Keyword research tools are built to measure what humans search for. An AirOps analysis of 548,534 retrieved pages found 95% of the queries ChatGPT actually generates while composing an answer have zero traditional search volume. If you are optimising for primary keywords alone, you are covering at most two-thirds of the AI citation surface. This post covers what fan-out query coverage means and why single-keyword optimisation can actively work against you.
How many sub-queries does ChatGPT generate from one prompt?
ChatGPT generates 25+ unique sub-queries from a single prompt before retrieving any content, then cites pages ranking for those expanded queries, not just the one the user typed.
The effect on citation surface is large. The AirOps analysis found 32.9% of cited pages were discoverable only through fan-out sub-queries, invisible through the original prompt and invisible through traditional keyword monitoring.
95% of those sub-queries had zero traditional search volume. Keyword tools do not fail because they are broken. They measure what humans search. They were never built to measure what an AI system queries internally while assembling an answer.
Does covering an adjacent fan-out query improve AI citation rate?
Covering one adjacent fan-out query more than doubles AI citation rate: pages ranking for both the main query and at least one fan-out query achieved a 51% ChatGPT citation rate, versus 20% for pages ranking only for the main query, per AirOps.
The mechanism is retrieval surface. A page ranking for only the main query is invisible to the 25+ sub-queries ChatGPT generates around it, while a page ranking for even one adjacent sub-query becomes reachable from a second entry point into the same conversation. That second entry point compounds as ChatGPT expands more sub-queries per prompt, not fewer.
| Share of fan-out sub-queries covered | ChatGPT citation rate |
|---|---|
| 26-50% of fan-out sub-queries | 38.2% |
| 100% of fan-out sub-queries | 34% |
Source: AirOps, analysis of ChatGPT fan-out sub-query coverage and citation rates.
Is covering 100% of fan-out sub-queries better than partial coverage?
Covering 100% of fan-out sub-queries measures something different from the earlier lift, which counted a page as covered once it ranked for its main term plus a single extra query, not a share of the full fan-out set. On that share basis, pages covering 26-50% of ChatGPT's fan-out sub-queries were cited at 38.2%, versus 34% for pages covering all of them: in this specific comparison, covering everything did not outperform covering the mid-range share.
Precision explains the gap: a narrowly focused page matches a smaller set of sub-queries more completely, while a page stretched across every possible sub-query answers each one less completely. The practical read: identify the highest-value adjacent queries and build for those specifically, rather than treating fan-out coverage as a checklist to exhaust.
Does optimising for one target keyword hurt AI citations for related topics?
Optimising for one keyword instead comes at a cost: AirOps calls this gain allocation skew, where a page's ranking gain on its main term comes at the expense of cross-query citations for the adjacent topics it also covers.
The 2x lift from covering one adjacent query runs in reverse here: a page tuned exclusively toward its main keyword gives up that same adjacent-query reach, trading a narrow ranking gain for a broader citation loss.
Pages optimised for a single keyword are effectively penalised in AI retrieval for the adjacent queries their topic logically covers. A hub-and-spoke content structure addresses the primary query and the fan-out queries it generates without that trade-off, because no single page is forced to carry every adjacent subtopic alone.
Does getting retrieved by ChatGPT mean getting cited?
ChatGPT retrieves roughly 6 to 7 pages for every 1 that appears in a final answer, discarding 85% of retrieved pages at that next stage; fan-out coverage gets a page into the retrieval pool, but retrieval alone does not mean citation.
Getting from retrieval to citation requires structural answer-extraction patterns: direct answers early in the page, title-query alignment, and clean readability. An Ahrefs analysis of 1.4 million ChatGPT prompts found 88% of citations originate from live search retrieval, not training data: the pages getting cited are pages ranking in traditional search for the specific fan-out query, then clearing content quality filters on top of that.
Does fan-out behaviour differ between AI platforms?
Fan-out processes diverge between AI platforms, not just their outcomes. Google AI Mode and Google AI Overviews retrieve and cite differently despite running on the same underlying search results, evidence their expansion logic differs: AI Mode cites only 12% of top-10 organic results, compared with 38% for Google AI Overviews, and despite 86% semantic answer overlap between the two, they share only 13.7% of cited URLs, meaning each is pulling from a different set of expanded queries against the same ranked pages. Sites outside the top 10 can still earn AI Mode citations through fan-out coverage on adjacent subtopics the top-ranked page does not address.
ChatGPT is also shifting in a way neither Google AI Mode nor Google AI Overviews is reported to: referral reach to unique domains peaked at 260,000 monthly in October 2025, then contracted to 170,000 by February 2026 per Semrush, while queries per session over the same window rose from 1.21 to 1.75, roughly 45% more prompts per conversation. That is a separate trend from per-prompt fan-out (each of those extra prompts still triggers its own 25+ sub-query expansion), but it means brands with comprehensive subtopic coverage face more fan-out events per visitor, not fewer, as conversations run longer.
Map your fan-out sub-queries, then build hub-and-spoke coverage for the highest-value gaps
Map the sub-queries AI systems generate for your primary topics by prompting an AI system directly with something like "generate 20 sub-queries an AI would generate when answering about [your topic]". Audit which sub-queries your existing content already covers, then build focused pages for the gaps that remain rather than attempting to cover every possible variation. Structure content in a hub-and-spoke architecture so no single page bears the gain-allocation cost of single-keyword optimisation.
The fan-out coverage data is observational, not proof of causation
The AirOps data linking fan-out coverage to citation rate is observational, not the result of a controlled experiment proving that adding coverage causes citations to rise; it shows only what coverage looks like on pages already earning citations. The finding that a 26-50% coverage share (38.2% citation rate) beats full coverage (34%) may partly reflect selection effects: pages that focus deliberately on specific subtopics may simply be higher quality overall, not made better by the coverage percentage itself.
The practical action holds regardless. If your keyword research stops at primary terms, you are planning for a fraction of what AI systems actually query. Map the fan-out queries, build focused coverage for the highest-value gaps, and stop tuning individual pages so tightly to one keyword that adjacent citations fall.