**TL;DR** — Across 24 client sites in April 2026 we audited Google AI Mode's query fan-out: when a user asks one question, AI Mode silently decomposes it into a fan of derived sub-queries, resolves each near-independently, and stitches the answer from the union of their citation sets. We reconstructed the fan-out behind every client citation and mapped which sub-query each citation actually answered. The surprise: 71% of client citations were earned on a narrow sub-query that was not the parent query and that the client had never explicitly targeted — only 29% were earned on the literal parent query. Parent queries fanned into a median of 6 sub-queries. The page that earns the citation is frequently not the page that ranks for the parent query; it is whichever page on the site most cleanly answers one narrow branch of the fan.
Why we ran the audit
Through 2025 and into early 2026 we modelled AI-search citation as a query-to-page contest: a user asks query Q, the engine picks N sources, your page is in the set or out of it. Google AI Mode broke that model. AI Mode does not run your query directly — it runs a fan of derived sub-queries, what Google itself calls "query fan-out," and the citation set of the synthesised answer is the union of the citation sets of those sub-queries. Through Q1 we kept seeing client pages cited on answers where the parent query had little to do with the page's actual topic, and pages we had optimised hard for the parent query absent entirely. The contest we were optimising for was not the contest the engine was running.
There is a second motivation, and it is about reporting. A client looks at a dashboard that says "we were cited on best CRM for startups" and reasonably assumes their CRM-comparison page won that query. Often it did not. AI Mode fanned the query into sub-queries — "how much does a startup CRM cost," "what CRM features do early teams actually need," "CRM data migration for small teams" — and the client page was cited only on the migration sub-query, off three paragraphs the editorial team barely thought about. The parent-query framing then tells that team to keep polishing the comparison table, when the citation came from somewhere else entirely. Misattribution at the query level produces misallocation at the editorial level, and the fan-out makes that misattribution the default.
How we ran the measurement
24 client sites — 9 SaaS, 7 publisher, 5 DTC, 3 B2B services — across April 2026. For each client we ran a 50-query commercial-intent basket through Google AI Mode twice daily via an instrumented browser session. AI Mode does not publish its fan-out, so we reconstructed it: for every parent query that produced a citation of a client URL, we captured the full answer prose, segmented it into discrete claims, and for each claim ran a reverse search to identify the most probable sub-query that claim was answering. We cross-checked the reconstruction against AI Mode's own visible follow-up suggestions and the expandable sub-headings it occasionally renders. The reconstruction is an estimate, not ground truth — we counted a sub-query attribution as confident only when three independent reviewers agreed, which held for 88% of citations.
Two normalisation moves matter. We dropped parent queries that produced no fan-out at all — short navigational and definitional queries that AI Mode answers directly, about 19% of the basket — because the sub-query framing does not apply to them. We also excluded citations where a client URL was cited on two or more sub-queries of the same parent, because the attribution is genuinely ambiguous; those were 7% of citations and we report them separately rather than forcing a single label. The numbers below are for single-sub-query citations on fan-out parent queries, which is the population where the misattribution problem actually lives and the one your editorial team can act on.
The shape of the fan-out
Parent queries that fanned out produced a median of 6 sub-queries, with P25 at 4 and P75 at 11. The fan widened with query breadth: a broad query like "best project management software" fanned into 9–12 sub-queries, while a narrow query like "asana vs monday pricing" fanned into 3–4. The synthesised answer cited a median of 7 distinct URLs — and the number that matters, those 7 URLs were spread across a median of 5 different sub-queries. No single source dominated the answer. AI Mode is not picking the best page for the query; it is picking the best paragraph for each branch of a decomposed query and stitching the branches together.
The misattribution headline followed directly: 71% of client citations were earned on a sub-query that was not the parent query and that the client had not explicitly targeted. Only 29% were earned on the literal parent query or a close paraphrase of it. Put plainly — roughly seven times in ten that a client was cited, they were cited for answering a question they did not know was being asked. This is why parent-query rank tracking has quietly stopped describing reality: the page that earns the citation is frequently not the page that ranks for the parent query, it is whichever page on the site most cleanly answers one narrow branch of the fan.
Driver one: sub-query specificity beats parent-query authority
The strongest pattern in the data: AI Mode resolves each fan-out sub-query close to independently, and on a narrow sub-query a narrow, specific page beats a broad, authoritative one. On the parent query "best CRM for startups," the most authoritative comparison pages were cited 34% of the time on the parent-style sub-query — but on the sub-query "CRM data migration for small teams," the page cited 61% of the time was a narrow how-to about migration, often from a domain with far weaker overall authority. The fan-out structurally rewards depth on a single sub-topic over breadth across the parent topic. A site with eight focused pages can collect eight sub-query citations off one parent query; a site with one excellent pillar page collects one or two.
This inverts a piece of 2024-era consolidation advice. The "merge thin pages into one comprehensive pillar" strategy — sound for traditional ranking, where one strong URL beats several weak ones — works against you in the fan-out era, because a pillar page answers every sub-query adequately and no sub-query excellently, and "adequately" loses every fan-out branch to a page that answers that one branch excellently. We have not reversed the consolidation guidance wholesale; thin pages are still thin. But we now ask, before merging two pages, whether each was winning a distinct fan-out sub-query. If they were, merging them spends two citations to buy one.
Driver two: the sub-query your page wins is the one in your H2s
We could predict which sub-query a page would be cited on with reasonable accuracy from a single signal: the page's H2 set. When a fan-out sub-query closely matched the text of an H2 on the page, that page was cited on that sub-query 3.1× more often than a page that answered the same sub-query in body prose without a matching H2. The fan-out, in effect, matches sub-queries to headings — it decomposes the parent query, then looks for pages whose heading structure advertises an answer to each branch. A page's H2 set is its menu of fan-out sub-queries it is eligible to win.
Operationally, we now treat the H2 outline of a commercial page as a deliberate list of fan-out sub-queries to claim, not as a readability convenience. Before publishing, we run the page's target parent query through a fan-out reconstruction, list the likely sub-queries, and check that the two or three most valuable each have a matching H2 with a tight answer paragraph directly below it — the post-H2 extraction pattern from our cited-paragraph audit compounds here, because winning the sub-query and being the extracted paragraph are two separate gates and the H2 helps with both. Pages edited this way picked up a median of 2.4 additional sub-query citations within six weeks, on questions the page had always answered in prose but never advertised in a heading.
Driver three: fan-out citations cluster on the same few domains
A discouraging finding for new entrants: across all sub-queries of a parent, the citation set was less diverse than the fan-out structure suggests it should be. AI Mode fanned a query into 6 sub-queries but drew its answers from a citation pool only modestly larger than a non-fanned answer — a median of 7 URLs for a 6-sub-query fan, against 4–5 for a direct answer. The same handful of domains recurred across multiple sub-queries of the same parent. A domain cited on one sub-query of a parent was 2.6× more likely than baseline to be cited on a second sub-query of the same parent. Fan-out rewards depth, but it also carries a within-parent incumbency effect: once AI Mode trusts a domain for one branch, it reaches for that domain again on adjacent branches.
The practical reading is that the fan-out is an opportunity for a site that already holds one citation on a parent query and a liability for a site holding none. If you hold one sub-query citation, building focused pages for the adjacent sub-queries is unusually high-yield, because the incumbency effect is working for you. If you hold none, a single broad page is unlikely to break in — the realistic entry point is one narrow, excellent page targeting the single sub-query where the incumbent pool is weakest, which is usually the most specific and least commercial branch of the fan. Win that branch first, then let the incumbency effect carry you into the adjacent ones.
What changed in our content checklist
Four additions. For every commercial-intent parent query, we now reconstruct the likely fan-out before briefing the page, and we treat the resulting sub-query list as the page's section outline. We write H2s that mirror the highest-value sub-queries close to verbatim wherever it reads naturally, because the H2-to-sub-query match was the strongest citation predictor we found. We stopped reflexively consolidating pages that each win a distinct sub-query, and we added a "which sub-query does this win" question to every merge decision. And we report citations at the sub-query level, so the client sees "you were cited on the migration sub-query of best startup CRM," not the misleading "you were cited on best startup CRM."
We dropped one habit. Through 2025 we built parent-query rank-tracking dashboards for AI search the same way we built them for the ten-blue-links era — one row per query, cited or not cited. In the fan-out era that row is an average over a hidden decomposition, and the average hides exactly the information an editorial team needs to act. The dashboard now carries a row per (parent query, sub-query) pair. It is roughly six times longer and far more useful, and it is the single reporting change that has most improved where our editorial hours go in 2026.
- 01Reconstruct the fan-out before you brief the page. Parent queries fanned into a median of 6 sub-queries, and 71% of client citations were earned on a sub-query, not the parent query.
- 02Write H2s that mirror the highest-value sub-queries. A sub-query that matched an H2 was cited 3.1× more often than the same answer buried in body prose.
- 03Stop consolidating pages that each win a distinct sub-query. A pillar page answers every sub-query adequately and no sub-query excellently — and "adequately" loses every fan-out branch.
- 04Report citations per (parent query, sub-query) pair. "Cited on best startup CRM" is an average over a hidden decomposition; the sub-query is where the editorial signal actually lives.
Where this argument breaks
For navigational, definitional and very narrow queries — about 19% of our basket — AI Mode answers directly without fanning out, and the whole sub-query framing collapses back to a straightforward query-to-page contest. For very small sites with only one or two commercial pages there is no portfolio to spread across the fan, and the analysis reduces to making that page as deep as possible on its single best sub-query. Our fan-out reconstruction is an estimate, not a published structure, and Google can change the decomposition logic without notice — the 71% figure is a snapshot of April 2026, not a constant. In Chinese-language search, 文心 and 通义 fan queries far less aggressively than Google AI Mode, so the misattribution problem is correspondingly smaller, though both engines are trending toward more decomposition. Outside those carve-outs, the fan-out is the most important structural change in how AI search picks sources since AI Overviews launched, and most teams are still optimising for a query the engine no longer runs.