Ask an AI engine one question and it does not run one search. It decomposes your question into the sub-questions it needs answered, runs those, reads far more than it uses, and assembles a response. The searches it runs are mostly not searches anyone has ever typed, which is a problem if your entire content plan was built from things people type.
AirOps put numbers on this in an analysis of 15,000 prompts. Those prompts produced 43,233 distinct queries, an expansion of nearly three to one. Almost 90% of prompts triggered at least two follow-up searches. And the queries generated in that expansion, the fan-out queries, are overwhelmingly phrases with no measurable search demand behind them.
What fan-out queries are
A fan-out query is a search the engine writes for itself. You ask which project management tool suits a distributed engineering team of forty. The engine does not search that string. It searches something closer to project management tools for distributed teams, then engineering team project management pricing tiers, then a couple of vendor-specific checks, then possibly a comparison between the two candidates it has short-listed.
Each of those is a machine-written query optimized for retrieval, not a human query optimized for typing. Humans compress, because typing is effort and search boxes reward brevity. Machines expand, because retrieval quality improves when the query is specific and there is no cost to writing five of them. That difference is the entire phenomenon, and it means the vocabulary of AI search is drifting away from the vocabulary of the keyword tools that measure it.
| WHAT THE USER ASKED | WHAT THE ENGINE SEARCHED | SEARCH VOLUME |
|---|---|---|
| Which PM tool for a 40-person distributed eng team | project management tools distributed teams | Measurable |
| project management pricing per seat engineering | Effectively zero | |
| comparison of the two shortlisted vendors | Effectively zero | |
| integration with existing developer tooling | Effectively zero |
Only the first row would appear in a keyword tool with meaningful volume attached to it. The other three are where the specifics get decided, and specifics are where a vendor either gets named or quietly drops out of the shortlist. Nobody loses a deal on the category question. They lose it on the pricing tier, the integration, or the scale limit.
The retrieval funnel, stage by stage
The most useful thing in the AirOps data is that it lets you see retrieval as a funnel with measurable drop-off, rather than as a black box. Across those 15,000 prompts, 548,534 pages were retrieved and 82,108 citations appeared in final responses. That is a citation rate of roughly 15%, which means 85% of everything the engine pulled was read and discarded.
AirOps, March 2026: citation rate by prompt intent, share of retrieved pages that made the answer.
Being retrieved is necessary and nowhere near sufficient. Most GEO advice is really about retrieval: be crawlable, be structured, be present. That gets you into the 548,534. The 82,108 is a different competition, decided by whether your page answered the specific sub-question cleanly enough to be worth quoting.
The intent split is worth noting too. Validation prompts, the ones where a buyer is checking a claim or confirming a decision they have already leaned toward, cite the least. Product discovery cites the most. If your content is all bottom-funnel reassurance, you are competing in the stage with the lowest citation rate.
Why keyword volume cannot see this
Here is the structural problem stated plainly. Ninety five percent of fan-out queries have zero traditional search volume, and 32.9% of the pages that got cited were discovered only through fan-out results rather than through the original query's results. So roughly a third of the citation opportunity sits behind queries your research tool reports as nonexistent.
A content plan built on volume thresholds will systematically exclude that third. Not because the planner made a bad call, but because the input data does not contain it. This is a measurement blind spot rather than a judgment error, which is why smart teams keep walking into it.
| PLANNING INPUT | WHAT IT CAPTURES | WHAT IT MISSES |
|---|---|---|
| Keyword volume | What humans type into a search box | Everything a model writes for itself |
| SERP competition | Who ranks for typed queries | Retrieval competition on sub-questions |
| Topic clusters from a tool | Semantically related typed queries | The decomposition an engine actually performs |
| Prompt testing | How engines answer real buyer questions | Nothing, this is the input that works |
The last row is the fix and it is unglamorous. You cannot buy fan-out volume data because there is no volume to measure. You can observe the decomposition directly by running your buyers' real questions through the engines and recording what gets cited and for which sub-question. That is a research method, not a tool purchase, and it is the only reliable input available right now. It sits naturally alongside the argument that topical authority comes from depth rather than breadth, because depth is what lets one topic survive several rounds of decomposition.
Ranking still matters, just not the way you think
It would be convenient to conclude that classic search is irrelevant here, and plenty of people have drawn exactly that conclusion out loud this year. The data says the opposite, and anyone selling generative engine optimization as a replacement for search engine optimization rather than a layer on top of it should read this part twice before their next pitch.
In the same dataset, 55.8% of cited pages ranked in Google's top 20 for something, and pages ranking first were cited about 3.5 times more often than pages outside the top 20. Kevin Indig's separate analysis of roughly 98,000 citation rows found the same shape: pages ranking first in Google were cited 43.2% of the time, again around 3.5 times the rate of pages beyond position 20.
“Ranking is not the mechanism, it is the qualifier. Engines retrieve from a pool that classic ranking helped assemble, then choose within it on entirely different criteria.”
So the two disciplines stack rather than substitute. Ranking well gets your page into the retrieval pool for the queries that do have volume. Writing extractable, self-contained answers to sub-questions gets you cited once you are there. A team that does only the first gets retrieved and discarded. A team that does only the second never gets retrieved. We have made a version of this case before in how AI engines decide what to cite, and this data is the cleanest quantitative support for it yet.
Indig's work adds one more constraint worth planning around: concentration. About 30 domains captured 67% of citations within a topic, and the top 10 took 46%. Meanwhile 58% of cited URLs were cited only once. The head is dominated by a handful of domains and the tail is enormous and shallow. If you are not one of the 30, your realistic path runs through the tail, which means many specific pages rather than a few broad ones.
Planning content for fan-out queries
The practical shift is from keyword briefs to question-cluster briefs. Same amount of work, different organising principle, and it survives contact with query expansion in a way a keyword brief does not. A keyword brief tells a writer what phrase to rank for. A question-cluster brief tells them which specific questions the page has to answer completely, which is the thing the engine is actually assessing.
Length is worth a sentence here because the data pushes against the current fashion for short answer-first pages. Indig found pages between 5,000 and 10,000 characters showed the biggest citation lift, and pages over 20,000 characters averaged 10.18 citations against 2.39 for pages under 500. Long pages contain more distinct passages, and more distinct passages means more chances to match a sub-question. That is not an argument for padding. It is an argument for covering the whole cluster in one place rather than splitting it into thin pages.
The format that does this best is still the comparison, for the same reason it always has: a good comparison naturally answers a dozen sub-questions with clear verdicts attached. We covered why comparison content earns the largest share of citations and the fan-out data explains the mechanism underneath that finding.
The first sprint
Two weeks, one topic, no new tooling. Week one: write twenty real buyer prompts for your highest-value topic, run each through ChatGPT, Gemini, and Google AI Mode, and record every source cited plus which part of the question it answered. Week two: build the decomposition map from what you observed, find the sub-questions where a competitor is cited and you are not, and write those sections into your existing best page rather than starting a new one.
Then re-run the same twenty prompts in thirty days and count sub-question wins, not traffic. Traffic will lag and may never fully arrive, since a citation frequently produces no click. The sub-question win rate is the leading indicator, and it is the one you can actually move. Teams running a structured content marketing programme can fold this into an existing brief template in an afternoon.
Sources: AirOps analysis of 548,534 retrieved pages across 15,000 prompts, via Search Engine Land, March 13, 2026 (coverage); Kevin Indig's citation domain analysis, via Search Engine Land, March 24, 2026.
See where you are cited today
A free snapshot audit of your rankings and AI citations before we ever talk.
Tyler leads work at the intersection of SEO and generative engines at Something Inc., helping B2B brands get ranked and cited across every major AI engine.