Query Fan-out Explained and Why it's the Citation lever
Key takeaways
|
Here's the mechanism behind almost every "we rank fourth but ChatGPT doesn't mention us" conversation I've had in the last year.
Your buyer types one question. The engine doesn't search it. It rewrites it into a batch of narrower questions, searches each of those in parallel, pulls the best passage for each, and assembles an answer. Then it cites the sources it pulled from. The head term you ranked for was never searched. Eleven things you didn't rank for were.
That's query fan-out. After more than 16 years watching search change, I think it's the single most important technical concept for B2B marketers to understand right now, because it explains why the Citation lever in the Engine Optimization Matrix behaves the way it does.
What is query fan-out?
Query fan-out is the process by which AI search engines expand a single user prompt into multiple related sub-queries, run those searches simultaneously, retrieve the best passages for each, and synthesize a single answer with citations. Google described it as the technique behind AI Mode and complex AI Overviews. ChatGPT, Perplexity, and Gemini all use a version of it. |
The 2026 Query Optimization Survey describes a five-stage pipeline that's consistent across the major engines: intent detection, query rewriting, query expansion, retrieval, and citation synthesis. Source selection happens at the last stage, after the expansion, which is why what you ranked for at stage zero matters so little.
A concrete example. A VP of Infrastructure types "best backup solution for VMware environments with ransomware protection." AI Mode might expand that into something like:
VMware backup vendors 2026
immutable backup storage ransomware
backup vendor Nutanix VMware support
backup recovery time objective comparison
Rubrik vs Veeam vs Cohesity
backup solution Gartner Magic Quadrant 2026
air-gapped backup cost enterprise
backup vendor SOC 2 compliance
Eight searches. Eight retrievals. One answer. The page that ranks first for the original phrase gets one shot. The vendor with a specific, citable passage on six of the eight sub-queries gets six.
How many sub-queries do AI engines run?
Google AI Mode runs 5 to 16 sub-queries per prompt, with 59% of prompts triggering 5 to 11 and complex B2B prompts averaging 9 to 11. ChatGPT runs roughly 2 to 3 on simple prompts and can run hundreds in Deep Research. Across engines, published averages cluster around 8 to 12 for consideration-stage queries. |
The depth differences matter for where you invest.
Google AI Mode is the widest. It expands aggressively, and it's the surface a buying committee uses for multi-turn research. Coverage breadth pays here.
ChatGPT is narrower on a single prompt but has a specific pattern worth knowing. In a July
2026 capture, 69% of ChatGPT's fan-out queries already contained a brand name, and 56% of its citations landed on vendor-owned pages.
The shortlist was set before the first page was fetched. That means the unnamed sub-queries ("best backup vendors 2026") select candidates from third-party surfaces, and the named sub-queries ("Rubrik immutable storage") confirm details from your own pages. Two different jobs, two different places to show up.
Perplexity behaves like a fast, citation-heavy search engine. It runs a moderate fan-out and cites more sources per answer than either of the others. I compared the three in Perplexity vs ChatGPT vs Google AI citations.
Why does query fan-out explain the ranking vs citation gap?
Because the engine never searched the query you ranked for. It searched a set of narrower questions and selected passages that answered each one. Mike King put the overlap between traditional rankings and AI citations at 25% to 39% in January 2026. A page ranking first for the head term but lacking a self-contained passage on any specific sub-query gets skipped in favor of pages that have one. |
This is the Citation lever in the Engine Optimization Matrix, and it's why I built the matrix with Citation as a separate column rather than an outcome of the other four.
Content, Schema, Distribution, and Authority get you into the candidate pool. Citation is whether a specific passage on your page gets selected for a specific sub-query. Fan-out is the selection process. So every cell in the Citation column ("whether your answer is lifted into the snippet," "whether your brand is named in generative outputs," "whether ChatGPT names you specifically") is a fan-out outcome.
The thesis I keep returning to, that AI citations don't correlate with top-10 rankings, isn't a mystery once you see the mechanism. It's the expected result of a system that retrieves passages for questions you didn't target.
And the ranking but not cited problem has a precise diagnosis. Your page answers the head term. It doesn't answer the sub-queries in a form the engine can lift.
What do fan-out sub-queries actually look like?
They're narrower, more specific, and more commercial than the original prompt. Analysis of fan-out traces shows engines injecting "best," "top," "vs," "reviews," and the current year. In B2B software categories, a Q2 2026 study of 1,000 prompts found 70.8% of citations pointed to "best" or "top" lineup pages and 51.6% of cited pages carried the current year in the title. G2 properties took 8% of citations; Reddit took 1.4%. |
Four sub-query types recur, and each points at a different piece of content.
Candidate selection. "Best X vendors 2026," "top X tools." Answered by third-party lineups: G2, analyst reports, comparison posts on trusted domains. You can't publish your way onto these. You earn a place through entity authority and review presence.
Specification. "Does X support Y," "X integration with Z." Answered by your own docs and product pages, if they have a direct sentence for it. This is where BLUF writing earns its keep.
Comparison. "X vs Y," "X alternatives." Answered by comparison content, yours or third-party. Your own comparison page gets cited more than you'd expect if it's honest and specific.
Validation. "X reviews," "X pricing," "X case study." Answered by G2, your case studies, and community threads. Unquotable case studies lose here.
The point of mapping these is that the fan-out is knowable. Not the exact strings, which vary per run (one study found 73% of fan-out queries change between runs), but the themes. A B2B category has a stable set of sub-intents, and you can inventory them.
How do you optimize for query fan-out?
Map the sub-intents a buyer's prompt expands into, audit which ones your site answers with a self-contained passage under a question-form heading, fill the gaps, and get the candidate-selection sub-queries covered on third-party surfaces you don't control. Then verify with the reasoning traces in ChatGPT and Gemini, which expose the actual sub-queries run. |
The workflow I run:
Extract the fan-out. ChatGPT and Gemini show their search steps in the reasoning trace. Run your category's core prompts ten times each, capture the sub-queries, and cluster them. For Google AI Mode, which doesn't expose them, infer from People Also Ask and the citation patterns. This replaces keyword gap analysis for a surface with no keyword tool.
Audit coverage. For each sub-intent, does a page on your site have an H2 phrased as that question and a direct answer in the first two sentences? Use the AEO content audit checklist. Mark each one covered, partial, or missing.
Fill gaps structurally, not with new pillars. Most gaps are fixed by restructuring an existing post so the sub-intent has its own heading and answer block. Chunked content is measurably easier to retrieve. Mike King's January 2026 data put the relevance gain from clear section boundaries at 9% to 15% in vector space. The blog post format I use exists for this.
Match phrasing. Cosine similarity above 0.88 between your passage and the sub-query correlates with dramatically higher citation rates in one analysis of 15,847 AI Overview results. In practice that means using the words the engine searches, not your product marketing vocabulary. Your FAQ section is the easiest place to do this, with FAQPage schema to reinforce it.
Cover the candidate-selection queries off-site. Lineup sub-queries resolve on G2, analyst pages, and trusted comparison sites. That's a distribution and review acquisition job, not a content job.
Track prompts, not keywords. Twenty to forty prompts across informational, comparative, instructional, brand, and transactional intent, checked on a cadence. The citation tracking and citation rate posts cover how.
One caution on the "cover every sub-query" advice you'll see elsewhere. It reads a correlation backwards. Brands cited across many sub-queries are cited because they're trusted, and the trust comes from the Authority and Distribution levers. Coverage without authority produces pages that answer the question and don't get selected. Do both.
Does query fan-out apply to Perplexity and ChatGPT or only Google?
All of them, at different depths. Google AI Mode is the most aggressive, at 8 to 12 sub-queries as standard. ChatGPT expands less on simple prompts, more on complex ones, and enormously in Deep Research. Perplexity runs a moderate fan-out with a high citation count. The mechanism is the same. The breadth and the source preferences differ. |
Which is why the EOM keeps AEO, GEO, and LLMO as separate rows. Same lever, different engines, different retrieval behavior. A page structured for AI Mode's wide fan-out is also structured for ChatGPT's narrower one. The reverse isn't always true.
If you want the fan-out map for your category built out and scored against your current library, book 30 minutes.
About the author
Harold Bell is Founder and CEO of MQL Magnet, a B2B content marketing agency serving enterprise technology brands. He's the creator of the Engine Optimization Matrix (EOM), author of Partner Over Product, and a Forbes Communications Council member.
Frequently asked questions
What is query fan-out?
The process by which AI search engines expand one prompt into multiple sub-queries, search them in parallel, retrieve passages for each, and synthesize a cited answer. Google named it as the technique behind AI Mode.
How many sub-queries does Google AI Mode run?
Five to sixteen per prompt. About 59% of prompts trigger five to eleven, and complex B2B prompts average nine to eleven.
Does ChatGPT use query fan-out?
Yes. It runs roughly two to three sub-queries on simple prompts, more on complex ones, and can run hundreds in Deep Research. About 69% of its fan-out queries already contain a brand name.
Why does my top-ranking page not get cited?
Because the engine didn't search your head term. It searched narrower sub-queries and selected passages that answered each one. Overlap between rankings and citations is estimated at 25% to 39%.
Can I see the sub-queries an AI engine ran?
In ChatGPT and Gemini, the reasoning trace exposes search steps. Google AI Mode and AI Overviews don't show them, so you infer from People Also Ask and citation patterns.
What modifiers do engines add to sub-queries?
"Best," "top," "vs," "reviews," and the current year are common. In B2B software, 70.8% of citations in one study pointed to lineup pages and 51.6% of cited pages had the year in the title.
How do I optimize content for query fan-out?
Map the sub-intents, give each one a question-form heading with a direct answer in the first two sentences, match the phrasing the engine uses, and cover candidate-selection queries on third-party surfaces.
Is query fan-out the same as keyword clustering?
Related but not the same. Clustering groups keywords by topic for ranking. Fan-out mapping identifies the sub-questions an engine runs for one prompt, which are narrower and more commercial than typical cluster keywords.
Does query fan-out change between searches?
Yes. One analysis found 73% of fan-out queries vary between runs of the same prompt. The themes are stable; the exact strings aren't. Optimize for themes.
How does query fan-out relate to the Engine Optimization Matrix?
It's the mechanism behind the Citation lever. Content, Schema, Distribution, and Authority get you into the candidate pool. Fan-out retrieval decides which passage gets cited for which sub-query.
