GEO, Measurement
When a buyer asks an AI one question, the engine quietly turns it into about ten. Seer Interactive measured 10.7 hidden sub-queries per prompt across 501 prompts on Gemini 3. Most sites answer the headline question and none of the ten. Here is how to find your gaps.
You type one question into an AI. Behind the glass, the engine does not go and look for that question. It writes a list of related questions, searches for all of them, and assembles an answer from whatever it finds across the lot.
The industry name for this is query fan-out. It is the single most useful thing to understand about how AI search picks sources, because it explains why a page that matches the buyer's exact wording still does not get cited.
Seer Interactive measured the fan-out directly by running 501 prompts through the Gemini 3 API with grounding forced on, so the interface returned the sub-queries it had generated. The average was 10.7 sub-queries per prompt, with a range from 3 to 28.
Two honest caveats before anyone builds a strategy on that number. It was measured on Gemini 3 specifically, not on all four engines, and the same research put ChatGPT's fan-out at roughly a fifth of Gemini's. It also jumped 78% from the previous Gemini version, which tells you the number is a moving target rather than a constant.
The shape holds anyway. One question in, many questions searched, one answer out. Even at the low end of the range, the engine is looking for things the buyer never typed.
Here is the failure this causes.
A buyer asks: "what is the best accounting software for a small UK building firm?" You sell exactly that. You have a page titled almost exactly that. You do not get cited.
Underneath, the engine has fanned out into something like: CIS deductions, VAT reverse charge for construction, subcontractor payments, integration with a common banking app, pricing tiers, whether it works on a phone on site, migration from spreadsheets, support hours, contract length, and reviews from other builders.
Then it assembles the answer from sources that covered those. You covered the headline. Someone else covered the ten. The answer names them.
This is also why competitors who look weaker than you on the surface keep appearing. They are not outranking you on the money question. They are the only source answering sub-question 7.
You do not need a tool for this. You need an afternoon and some honesty.
One. The single question that, if a buyer asked it and got your name, would be worth the most. Write it in the buyer's words, not yours. "Best accounting software for a small UK building firm", not "construction sector financial management solutions".
Now be the buyer, not the seller. What would you actually need to know before choosing? Aim for ten. Force yourself past the easy five, because the useful ones are at the bottom of the list.
Two shortcuts that help. Ask each of the four engines the money question and read which specifics they raise unprompted, because those are fan-out results leaking into view. And read your own sales calls or enquiry emails for the questions people ask before they buy, which are the same questions in different clothes.
For each of the ten, name the URL on your site that answers it. Not "we mention that somewhere". A page where a reader arriving cold gets a straight answer.
Most businesses doing this exercise get one match out of ten, sometimes two. That is not a sign the business is bad. It is a sign the site was written for someone who already decided to look at you.
Not all ten are equal. The ones nearest the moment of decision are worth more. A page answering "does it handle CIS deductions" earns you a citation at the point a buyer is choosing. A page answering "what is accounting software" earns you a citation nobody acts on.
Write plainly, answer the question in the first two sentences, and keep it as text. Engines lift text. See what AI search engines actually read on your site if you are unsure what survives the trip.
Fan-out is also why we do not test a business with a handful of brand-name questions and call it a scan.
We test 5 buyer-intent buckets, mapped to the stages of a purchase:
Split that way, the pattern that shows up most often is a business that is visible at the top of the funnel and invisible at the bottom. Plenty of citations for the broad discovery questions, because they wrote blog posts. Nothing at all for comparison and high intent, because nobody writes the awkward pages.
A single blended visibility score hides that completely. So does a scan built only from brand questions, which mostly tells you whether the engine can read your homepage. The buckets exist so you can see which stage of the buying journey you disappear at, which is the difference between a number and an instruction.
It is the same reason we split AI traffic into two numbers instead of one, and the same reason a mention is not automatically worth anything until you know which of the three kinds of mention it is.
Fan-out is not a licence to publish ten thin pages this week.
Ten pages that each half-answer a question do less than three that answer their question properly, and mass-produced filler is exactly what engines are getting better at discounting. If you can only do three of the ten well, do three.
It also does not replace the off-site work. Sub-queries about reviews, comparisons and reputation get answered by pages you do not own, which is why mentions behave like the new backlinks. Your own site can only win the sub-questions where you are a credible first-party source.
And none of it matters if the engines cannot fetch you at all. Check the door before the decoration: is your site blocking ChatGPT.
Take one question. Write the ten behind it. Count how many of your pages answer them. That number, honestly counted, is a better brief than most content plans you will be sold.
Then find out which of the engines already name you and which do not. Run a free scan at whocanfindme.com. No signup. Takes about ten seconds.
The same question, same engine, same day can produce different answers. That is the nature of AI search, not a bug. The 3 rules we use to avoid showing you fake statistics, and how to read a visibility score without fooling yourself.
Leer másClaude searches the web through Brave. Gemini leans hard on YouTube. Perplexity is the most citation dense of the four. Where each engine gets its sources changes what you should do, and it is why we score every engine separately instead of blending an average.
Leer más