Google documents that AI Mode "may use a query fan-out technique," and a granted patent describes eight sub-query type categories. The specific number of sub-queries issued per real question — the "8-16" figure that circulates constantly — has never been published with a disclosed method. This corpus is designed to infer and publish that data directly, with its limitations stated plainly.
Why this replaces folklore with data
"AI Mode fans a query out into 8-16 sub-queries" appears across dozens of SEO articles. It's always stated with the same confidence, never with a visible source. Our own trace, documented in the fan-out mechanism page, found the underlying patent describes eight sub-query types, not a count of instances per query. The repeated figure appears to have mixed those two things up. This corpus doesn't assume the folklore is wrong. It simply replaces the assumption with a measurement.
What is currently known
Google's own documentation confirms the fan-out mechanism exists and names eight sub-query type categories in a granted patent (US11663201B2, filed 2018).
The actual number of sub-queries issued per real-world question, and how that number varies by question complexity, is not published anywhere we could locate with a disclosed method.
Corpus design
- 01 Issue query From the published 1,000-query set
- 02 Capture full response Including any observable sub-answer structure
- 03 Infer sub-query boundaries From citation clusters and answer segments
- 04 Classify each sub-query Against the 8 documented patent categories
- 05 Publish the corpus Query → inferred sub-queries → citations, open
How a fan-out will be captured
AI Mode doesn't expose an internal sub-query list through any known public interface. So this corpus works from observable proxies: distinct answer segments, citation clusters that don't overlap with each other, and structural cues in the response, headers, comparison tables, distinct topic shifts. These are inferred sub-query boundaries, not a direct readout of Google's internal process. That's a real limitation, stated explicitly rather than glossed over.
A worked inference example
Here's the inference process made concrete. For a hypothetical query like "should I upgrade to a mirrorless camera," an AI Mode response might present distinct segments: the general case for switching, a specific comparison of leading mirrorless models, pricing considerations, and a section on compatible lenses and accessories. Each draws on a different, largely non-overlapping set of cited sources.
The inference method treats each such segment as one inferred sub-query. It's distinguished by structural cues, a new heading or clear topic shift, and by citation clustering, a distinct set of sources not shared with the surrounding text. Classifying that segment against the patent's eight documented sub-query types, here, likely "comparisons," "pricing," and "related products," produces the structured record the corpus ultimately publishes.
Pre-registered hypotheses
| # | Hypothesis | Prediction |
|---|---|---|
| FO1 | Inferred sub-query count correlates positively with question complexity (measured by word count and entity count) | Supported |
| FO2 | The commonly-cited "8-16" range is roughly consistent with observed inferred counts | Supported |
| FO3 | Pages covering multiple angles of a topic (comparison + specification + pricing) appear in more inferred sub-query clusters than single-angle pages | Supported |
FO2 is included specifically to give the existing folklore a fair test, not to assume it's wrong by default. If the commonly repeated range turns out to be roughly accurate, that's worth publishing plainly. Replacement isn't the goal in itself; an accurate answer is.
'AI Mode fans a query into 8-16 sub-queries' is repeated everywhere with no visible source. We're not assuming it's wrong — we're building the corpus that would let anyone actually check.
Share on XHow the inference method will be validated
Two checks are planned before treating the inferred data as reliable. First, a human-review sample. A subset of responses gets manually reviewed and independently segmented by a person blind to the automated inference's output. The two get compared for agreement. Low agreement would signal the automated method needs revision before the full corpus is trusted.
Second, a stability check. The same query runs multiple times, checking whether the inferred sub-query structure stays reasonably consistent. Wildly different inferred structures for the same question would suggest the method is picking up noise, not a genuine underlying pattern.
What gets published
The full corpus: query → inferred sub-query segments → citations per segment, as a downloadable file, alongside the inference methodology so the classification logic itself can be audited and disputed.
Why Google itself might find this useful too
Worth noting explicitly: an independent, methodologically transparent measurement of fan-out behavior isn't only useful to site owners trying to understand the mechanism from the outside. A company as large as Google runs countless internal experiments on its own products. But an external, reproducible dataset provides a different kind of value: a check against internal assumptions, visible to anyone, that doesn't depend on any single organization's internal reporting incentives.
This isn't a claim that Google needs or has requested this corpus. It's an observation about why open, third-party measurement of a company's own product can have value even to that company.
What to do before the corpus ships
You don't need the finished corpus to act on the underlying idea. If your topic has real adjacent angles, comparisons, pricing, related products, cover them on the same page or in clearly linked companion pages, rather than assuming a single narrow answer is enough.
That structure is a reasonable bet regardless of the exact sub-query count this study eventually confirms. A page that already anticipates the obvious follow-up questions is well positioned whether the real number turns out to be 8, 16, or something else entirely.
Limitations
- This is inference, not direct observation. Without access to AI Mode's internal query log, sub-query boundaries are reconstructed from observable output, which will contain some misclassification.
- The product can change its behavior without notice, making any specific count a snapshot rather than a permanent architectural fact.
Namdev, R. (2026). The Query Fan-Out Corpus (v1). Retrieved from https://ritiknamdev.com/blog/query-fan-out-corpus-study Published under CC BY 4.0 — reuse freely with attribution.
See Google AI Mode statistics and query fan-out, explained for the mechanism this corpus will measure directly, and the AI Citation Index for where this study sits in the wider roadmap.