The Fan-Out Machine: Your Content Is Being Retrieved by Questions Nobody Typed

Ask Google AI Mode "what is the best CRM for a 10-person startup with a $500 monthly budget?" and you will get one clean, synthesized answer. Behind it, the system did not run one search. It ran a portfolio: sub-queries for pricing models, ease of use, integration limits, migration difficulty, and more — each executed separately, each retrieving its own passages, all fused into a single response with citations.

This is query fan-out, and it is the single most consequential mechanism in AI search that most content strategies still ignore. The direct answer to "why doesn't ranking #1 get me cited in AI Mode?" is: because the ranking was computed for the question the user typed, while the citations were assembled from a dozen hidden questions nobody typed.

The Mechanics: One Prompt Becomes 8-16 Searches

Google described the architecture at I/O 2025, and the industry has since measured it. AI Mode decomposes a user query into 8 to 16 sub-queries — commonly averaging around 10 — executes them in parallel against the web index (plus the Knowledge Graph, Shopping, News), and synthesizes a long-form answer across the union of what comes back.

Measured fan-out counts, by platform:

Two structural facts fall out of that table. First, Google is the fan-out engine: nearly every AI Mode query triggers expansion, at roughly triple ChatGPT's depth. Second, the platforms are not interchangeable — the same prompt produces very different retrieval surfaces, which is one more reason cross-engine citation overlap stays low (the ~11% figure this series has cited before).

The Eight Sub-Query Types

Google's patent (US 11,663,201 B2) documents eight synthetic query types, with Mike King's reading at iPullRank remaining the field's standard reference. Applied to one starting prompt — "best physiotherapist for running injuries in Toronto" — they are:

1. Equivalent — rephrasings of the same intent ("top physio running injury Toronto")

2. Specification — narrower versions ("shockwave therapy for runner's knee Toronto")

3. Generalization — broader nets ("physiotherapy for sports injuries Ontario")

4. Follow-up — the logical next question ("do I need a referral for physiotherapy in Ontario?")

5. Canonicalization — resolving entities and names to their standard forms

6. Translation — cross-language variants of the intent

7. Entailment — implications that must hold if the answer is true

8. Clarification — ambiguity-resolving variants

The strategic weight of this list is easy to miss: six of the eight types are questions your keyword research has no record of, because they were never searched by humans at volume. Fan-out sub-queries routinely show zero monthly search volume in Semrush or Ahrefs — which means keyword-first content planning is now structurally blind to most of the retrieval that determines AI citations.

Passage-Level Retrieval: The Page Was Never the Unit

Each sub-query performs passage-level retrieval — the system scans for specific text blocks that answer the narrow question, not pages that rank for the broad one. The measurable consequences:

• Moz's 40,000-query analysis found 88% of pages cited by AI Mode do not rank in the organic top 10 for the original prompt. (A Surfer SEO study puts the equivalent figure at 67.82%.) The passage matters more than the page's standing.

• iPullRank's work identifies extractability — heading-query match, schema markup, self-contained answer blocks — as one of the biggest predictors of which passages survive the streamlining stage, where retrieved candidates are cut down to synthesizable size.

• The brand filter still exists: Ahrefs' 75,000-brand study found a 0.466 correlation between brand search volume and AI Mode mentions (stronger than the 0.392 for AI Overviews). Fan-out widens the door for unknown pages; recognized entities still clear the trust filter more easily.

• One 2026 analysis found brands are 161% more likely to be cited when they rank for both the fan-out queries and the main query — coverage of the constellation, not just the head term.

What Survives the Machine: A Practical Checklist

Fan-out rewards content that is retrievable by question and quotable in isolation. Concretely:

1. Write every H2/H3 as a question a fan-out sub-query would ask — specification, follow-up, and clarification questions included, not just the head term.

2. Follow each heading with a 15-30 word answer that survives isolation. If the model lifts that paragraph alone, it must still make sense — no "as mentioned above," no build-up.

3. Mark up the passages. Structured data remains the highest-leverage technical layer for Google's surfaces (schema-marked pages are cited roughly 2.3× more often in AI Overviews).

4. Reverse-engineer your fan-out before building. Simulators (Qforia and similar, using a Gemini API key) return 20-30 candidate sub-queries sorted by intent; Search Console long-tail queries and People-Also-Ask patterns approximate them; turning sub-queries into standalone assets — a comparison table, an explainer, a budget roundup — beats cramming them into one guide competing against itself.

5. Cover the constellation, not the star. The compounding asset is a topic cluster whose pages collectively answer the specification, follow-up, and entailment questions — because that is literally the portfolio the machine retrieves.

None of this guarantees citations — the brand-authority filter and the synthesis stage are outside anyone's control. But it aligns the work with the actual unit of retrieval, which keyword-first planning structurally misses.

Limitations

Sub-query counts come from small-to-mid prompt panels (501 to 9,000 prompts) using API access that may differ from production behavior; treat ranges, not point estimates, as the finding. Google has not published official production fan-out counts; the patent describes capability, not deployed parameters. Passage-survival predictors (extractability, heading-query match) are inferred from industry experiments, not Google documentation. Fan-out behavior varies by query type — navigational and transactional prompts fan out less than research prompts. All figures as of September 2026.

Frequently Asked Questions

What is query fan-out?

The mechanism AI search engines use to decompose one user prompt into multiple parallel sub-queries (8-16 in Google AI Mode, 2-4 in ChatGPT, 1-2 in Perplexity), retrieve passages for each, and fuse the results into a single cited answer. Google's patent calls it query variant generation.

How many sub-queries does Google AI Mode generate per query?

Measured averages run 9-11, with a documented range of 3-28 (Seer Interactive, 501 prompts, November 2025; Nectiv, 9,000 prompts, December 2025), and roughly 95% of queries trigger fan-out at all. Exact counts vary by query complexity and are not officially published by Google.

Why doesn't ranking #1 on Google get me cited in AI Mode?

Because the citation decision happens across a dozen hidden sub-queries, each retrieving passages independently. Moz's 40,000-query analysis found 88% of AI Mode-cited pages do not rank in the organic top 10 for the original prompt. Ranking remains useful — it feeds candidate pools and the brand-authority filter — but it is no longer the citation mechanism itself.

How do I find the sub-queries for my topic?

Three practical routes: fan-out simulators that return 20-30 candidate sub-queries from a Gemini API key; mining your Search Console long-tail queries, which often mirror sub-query patterns; and applying the eight patent types (equivalent, specification, generalization, follow-up, canonicalization, translation, entailment, clarification) manually to your head terms. Then build standalone assets for the questions that matter — not one page trying to answer everything.

Does fan-out mean keywords don't matter anymore?

No — it means the keyword was always the wrong unit. Head terms still anchor clusters and brand authority still filters citations (0.466 correlation with AI Mode mentions, Ahrefs). What changes is that the pages which win passages are the ones answering the hidden questions: the specification, follow-up, and entailment variants that never showed up in keyword research.

Last updated: September 9, 2026 Sources and method note: Fan-out counts from Seer Interactive (501 prompts via Gemini 3 API, November 2025), Nectiv (9,000 prompts, December 2025), and Qwairy (102,000 queries, 2025–2026), compiled in Logik Digital's 2026 fan-out guide; eight sub-query types from Google patent US 11,663,201 B2 as analyzed by iPullRank; top-10 citation decoupling from Moz (40,000 queries) and Surfer SEO; brand correlation from Ahrefs 75,000-brand study (2025–2026); extractability findings from iPullRank; +161% fan-out-coverage finding from Masthead Media's 2026 analysis. Industry sources use differing prompt panels; treat all figures as directional and dated.