TL;DR
Fan-out is a query-generation step that runs before retrieval. The sub-queries are written by the model, from its own prior — not derived from your rankings and not read off your site.
In EMGI’s July 2026 capture, 69% of ChatGPT’s fan-out queries already contained a specific brand name, and 56% of its citations landed on vendor-owned pages. The shortlist was set before a single page was fetched.
So a decomposition splits into two families: unnamed sub-queries, which select the candidates and are answered off-domain, and named sub-queries, which confirm the detail and are answered on your own pages.
The popular “cover every sub-query” advice reads a correlation backwards. Ranking for many fan-out queries is a symptom of already being trusted, not a method of becoming trusted.
Two instruments below: the Fan-Out Anatomy (the six roles a sub-query plays, and who is a valid answer to each) and the Co-Fetch Window (server-log forensics that shows which of your pages an engine grouped into one decision).
1. The mechanism, stated precisely
What is query fan-out? Query fan-out is the step where an AI search system sets your question aside, writes a set of narrower questions of its own, runs a separate retrieval for each, and assembles one answer from the results. Google named the technique when it launched AI Mode, describing it as breaking a question into subtopics and issuing a multitude of queries simultaneously on the user’s behalf.
The word doing the heavy lifting there is writes. The sub-queries are generated, not derived. The 2026 Query Optimization Survey describes a consistent five-stage pipeline across the major engines — intent detection, query rewriting, query expansion, retrieval, citation synthesis — with source selection happening only at the final stage. Google’s patent application US20240289407A1 calls the generation step “prompted expansion”: a language model receives structured instructions to produce queries showing intent diversity, lexical variation and entity-based reformulations.
Notice what is absent from that description. The index. The expansion model is not told what exists on the web; it is told what a good set of questions looks like. That single architectural fact is the source of everything awkward that follows, and it is the part the optimisation advice tends to skip.
How wide does the fan get?
Wider than most content plans assume, and it varies enormously by engine. Ekamoira’s analysis found 59% of prompts trigger between five and eleven simultaneous sub-queries, with nine to eleven typical for complex questions. Peec.ai, working across 20 million search queries, put ChatGPT considerably lower at 2.3 to 2.8 sub-queries per prompt — while the average sub-query doubled in length from six words to twelve between October 2025 and early 2026. EMGI’s July 2026 pilot recorded 127 fan-out queries across 48 SaaS buying prompts on ChatGPT: 6.3 per prompt that searched at all.
How many sub-queries does one prompt generate? Between two and roughly a dozen, depending on the engine and the complexity of the question, with nine to eleven typical for a complex prompt on Google AI Mode and two to three typical on ChatGPT. Anyone quoting a single figure is quoting one engine on one prompt class.
The qualifier almost nobody quotes
That last clause matters more than any of the averages. A large share of prompts never fan out, because they never retrieve. In the same EMGI capture, ChatGPT ran a live web search on 42% of the SaaS buying prompts (20 of 48), Claude on 81% (13 of 16) and Perplexity on every one. Gemini searched on 6% — one prompt in sixteen — answering the rest from memory, including “best HR software for small business”. Navigational, transactional and straightforward factual questions typically skip retrieval altogether, because synthesising diverse sources adds nothing to them.
Where no retrieval runs, the answer is composed entirely from the model’s weights, and nothing on your page reaches it. That is not a footnote to fan-out strategy. It is a boundary drawn around the whole discipline, and it belongs in the budget conversation before any of the tactics do — the same way the basic mechanics of how links confer authority belong in it before any single tactic gets funded.
2. The consensus playbook, stated at full strength
The prevailing 2026 advice on fan-out is coherent, well evidenced, and worth stating properly before disagreeing with any of it.
Surfer’s December 2025 study of 173,902 URLs across 10,000 keywords found that 68% of pages cited in AI Overviews were not in the top ten organic results, and reported a Spearman correlation of 0.77 between the number of fan-out queries a page ranks for and its probability of being cited. ALM Corp’s analysis across 10,000 keywords and 33,000 fan-out queries found pages ranking for both the main query and fan-out queries were 161% more likely to be cited than pages ranking for main keywords alone; ranking for fan-out queries by themselves was worth 49%. Pages covering both accounted for 51% of all citations, against under 20% for main-query-only pages.
The gap keeps widening. Ahrefs, across 863,000 keyword SERPs and 4 million AI Overview URLs, found only 38% of cited pages also rank in the top ten for the same query — down from roughly 76% in July 2025. Mike King, speaking at SparkToro Office Hours in January 2026, put the overlap between traditional rankings and AI citations at 25–39%. Cyrus Shepard of Zyppy synthesised 54 separate AI citation experiments and scored fan-out rank at 9.3 out of 10.
Trigger rates give the advice its urgency. Seer Interactive, working across 49,000 tracked queries, found comparison queries trigger AI Overviews 95% of the time and question-format queries 86% of the time — which is to say the surfaces most exposed to fan-out are precisely the mid-funnel and bottom-funnel queries most B2B and considered-purchase teams have spent a decade building pages for. SparkToro put 68% of Google searches ending without a click in the first four months of 2026. The exposure is real and it is concentrated exactly where the investment already sits.
The playbook that follows is equally clear: map the question space, build tight topical clusters, write self-contained atomic answers under question-shaped headings, and use a simulator — iPullRank’s Qforia, DEJAN’s fan-out generator, WordLift’s DSPy-based tool, Profound’s captured fan-outs — to produce the sub-query list you write against. Most of the AI visibility platforms now on the market sell some version of it.
None of that is wrong. It is incomplete in a way that costs real money, and the evidence for the gap sits inside the same datasets.
3. The correlation runs the other way
Look closely at what the headline studies actually measure. They measure whether a page ranks for fan-out queries. Ranking is not an input you control by writing; it is an outcome. And it is an outcome of the same underlying property — established authority, corroboration, entity recognition — that also drives citation.
So the 0.77 correlation is a correlation between two symptoms of one cause. Adding eleven sub-headings to a page does not make it rank for eleven narrow queries. It makes it a longer page. The coverage advice quietly assumes away the hard part: that you can already rank for narrow commercial queries at all, which in any contested market is a function of the referring-domain profile behind the page far more than its subheading structure, not of its subheading count.
The tell is inside Zyppy’s own synthesis. In the same set of 54 experiments, search rank scored 9.4 and URL accessibility 9.5 — both marginally above fan-out rank at 9.3. The strongest available meta-analysis in the field says classic ranking is at least as decisive as fan-out coverage, and that a page being fetchable at all matters slightly more than either. That is not the story the coverage playbook tells, and it is drawn from the playbook’s own evidence base.
For example: two suppliers publish the same eleven-section guide to the same buying decision in the same week. One sits on a domain that eight trade publications, two accreditation registers and a manufacturer’s partner page have referenced over three years. The other does not. Both pages “cover the fan-out”. Only one of them ranks for the sub-queries, and only one gets cited. The content was never the variable.
There is a second, quieter problem with reading the studies as instructions. They are measurements of what cited pages look like, taken after the fact. A page cited by an AI Overview will, on inspection, tend to rank for many related queries — in the same way that profitable companies tend to have large finance teams. The observation is accurate. The instruction “hire accountants” does not follow from it. Coverage is what winning looks like from the outside; it is not the mechanism that produces the win.
4. The naming gate: the shortlist is set before retrieval runs
EMGI’s July 2026 capture reported four numbers together, and they are far more interesting as a set than individually. Across 48 SaaS buying prompts on ChatGPT: 69% of the fan-out queries contained a specific brand name. 86% contained a year. 51% asked about pricing. And 56% of the citations pointed to vendor-owned pages rather than third-party editorial.
Read those in sequence. If more than two-thirds of the sub-queries name a brand, the candidate set was decided before retrieval ran. The engine is not shopping. It is checking. It arrived at the search step already holding a list, and it spent its retrieval budget confirming prices, dates and specifications for names it brought with it. Which is precisely why vendor pages take the majority of citations: a vendor page is the correct answer to “what does Brand X cost in 2026” and a poor answer to “who are the best suppliers”.
A UK agency capture published in July 2026 by Little Green Agency makes the same point without the SaaS framing. Given the prompt “best builders leicester”, ChatGPT expanded it into four searches — and the Federation of Master Builders, Checkatrade and TrustATrader appeared inside the query strings themselves, before any page had been fetched. The model named its intended sources in advance. It was not discovering trusted intermediaries; it was going to fetch the ones it already trusted.
That is the naming gate. A sub-query that names you can only be generated if the model already holds your name — and nothing on your own website causes that to happen. Your on-site estate is superb equipment for the second half of the decomposition and structurally irrelevant to the first.
THE FAN-OUT ANATOMY
The six roles a sub-query plays in a real decomposition, and who counts as a valid answer to each.
| Sub-query role | Typical form | Names a brand? | Valid answer | What it decides |
| Scoping | “what is X used for”, “types of X” | No | Either | Which category you are resolved into — and therefore which later contests you enter |
| Enumeration | “best X suppliers UK 2026” | No | Third party | Whether you are a candidate at all. Nothing downstream rescues a failure here |
| Comparison | “X vs Y”, “alternatives to X” | Both names | Third party | How your difference is described, and in whose vocabulary |
| Specification | “X pricing”, “does X support Y” | Yours | First party | Whether your own detail is accurate, current and extractable |
| Verification | “X reviews”, “is X accredited” | Yours | Third party | Whether you survive synthesis after being retrieved — the silent drop |
| Constraint | “X for [specific condition]” | Usually no | Either | The uncontested citation. Rarely written by anyone, so rarely competed for |
Three of the six roles admit only third-party answers, by construction rather than by preference. You cannot be the valid source on whether you belong on a shortlist, how you compare with a rival, or whether your claims check out. One role — specification — is reliably yours. Two are contestable. And the one role you own outright is only generated once the unnamed roles have already produced your name.
Key takeaway
Fan-out does not hand you nine extra chances. It runs nine contests you must already be eligible for. Publishing raises your ceiling on one of the six roles; the other five are decided on other people’s domains, and they are re-decided on every prompt.
5. What this does to content architecture
The retrieval unit is a passage, not a page
Retrieval at this layer works on chunks, aggregated by chunk-level semantic similarity rather than page-level relevance, and AI Mode may pull up to five chunks either side of a matched one to supply context. Two consequences follow immediately, and they point in opposite directions from the usual advice.
First, a thin page built to answer one sub-query is the wrong shape in both directions: too slight to establish who is speaking, and too coarse to be the retrieved unit anyway. Second, the useful metric is not pages published but answer density — how many distinct sub-queries a single document can answer, each in a self-contained block that survives being lifted out. The discipline is closer to writing for featured snippets than to writing long-form guides: state the answer, then explain it.
For example: a single well-corroborated 2,000-word document that answers pricing bands, lead times, compliance intervals and two named edge cases in four self-contained blocks is eligible for four sub-queries. Four 500-word pages carrying the same material are eligible for four sub-queries as well — but each competes for context, none establishes the entity, and the internal links between them do work that a single heading would have done for free.
Publish the number, not the build-up
For example: “how often does local exhaust ventilation need testing” has a statutory answer expressed as an interval. A passage that states the interval in its first sentence is retrievable. One that reaches it in paragraph nine, after a history of the regulations, is not — the chunk that gets matched will be the history.
Recency is mechanical too. If 86% of ChatGPT’s fan-out queries carry a year, undated pages lose every temporal variant of every sub-query, regardless of quality. And if the page cannot be fetched and parsed cleanly — Zyppy scored URL accessibility highest of the three factors it tested — none of it applies; client-side rendering that hides content from crawlers removes you from contests you would otherwise win.
Structure so a block survives removal
The practical test for any passage is whether it still makes sense with the two paragraphs on either side deleted. If it depends on an earlier definition, an implied subject, or a heading three sections up to be intelligible, it will be lifted, misread and either dropped at synthesis or attributed to somebody who wrote it more plainly. Name the subject inside the block. Repeat the entity rather than pronouning back to it. Put the number, the interval or the price in the opening sentence and the caveats after it, not before.
Your own pages compete with each other
One more consequence, rarely mentioned: near-duplicate pages targeting neighbouring sub-queries split the evidence for the same chunk match. Consolidating five thin pages into one dense document usually raises answer density and reduces internal competition at once. It also costs long-tail rankings, which is a real trade rather than a free win, and should be decided deliberately rather than as a side effect.
6. Seeing the decomposition without buying it
Simulators generate a plausible decomposition. They do not return the decomposition, and the difference is material: Ekamoira found roughly 73% of fan-out queries change between runs of the same prompt. Two free first-party sources get you closer.
The first is the Gemini API. With grounding enabled, the response includes a groundingMetadata object containing a webSearchQueries array — the actual searches issued. The honest limitation is that this is the API’s fan-out, not consumer AI Mode’s, and the two are not guaranteed to match. Perplexity’s API, by contrast, exposes citations but not its internal sub-queries, a constraint EMGI stated plainly in its own write-up.
Beyond the API, the engines vary in how much they volunteer. Perplexity exposes its search steps directly in the interface. Google’s AI Mode shows the sources it considered, which lets you infer the branches it took even without the query strings. ChatGPT hides its sub-queries in the product, though they remain visible in the network response behind a conversation, and a small industry of browser extensions has grown up around extracting them. All of these are samples of a non-deterministic system: run the same prompt twice and you will get two overlapping but different sets, which is a reason to run each prompt several times and union the results rather than to treat any single capture as the answer.
The second source is one you already own, and nobody can sell it to you.
THE CO-FETCH WINDOW
1. Filter your access logs to retrieval-time agents: Claude-User, OAI-SearchBot and ChatGPT-User, PerplexityBot, Bingbot. Verify them — a user-agent string is an assertion, not an identity. Check published IP ranges or a Web Bot Auth (a signed-request scheme for verified agents) signature.
2. Group requests by agent, source network and a rolling 90-second window. Each group is one decomposition, executed against your estate.
3. Read three things off the groups. Burst width: how many of your URLs the engine treats as belonging to one question. Repeat membership: which single URL appears in the most bursts — that page is carrying your visibility. Orphans: URLs fetched alone every time, and URLs never fetched at all.
4. Interpret. A median width of one means you supplied a single fragment and someone else supplied the rest. Width of three or more means the engine tried to build several roles from you. Never-fetched URLs sit outside every decomposition you appear in — which is a content-strategy finding, not a technical one.
5. Accept the limit. Google’s AI Mode serves largely from its own index, so most of its fan-out will not appear in your logs. This reads the retrieval-time engines directly and stands as a proxy for the rest — an imperfect one, and better than a simulation.
7. Why fan-out raises the return on earned coverage
Here is the arithmetic that should drive budget. Take a ten-sub-query decomposition with a typical role mix. Specification is yours. Scoping and constraint are contestable and you can often take them. That is roughly three roles reachable by publishing — a ceiling, not a target. Enumeration, comparison and verification are not reachable by publishing at any volume, because a first-party document is not a valid answer to them.
The asymmetry compounds. Your on-site estate is a stock: written once, retrieved repeatedly. The third-party roles are a flow: re-answered from scratch on every prompt, from whatever the web currently says. Every additional sub-query the engines generate is another draw against that flow. Fan-out therefore multiplies the value of off-domain corroboration while your own publishing saturates.
Where each unreachable role is actually won
Enumeration is decided on roundups, buyers’ guides, supplier directories and trade listings — which is why placements inside existing ranked roundups and the local citation and directory layer behave so differently now than they did as ranking tactics; they are candidate-set entries. Sector launch surfaces such as Product Hunt and its equivalents function the same way for newer entrants. Comparison is won on independent editorial, which is the durable case for contributed articles on third-party publications.
Verification is the one most teams under-fund, and it is the role that quietly drops you after retrieval. It is answered by dated third-party mentions, accreditation registers, trade-body membership pages and named expert quotes — the reason journalist-request platforms and trade body and event sponsorships produce citation value out of proportion to their referral traffic. Recency-shaped sub-queries reward timely commentary tied to live news, because a mention carrying this quarter’s date answers the temporal variant that an evergreen page cannot.
Scoping deserves its own line. It determines which category the engine resolves you into, and therefore which contests you are entered for at all — the practical face of entity authority and how it is measured. If you sell across borders, the scoping sub-query is also where market context is fixed, which is the underrated argument for market-specific link acquisition and for separate European link estates or dedicated South Asian coverage rather than one global estate serving every market.
One measurement warning follows directly. Placements that win enumeration and verification sub-queries frequently send no referral traffic at all — a directory entry, a register listing, a member page — and under a click-based evaluation they look like waste. They are not being read by people. They are being read by a retrieval system, once per prompt, as evidence about whether you exist and whether you check out. Judging them on sessions is judging a reference on its footfall. If your reporting cannot separate “earned a click” from “was consulted”, it will systematically defund the placements doing the most work in a fan-out world.
Key takeaway
Publishing is a stock and corroboration is a flow. Fan-out draws against the flow on every single prompt, which is why a bigger content estate and a static referring-domain profile produces a rising cost per citation — and why the tactics that earn third-party coverage now sit upstream of content planning rather than beside it.
8. Worked example: Marden & Bly Extraction, Sheffield
Marden & Bly is a hypothetical but deliberately ordinary case: a Sheffield supplier and installer of industrial dust and fume extraction systems, £5.9M turnover, 41 staff, selling into manufacturers who must meet statutory testing duties under COSHH. Its website carried 180 URLs, 62 of them sector landing pages built years earlier for paid search.
November 2025 — baseline
AI-referred sessions ran at roughly 240 a month. Across 40 tracked commercial prompts across four engines, the company appeared in 9. The marketing lead’s proposed fix was the standard one: run a fan-out simulator across the 40 prompts and commission a page for each sub-query it returned. That plan would have added an estimated 90 pages.
December 2025 — the co-fetch read
Before commissioning anything, six months of server logs were filtered to verified retrieval agents: 4,100 requests, grouped into 780 bursts. Median burst width was 1.4 URLs — the engines were taking a single fragment and building the rest of the answer elsewhere. Only three URLs appeared in more than twenty bursts: the LEV testing FAQ, a COSHH compliance guide and the homepage. And 118 of the 180 URLs had not been fetched once in six months, almost all of them the paid-search sector pages.
January–February 2026 — reconstruction
Twelve head prompts were run ten times each across three engines, with cited URLs and visible sub-queries unioned. That produced 94 distinct sub-queries, classified by role: 11 scoping, 14 enumeration, 21 specification, 12 comparison, 18 verification, 18 constraint.
The estate answered 19 of the 21 specification sub-queries — 90%, and its best number. It answered 6 of 11 scoping and 4 of 18 constraint. It answered none of the 44 enumeration, comparison and verification sub-queries, and could not have. Answerable fraction: 29 of 94, or 31%, concentrated almost entirely in the one role that only fires after somebody else has named you.
March–June 2026 — the work
Three streams ran in parallel. The 62 sector pages were consolidated into 14 documents, each carrying constraint answers as standalone 120–180 word blocks with the figure in the first sentence. Of the 18 constraint questions, 11 had no adequate answer anywhere on the UK web; those were published with real numbers — test intervals, indicative hood-face velocities, typical installation downtime, indicative costs. On the earned side: two trade titles, the accrediting body’s member directory, three plant-engineering roundups, a regional manufacturing network, a safety publication, and partner pages on two machine OEMs whose equipment they install alongside.
July 2026 — results, and the parts that went badly
Citations rose from 9 of 40 tracked prompts to 23. Median burst width went from 1.4 to 2.6. Constraint answers appeared in 14 of the 23. AI-referred sessions moved from 240 to 690 a month, referring domains rose by 31, and enquiries were up 21% year on year.
Total organic clicks fell 11% over the same period, and the consolidation is the obvious suspect — it cost more than 40 long-tail rankings, knowingly. The 11 constraint pages took five months rather than the planned two, because engineering sign-off on published figures was slow, and two had to be withdrawn and reworded after a professional body objected to an over-general claim. Two of the earned placements were nofollow directory listings that delivered zero referral clicks and appeared consistently in verification retrievals. And the co-fetch analysis showed almost nothing about Google, their largest surface, which nearly killed the method in month two.
One number is worth isolating. Of the 23 prompts they were cited in by July, 14 involved a constraint answer — a question with no established incumbent — while their heavily invested specification pages contributed to citations only where the engine had already named them. The cheapest wins came from the role nobody was competing for, and the most expensive pages worked only after the earned placements had done their job. The sequencing mattered more than the spend.
9. Where this argument could be wrong
The strongest counter is not that fan-out is unimportant. It is that the naming-gate claim rests on thin evidence while the coverage claim rests on thick evidence.
That is fair, and worth stating at full force. EMGI describes its own capture as a pilot rather than a population study: 96 runs, SaaS only, API behaviour that may not match the consumer apps, and non-deterministic systems sampled once. Against it sit 173,902 URLs from Surfer, 863,000 SERPs from Ahrefs and a 54-experiment synthesis from Zyppy, all pointing towards coverage. There is also a genuine confound: a sub-query can contain a brand name because the user named it, not because the model did.
Four things bound that objection without dissolving it. First, the confound cuts in a specific direction — in the Little Green Agency capture the prompt was “best builders leicester”, which names nobody, and the model still produced three named intermediaries inside its own query strings. Second, 86% of ChatGPT’s fan-out queries carrying a year is evidence of the model rewriting toward its own template rather than the user’s words. Third, the coverage studies measure ranking, which is an outcome; Zyppy’s own numbers put search rank at 9.4 against fan-out rank at 9.3, so the field’s best meta-analysis does not actually claim coverage displaces authority. Fourth, new entrants demonstrably do break in — through the unnamed roles, exactly as the argument predicts. YouTube being Perplexity’s single most-cited domain in the EMGI set is an off-domain surface winning, not a vendor page winning.
There is a fifth bound the argument imposes on itself. Fan-out only governs the prompts that retrieve at all, and Gemini answered fifteen of sixteen prompts from memory in the same capture. Where the engine never searches, neither coverage nor corroboration reaches the answer within a session — only what the model absorbed during training does. That is a smaller surface for on-page work than the coverage playbook implies, and a larger one for long-run reputation than either camp usually admits.
Three findings would materially damage the argument. A large-sample replication showing brand-name presence in sub-queries is driven by user phrasing rather than model generation. Evidence that unnamed enumeration sub-queries routinely cite vendor-owned pages. Or an engine publishing first-party data showing decompositions vary with the live index rather than the prior. Watch for all three; the second is the one that would hurt most.
10. The Monday checklist
- Pull six months of access logs, filter to verified retrieval agents, and group into 90-second bursts. Record median burst width and your repeat-membership URL.
- List every URL never fetched by a retrieval agent. That is your consolidation candidate list, not your refresh list.
- Take your ten highest-value prompts. Run each five times across three engines and union the sub-queries you can capture.
- Classify every captured sub-query into the six roles. Count how many name your brand and how many do not.
- Calculate what share of the non-naming sub-queries your own site could ever answer. Stop budgeting content against the rest.
- Find the constraint questions with no good answer anywhere. Publish those with real figures in the first sentence of each block, and date them.
- For every enumeration and comparison sub-query, name the specific third-party page that currently wins it, and decide whether the route in is editorial, directory, accreditation or partner.
- Set one verification target per quarter — a register entry, an accreditation, a named quote in trade press — and track it in the same report as citations, alongside the wider 2026 link and citation benchmarks.
The instinct fan-out rewards is not comprehensiveness. It is knowing which questions you are disqualified from answering, and buying your way into those with somebody else’s credibility rather than more of your own words. The engine wrote nine questions before it looked at your site. You can answer three of them well. The other six are a link-building problem wearing a content-strategy costume — and that is also why what an AI system was trained on and what it retrieves live pull in the same direction, and why the factors behind AI product recommendations keep resolving to corroboration rather than to copy.
