Skip to content
Content strategy - GEO

Read the fan-out. Write the answer. Win the citation.

When someone asks an AI engine a question, the engine does not run that question. It runs a spray of its own smaller searches, reads what comes back, and writes an answer from the sources it found. Those hidden searches are the actual brief. If you can see them, you know precisely what the engine wanted answered and which of your pages failed to answer it. So why are you still guessing at keywords for a query nobody typed?

George, the Baseline Labs mascot, holding a paper fan

One question in, two dozen searches out

Google names this behaviour in its own documentation. Both AI Overviews and AI Mode "may use a 'query fan-out' technique, issuing multiple related searches across subtopics and data sources", per Google Search Central. Its AI optimisation guide works a single query into three named fan-out sub-queries, and the AI Mode launch post describes a plan-search-adjust loop, using a question about sleep tracking that fans into separate searches for smart rings, smartwatches and tracking mats. Scale it up and the numbers get serious: Deep Search "can issue hundreds of searches" and return a fully cited report in minutes, according to Google's I/O 2025 update. On the visual side, a Google engineering director put it at "a dozen searches" in the time it takes to do one, from a single photo of a garden.

1
Ingest the prompt
The user asks one question. This is the last thing in the process a human writes.
Best trail shoes for wide feet
2
Turn it into fan-outsThe key step
The LLM chooses these searches itself - nobody typed them. This is the step the audit reads, and the step you can write for.
Best trail shoes 2026 Trail shoes wide toe box Trail vs road shoes Shoe brands for wide feet
3
The fan-outs search and return
Each search runs and brings back its own pages. The accented page answers the accented fan-out.
4
Ingest the results
Everything the searches returned is read back into the model as context.
5
Write the answer, with sources
One answer, citing the pages that answered. Citation [2] is the page that answered the accented fan-out.
[1][2][3]
The five steps behind every AI answer, on a made-up question. Step 2 is the fan-out: the model, not the user, decides what gets searched. That makes it the one step you can read and write for - which is the rest of this post.

Here is what that looks like on a real brand. In an Ireland scan, the third of three questions was "What stout do Dubliners actually drink". That one question produced 24 sub-queries and 50 citations underneath it. Gemini fanned it into six searches, including "Most popular stout in Dublin" and "Craft stout Dublin". ChatGPT ran the question unchanged as a single search. Across the whole run, reddit.com came back eight times as a source, more than any other domain.

24
Sub-queries behind one question in a real Ireland scan
50
Citations returned under that single question
8
Times reddit.com was cited, the top domain of the run
What stout do Dubliners actually drink
True fan-out Proxy
Gemini6 searches the engine reported running
Most popular stout in Dublin What stout do people drink in Dublin Dubliner's favorite stout Guinness popularity in Dublin Craft stout Dublin What stout do Dubliners drink
ChatGPT1 search, the question unchanged
What stout do Dubliners actually drink
Google12 proxies - People Also Ask + related searches
What beer do the Irish really drink? Which Irish stout is considered the best? Do Irish actually drink Guinness? What stout is better than Guinness? What stout do Dubliners actually drink in Ireland What stout do Dubliners actually drink reddit What stout do Dubliners actually drink Guinness Irish Stout brands Traditional Irish beer brands Murphy's Irish Stout O'Hara's Irish Stout Killian's Irish beer
Perplexity5 proxies - related questions
Do Dubliners prefer Guinness over Beamish or Murphy's? Which Dublin pubs pour the best pint of stout? What makes a perfect pint of Guinness in Dublin? Is Guinness really the Most popular stout in Dublin? Are there local stouts Dubliners drink besides Guinness?
All 24 sub-queries the four engines produced for one question, as captured in the Ireland Fan-out Audit run. Gemini and ChatGPT report their own searches; the Google and Perplexity rows are proxies, shown dimmed because they are inferred rather than reported.

Note the difference between the two engines. One question became six searches on Gemini and one on ChatGPT. If you optimise for the question you think a customer asks, you are writing for the one case out of two where the engine did not decompose it.

True fan-out, and honest proxies

Not every engine will tell you what it searched. Gemini and ChatGPT report the searches they run, and those are capturable through the providers' APIs. That is true fan-out: the engine's own account of its work. Perplexity and Google do not expose theirs, so the best available signal is a proxy, which is related questions on Perplexity and People Also Ask plus related searches on Google. Proxies are useful. They are not the same thing, and a report that mixes them into one list is selling you a confidence you have not got. The Fan-out Audit labels each source separately and never blends them.

EngineWhat can be captured
GeminiTrue fan-out, the searches the engine reports running
ChatGPTTrue fan-out, the searches the engine reports running
PerplexityProxy only, related questions
GoogleProxy only, People Also Ask and related searches

Two caveats. First, API-side behaviour is close to the consumer apps but not identical, and Google itself warns that AI Mode and AI Overviews "may use different models and techniques, so the set of responses and links they show will vary" (Search Central). Second, be sceptical of anyone advertising a full view of hidden queries with no method attached. One vendor markets "dark query discovery" that "maps fan-out sub-queries with zero Google search volume" and publishes no sample size or methodology anywhere on the page. Ask which engines expose what, and treat a non-answer as an answer.

The reason you need to observe fan-out rather than infer it is that your keyword tool structurally cannot see it. Seer Interactive forced grounding across 501 prompts and found 95% of the fan-out queries Gemini generated had zero global search volume. Ahrefs reports the same, noting that over 95% receive no recurring searches even against a database filtered from 110 billion discovered keywords down to 28.7 billion. Research from 85sixty and Nectiv both landed on about 95%, with Nectiv's analysis covering 60,000-plus fan-out queries. The mechanism is length: fan-out phrases average 5.5 words on ChatGPT and 9.1 on Gemini against roughly 3.4 for a classic Google search, which puts them below the threshold volume tools bother to track. DataForSEO's sample of 100,000 ChatGPT prompts and 100,249 fan-out queries shows the same shape, with fan-out queries clustering at 61 to 90 characters against 31 to 60 for the original prompt, and states the obvious consequence: "you can't manually explore or fetch fan-out queries because they are hidden from users". We covered why keyword research is blind to this in stop optimising for keywords, start optimising for sub-questions. This piece is about what to do once you can see them.

Average words per query
Typed Google search
3.4
ChatGPT fan-out query
5.5
Gemini fan-out query
9.1
Average query length in words, from the Seer Interactive, Ahrefs and DataForSEO samples above. Volume tools are built around the short grey shape; the two blue shapes are what actually retrieves your pages.

A citation on one engine is not a citation on three

Growth Memo analysed 3.7 million URL citations from a 20,000-prompt random sample across three engines. Only 2.37% of cited URLs appeared on all three, and 91.07% appeared on exactly one. Citation turns out to be three separate races with almost no shared podium, which is why an audit that shows you each engine's fan-out separately is worth more than an aggregate score.

2.37%
Of cited URLs appear across all three engines
91.07%
Appear on one engine only

The same analysis shows which page types travel. Overlap is lowest for informational queries at 2.0% against 2.4% for commercial, and lowest of all for homepages at 1.1%, with product pages barely better at 1.2%, against 2.3% for guides and tutorials. Your homepage is the page least likely to be cited by more than one engine. The explainer that answers one sub-question properly is roughly twice as likely. That is a content brief, not a branding note. It also matches what shows up in citation logs elsewhere: Khalid Hamadeh's GrantCompass case study, built on 192,924 Copilot citations, found the search volume of most page-retrieving queries is zero.

What you actually do

Four steps, in this order. The order matters, because without a baseline you have opinions rather than a measurement.

1. Take the baseline. Run a visibility report on the questions that matter to your business, before you change anything. This is your before picture: who gets cited on your topics today, how often you appear, and on which engines. Save it. You will compare against it.

2. Read the fan-out for the gaps. Run a fan-out audit on the same questions. It captures the sub-queries each engine ran and the sources they cited back, then checks every sub-question against your own pages. The sub-questions nothing on your site answers come back as a writing brief. You also get a most-cited-domains ranking across the run, which tells you who currently owns the answer. In the Dublin example that was reddit.com, which is a useful thing to learn before you write a landing page nobody will cite.

3. Write to the gaps. Take the unanswered sub-questions and answer them directly, in the words the engine used, on pages that exist to answer them. Given that guides and tutorials outperform homepages on cross-engine citation by roughly two to one, put the answer in a page that is structured as an answer. One sub-question per section, stated as a heading, resolved in the paragraph underneath.

4. Re-run and compare. After the new content is live and crawled, run the visibility report again on the same questions and diff it against the baseline. The change in citations is your result. If a gap you filled did not move, that is information too: either the engine is not fanning to that sub-question any more, or something else answers it better, and the fan-out audit will show you which.

Then repeat. Fan-out is not stable, models change, and a brief written six months ago describes an engine that no longer exists.

Run a fan-out scan See the searches engines run on your questions

Fan-out as research, before you write anything

The loop above assumes you already have content to audit. The tool works just as well pointed at an empty page. Take the area you are thinking of writing about, ask the questions a customer would ask, and read the fan-out. You get a map of what people and machines actually want to know about that topic, in their own phrasing, with the currently cited sources next to each branch. That is a better commissioning brief than a keyword list, because the branches are the sub-questions the engine has already decided the answer needs, and roughly 95% of them will never appear in a volume tool.

Used this way it also settles arguments early. If every engine fans your topic into six sub-questions about price and two about specification, the piece your team wanted to write about specification is the wrong piece. Better to learn that from the fan-out than from four months of no citations.

Questions people ask about this

Is the fan-out captured through an API identical to what the consumer app runs?
No. It is close, but it is API-side behaviour, and Google states outright that different surfaces may use different models and techniques, so the responses and links will vary. Treat the fan-out as a strong signal about what the engine wants answered, not as a byte-for-byte transcript of one person's session.
Why can I not just pull these queries into my keyword tool?
Because around 95% of them have no search volume to look up, in three independent samples. They are machine-generated phrasings, longer and more specific than anything a person types, and volume-ranked tools filter exactly that shape out. The queries are real and they retrieve pages. They are simply not keywords.
How many sub-queries should I expect per question?
It varies by engine and by question. In the Ireland scan, one question produced 24 sub-queries across the engines, with Gemini contributing six and ChatGPT running the question unchanged as one search. Google's own examples range from three sub-queries for a simple query up to hundreds for Deep Search. Plan for a handful per engine on an ordinary commercial question.
Do I need a new page for every gap the audit finds?
No. Most gaps are sections, not pages. Group related sub-questions into one guide and give each its own heading and a direct answer underneath. Cross-engine citation data favours guides and tutorials over homepages and product pages by about two to one, so depth in one well-structured page usually beats a scatter of thin ones.

Sources: Google Search Central AI features and AI optimisation guide, the Google blog on AI Mode, its I/O 2025 Search update and its visual search explainer, Seer Interactive, Ahrefs, WebSearchAPI and Nectiv, DataForSEO, Khalid Hamadeh, Ekamoira and Growth Memo, drawn from publicly available reports and studies from 2025-2026. Scan figures are from a real Ireland Fan-out Audit run on baselinelabs.ai.

References (12)
  1. Google Search Central (developers.google.com)
  2. AI optimisation guide (developers.google.com)
  3. AI Mode launch post (blog.google)
  4. Google's I/O 2025 update (blog.google)
  5. "a dozen searches" in the time it takes to do one (blog.google)
  6. no sample size or methodology anywhere on the page (ekamoira.com)
  7. 95% of the fan-out queries Gemini generated had zero global search volume (seerinteractive.com)
  8. 110 billion discovered keywords down to 28.7 billion (ahrefs.com)
  9. both landed on about 95% (websearchapi.ai)
  10. 100,000 ChatGPT prompts and 100,249 fan-out queries (dataforseo.com)
  11. 2.37% of cited URLs appeared on all three, and 91.07% appeared on exactly one (growth-memo.com)
  12. 192,924 Copilot citations (khalidhamadeh.com)

See how AI search sees your site

Free AI visibility check in under a minute

Check my site
George
Online
0%