Query fan-out: why 68% of AI citations rank outside Google's top 10

One AI search prompt fans out to 8-30 retrieval queries before any answer is composed. Google's Deep Search mode for AI Mode can fire several hundred. In a November 2025 study of 173,902 URLs across 10,000 keywords, Surfer SEO found that 67.82% of pages cited in AI Overviews don't rank in Google's top 10 for either the main query or its fan-outs. Pages ranking only for fan-out variants are 49% more likely to be cited than pages ranking only for the main query. Pages ranking for both are 161% more likely. The retrieval surface for AI search is roughly 10-16x wider than the keyword surface most marketing teams still measure - and that gap is the mechanical reason rank tracking now under-counts citation exposure.

Why this matters

Most marketing teams measure AI search the same way they measured Google: a list of priority head-terms, position tracking, citation share against those same prompts. That model assumes the system answering the user ran one search. It didn't. ChatGPT, Perplexity, Google AI Mode, and Microsoft Copilot all decompose the user's prompt into 8-12 parallel sub-queries (more for complex prompts), retrieve content for each, and compose afterwards. The retrieval surface that decides whether you get cited is 10-16x wider than the keyword surface you're tracking. Most teams are scoring the wrong game.

We covered the consequence of this in the SEO-AEO gap insight: only 12% of URLs cited by ChatGPT, Perplexity, and Copilot rank in Google's top 10 for the original query. That gap isn't noise. It's the predictable output of a retrieval architecture that fans queries out before it ranks anything. This piece is about the mechanism that produces the gap, and the small set of decisions teams can make to compete on the wider surface.

Method

This is a synthesis of seven publicly cited studies and platform disclosures covering May 2025 through January 2026. Sources are linked inline. Where third-party studies disagreed on numbers we used the most conservative figure. Where platforms (OpenAI, Anthropic) have not disclosed orchestration details, we noted the absence rather than guessed.

• Surfer SEO query fan-out study (November 28, 2025): 173,902 URLs across 10,000 AIO-triggering keywords, 33,000 fan-out queries extracted and analyzed.
• iPullRank query fan-out analysis (December 11, 2025, Lazarina Stoy): query expansion taxonomy across Google AI Mode, ChatGPT, Perplexity, and Copilot.
• Google I/O 2025 AI Mode disclosure (May 2025): custom Gemini 2.5 model trained for fan-out; Deep Search can issue several hundred sub-queries per prompt.
• Search Engine Land patents analysis: coverage of patent WO2024064249A1 (Systems and Methods for Prompt-Based Query Generation for Diverse Retrieval) and the synthetic query training approach.
• Semrush query fan-out research (2025-2026): commercial SERP behavior under AI Overviews, fan-out optimization experiments.
• iPullRank x SimilarWeb prompt-length data (December 2025): AI search prompts average 70-80 words vs. 3-4 words for traditional Google queries.
• Search Engine Land query fan-out guide: cross-platform synthesis of fan-out behavior across Google, ChatGPT, Perplexity, and Copilot.

We did not run a controlled experiment. Treat all numbers below as the best public picture available at this moment - the underlying systems retune frequently, and most platforms publish little about their orchestration.

Findings

1. One prompt becomes 8-30 retrieval queries on average. Deep Search fires hundreds.

The publicly reported range across AI search systems is consistent: simple prompts fan out to 4-8 sub-queries, complex prompts to 12-20, and Google's Deep Search mode can issue several hundred sub-queries on a single prompt, per Google's I/O 2025 keynote. Mike King's Qforia tool, which replicates Google's fan-out logic against a Gemini prompt informed by patent WO2024064249A1, returns 20-30 example sub-queries for typical commercial prompts.

The shape: one user input, dozens of internal queries running in parallel against the live web, the knowledge graph, shopping graphs, and platform-specific indexes. A separate retrieval pass for each. A separate citation list for each. The composer model then synthesizes a single answer with 3-12 visible citations, drawn from across all of those passes.

2. Eight types of fan-out queries cover most of what gets generated.

iPullRank's December 2025 analysis catalogued eight synthetic query variants AI search systems generate during fan-out:

• Equivalent: alternative phrasing for the same question.
• Follow-up: the next logical question after the original.
• Generalization: a broader version of the prompt.
• Specification: a more detailed or constrained version.
• Canonicalization: a standardized industry phrasing.
• Language translation: cross-lingual retrieval for multilingual queries.
• Entailment: a question logically implied by the original.
• Clarification: a question presented to confirm intent.

For a marketer, this is the keyword research framework for AI search. If you only own pages for the head term, you compete in slot one of the fan-out. The other seven slots go to whoever covered the equivalent phrasing, the comparison, the follow-up, and the implied next step.

3. 67.82% of pages cited in AI Overviews don't rank in Google's top 10 at all.

Surfer SEO's November 2025 study of 173,902 URLs across 10,000 AIO-triggering keywords found that 67.82% of cited pages were not in Google's top 10, for either the main query or any of the fan-out variants. They were ranking deeper, sometimes for adjacent topics or specific entities, sometimes pulled from page two or beyond. The retrieval surface for AI search is structurally separate from the SERP.

Of the 32.18% of citations that did rank top 10 for something, only the top 3 AIO citation slots had a clean correspondence with traditional ranking: 54.14% of top-3 citations matched a top-10 rank for the main or a fan-out query. Below position 3 the correspondence weakens fast. Position-driven content strategies recover roughly half of the citations in the most prominent slots, and roughly nothing in the long tail of in-answer references.

4. Ranking for fan-out queries is 49% more citation-likely than ranking for the main query. Ranking for both is 161% more likely.

Same Surfer study: pages ranking for fan-out variants of an AIO-triggering keyword are 49% more likely to be cited than pages ranking only for the main keyword. Pages that rank for both the main query and its fan-outs are 161% more likely to be cited than pages ranking only for the main keyword. The Spearman correlation between fan-out rankings and AIO citations is 0.77, which is strong by SEO study standards.

The arithmetic is straightforward. If a prompt generates 12 retrieval queries, owning even one of the 11 sub-queries gives you 11 separate chances to be selected, vs. one chance from owning the main query. Owning more variants compounds. Coverage depth on a topical cluster outperforms a single dominant page on the head term, and the data now puts a number on it.

5. AI prompts are 17-26x more complex than search queries, which changes what gets fanned out.

iPullRank x SimilarWeb telemetry from December 2025: AI search prompts average 70-80 words, vs. 3-4 words for traditional Google queries, a 17-26x increase in query complexity. Longer prompts contain more entities, more constraints, and more implicit intent, which means the fan-out generator produces more, narrower sub-queries.

For a marketing team this means head-term tracking isn't just incomplete. It's measuring a different unit. The query you can rank for in Google has 4 words. The prompt that triggers your citation in ChatGPT has 70. They are not the same demand, and they reward different content.

6. Google's fan-out is a custom Gemini 2.5 model. The other vendors haven't disclosed details.

At Google I/O 2025, Google confirmed AI Mode uses a special-purpose Gemini 2.5 model trained specifically to generate fan-outs. Patent WO2024064249A1 (Systems and Methods for Prompt-Based Query Generation for Diverse Retrieval) describes training a query expansion model on synthetic query-document pairs.

OpenAI has not published orchestration details for ChatGPT search or for Atlas. Anthropic has not described how Claude's web search composes sub-queries. Perplexity has documented its Copilot mode (Pro Search, Deep Research) as multi-stage plan-and-execute, but the per-query expansion logic is not public. Treat the 8-30 range as platform-specific within those bands; do not assume identical behavior.

What this means for marketers

• Stop tracking head terms in isolation. Pair every priority head-term with the eight fan-out variants. Track citation share, not just position, across the whole cluster.
• Build for sub-queries, not for the prompt. A page that perfectly answers one fan-out (a comparison, a follow-up, a constraint case) often outperforms a generalist page on the head term.
• Topical depth beats topical breadth. The 49% / 161% uplifts reward covering more of one topic's fan-out tree, not adding new topics. The retrieval surface compounds inside a cluster.
• Treat keyword research as prompt research. A 70-word prompt is the unit. Sample real prompts (yours, your competitors', the ones your sales team gets asked) and decompose them into the eight variants instead of starting from a 4-word seed.
• Move measurement with the architecture. Citation share by prompt cluster and by fan-out variant is the metric. Average position is now a lossy proxy.

If you only own the head term, you compete for one of thirty slots. If you own the fan-out, you compete for thirty.

Limitations

Three caveats. First, most public data on per-query expansion comes from a small number of analyses (iPullRank, Surfer, Semrush, Search Engine Land's syndicated reporting, NoGood). The platforms themselves disclose little. Independent replications at scale would strengthen the picture, and we expect more in the next two quarters as more SEO teams build fan-out instrumentation.

Second, the Surfer 67.82% / 49% / 161% numbers come from one study at one moment in time on AI Overview citations specifically. AIO behavior shifts week-to-week as Google retunes Gemini and adjusts fan-out aggressiveness. The directional finding (rank-for-fan-outs beats rank-for-head) is robust across every smaller study we reviewed; the exact magnitudes will move.

Third, this insight covers retrieval. It does not cover the ranking step inside an answer (which retrieved passages the composer cites) or the synthesis step (how the model attributes credit). Those add filters between fan-out and citation. We're explaining the funnel's first stage, not the whole funnel. For the citation and synthesis side, see our citation engineering playbook.

Where this goes next

Query fan-out is the architectural reason AI search rewards different content than Google search did. Marketing teams that internalize the eight-variant taxonomy and ship coverage across it will compound visibility as fan-out depth grows. Teams still tracking 20 head terms in a rank tracker will keep posting flat dashboards while their competitors get cited more often, on more prompts, in more verticals.

Oz's research and content agents are built for the fan-out world. They decompose prompts into the eight variant types, audit cluster coverage against citation data, and ship the missing pages. Built by Passionfruit, refined across 500+ brand engagements, now in product form. If you want to be cited when prompts get longer and retrieval gets wider, join the waitlist. For the operator essay on where this fits in the broader shift, see Agentic marketing vs marketing automation.