-turn; earlier turns can inform later retrieval | Effectively one-shot per query |
| Fan-out | Documented as possible | Documented as possible |
| Overlap in sourcing | Not independently measured. Registered as an open question. | |
| Published citation measurement | None located | Ahrefs top-10 overlap; Semrush AI Overviews study |
Whether the two draw on materially the same retrieval pipeline, a shared-but-tuned one, or genuinely separate systems is something Google has not stated precisely and no independent study has disentangled. Sourcing on the AI Overviews side has at least been studied publicly, in Semrush's study, in Ahrefs' top-10 analysis and in reporting on that overlap dropping. Nothing equivalent exists for AI Mode. Click behaviour is in the same state — studied mostly for AI Overviews, in Ahrefs' click-loss update and on the CTR statistics page.
Where the "8 to 16 sub-queries" figure goes wrong
This topic has generated more confidently repeated numbers than almost any other in AI SEO, and four errors account for most of them.
Confusing types with counts. Eight documented sub-query types is a taxonomy. It is not a statement that eight searches run. Most circulating figures start with this slip.
Treating a qualitative phrase as data. Turning "ten, twenty, or more" into a measured average is an invention.
Attributing industry shorthand to Google. "Query fan-out" became the de facto name through industry usage, and quotes built around it are frequently attributed to Google without a source.
Assuming the count is constant. Even if a count were measured once, there is no reason to expect it fixed across query type, complexity or time. A simple factual question and a complex purchase question have no reason to fan out identically.
The general pattern — a plausible figure detached from its method, then repeated — is catalogued in where AI SEO statistics come from. The honest position here is short. The mechanism is documented; the magnitude is not. A page that admits the second half is more useful than one filling the gap with a plausible range. We would correct this gladly if someone can point to an origin for it.
Why counting sub-queries is hard
The obstacles are structural rather than a matter of effort. The process is not exposed — intermediate searches happen server-side and no public interface reports them. Citations undercount retrieval. A sub-query can run, return results and contribute nothing cited, so counting cited sources sets a floor on activity and never a total. One source can serve several sub-queries, so citation count and sub-query count are not convertible into one another. Outputs vary between identical runs, making a single observation a sample of one from an unknown distribution. And the system is a moving target: any count measures a product version, valid until the next unannounced change.
We do not know of a method that observes sub-queries directly from outside the system. Absent that, every published count is inference. We would rather say so than add another estimate.
How to observe fan-out yourself
You cannot see the sub-queries. You can see their fingerprints.
Pick a question with obvious implied angles. A purchase question works best, since it naturally implies comparison, price and specification. Record the full citation set rather than just whether you appear — every cited URL and what each page is actually about, because the mix of topics across citations is your main evidence. Classify each citation against the eight documented types, labelling that as inference in your notes.
Repeat the same question several times and note which citations are stable and which rotate. That is the question the citation half-life study asks over longer windows. Then ask the same question three ways — narrow, broad and comparison-framed. If the citation mix barely changes, the implied angles may be generated from the topic rather than your phrasing.
Then write down what you cannot conclude: not the sub-query count, and not which sub-query produced which citation. You are reading outputs and inferring a process, and the inference could be wrong.
What does not transfer between engines
| Does not transfer | Why not |
|---|---|
| The eight documented types | They come from one company's patent. Another system's decomposition has no reason to match that taxonomy. |
| The underlying index | Fan-out here issues searches against Google's infrastructure. An assistant renting a different index searches a different web. |
| Number of sub-searches | Cost and latency budgets differ per operator, so the number of searches each can afford will too. |
| Whether decomposition happens at all | Some systems answer directly from a model with a single retrieval pass, or none. |
| Multi-turn context | A conversational surface can carry earlier turns into later decomposition. A one-shot answer cannot. |
What plausibly transfers is the strategic implication: if several systems decompose questions, covering the adjacent angles of a topic has more surfaces to pay off on. It remains a hypothesis on every one of them. Each system has its own sourcing profile — ChatGPT, Claude, Perplexity and Gemini do not draw on the same web, as Ahrefs' overlap analysis and Profound's platform citation patterns each find in their own way. Which domains end up favoured is tracked in most-cited domains in AI search.
What this means for content, tentatively
Hypothesis Suppose a question really does fan out into comparison, specification, pricing and how-to sub-queries. A page covering only the narrow original question then has fewer chances to be pulled into any single sub-query's results than one covering the surrounding cluster. That is testable and untested — nobody has published a study connecting topical coverage breadth to fan-out capture rate.
What has been studied is which properties are associated with being cited at all: Zyppy's ranking-factor analysis, Ziptie on original research, and on this site the schema markup study and brand mentions versus backlinks. Coverage breadth as a fan-out variable is untouched by all of them.
"Cover more angles" also is not free. Five half-answers can lose to one focused page; price and specification sections go stale fastest; and two of your own pages covering the same angle create ambiguity rather than coverage. The proportionate response is to cover the obvious follow-ups on pages you were writing anyway.
Who this applies to, and who it does not
Considered purchases and complex how-to topics fan out naturally, because comparison, price and prerequisite questions are genuinely implied by them. Simple factual queries have little to fan out into, and navigational or branded queries have none — someone looking for your login page needs no comparison sub-search. Local queries are a separate case again. And before restructuring anything, check whether your customers use this surface at all; the market-share page covers why those figures are harder to read than they look.
Confounds, and the null results we would publish
Anyone measuring fan-out, us included, walks into the same traps. Non-determinism is the worst: the same question can produce different citations on consecutive runs, so a study sampling each question once measures noise as much as system. Personalisation and location mean a study run from one machine describes that machine's view. Product drift during a collection window long enough to gather a decent sample is near-certain. Query set construction can design its own result — choosing questions that obviously imply subtopics will find evidence of subtopic retrieval, so a neutral set has to include questions where fan-out would be pointless. And observer contamination cannot be ruled out.
Four outcomes would undercut the strategic advice on this page, and each would be filed in the null results registry. Broader topical coverage showing no capture advantage. Citation mixes not varying by question type, which would weaken the inference that different angles are retrieved at all. Focused pages outperforming comprehensive ones — a real possibility, and the opposite of received advice. Or sub-query counts turning out to be small and fixed, which would deflate a large amount of published strategy, including some of ours. A page that names the results that would refute it is a page you can check.
Open question Three gaps stay open: the actual sub-query count per question across query types; which sub-query earned a given citation, since attribution within a fan-out is not publicly observable; and whether broader coverage measurably increases capture. Closing them is the purpose of the Query Fan-Out Corpus, registered as a flagship study in the AI Citation Index, with the wider programme in the dataset strategy.
A reasonable bet while the corpus is built
You do not need a confirmed sub-query count to act sensibly. Look at your most important page and ask which of the eight documented types it actually covers. A page answering only the literal question typed covers one at most; a page that also answers the two or three most obvious follow-ups covers more, at no cost beyond the writing time. That is a reasonable bet regardless of how many sub-queries the mechanism turns out to run.
Sequencing matters more than the tactic. A technical GEO audit comes first, because a page an AI crawler cannot render cannot be retrieved for any sub-query at all. The standards this page is held to, and who is behind it, are on the about page.
Mark up the angles your page already covers before adding new ones — retrievability precedes coverage. The Query Fan-Out Corpus results ship first through the newsletter.
Namdev, R. (2026). Google AI Mode statistics and query fan-out (v2). Retrieved from https://ritiknamdev.com/blog/google-ai-mode-statistics Published under CC BY 4.0 — reuse freely with attribution.
This is the mechanism behind one of the flagship studies planned for the Query Fan-Out Corpus, part of the AI Citation Index.