Query fan-out: the buyer questions that decide if AI recommends you

A buyer does not ask ChatGPT "what is the best CRM". They ask "which CRM works for a solo consultant who invoices in pesos", "does X integrate with Y", "how much is Z for three users", "X or Y for a small real estate office". The model recommends at the level of that question, and a brand can be invisible in the generic prompt and first in three specific ones.

Google calls this mechanism query fan-out. Its own documentation for AI Overviews and AI Mode says the system "may use a 'query fan-out' technique, issuing multiple related searches across subtopics and data sources, to develop a response" (Google Search Central, AI features). Buyers do the same thing by hand, one specific question at a time. If you only measure the generic prompt, you measure the one question almost nobody wins.

What one generic prompt told us over 373 live queries

For the first months of LLM Audit, the live check asked each model a single question per brand: "what are the best options for [category]". Over 30 days that produced 373 live queries. The brand being audited was named in 3 of them.

That number did not mean every brand was invisible. It meant the question could not discriminate. "Best options for project management" returns the same four names whoever you are, so the audit told almost everyone the same thing and taught nobody where they could actually appear. A measurement that gives every brand the same verdict is not a measurement.

Our methodology still runs that generic prompt, because it is the same question for every brand and it feeds the public index. But it stopped being the headline. The headline now comes from the questions a buyer types before they know you exist.

The four families of buyer questions

When we read a product's own site (home, pricing, the pages it compares itself on) and generate the questions its buyer would type, they fall into four families. The split matters because each one measures something different.

Use case. "What [category] works for [situation]". The buyer describes their problem without naming anyone. This is where discovery happens and where most brands are absent.

Comparison. "[Brand] or [competitor] for [segment]". The buyer already has a shortlist. If you have no named competitor, the comparison is against the default alternative: doing it by hand, cooking at home, staying with the current tool.

Objection. "Is [brand] worth it for [segment]", "how much does [brand] cost", "does [brand] do X". The buyer is close to a decision and the model needs citable content about your product to answer at all.

Context. "[Category] that works in [country, currency, language, stack]". This is the family local businesses win by default and never measure. A tarot reader in Buenos Aires does not compete with the global list; she competes with whoever the model names when the question is asked in Spanish with a city in it.

Up to twelve of these questions in a free run, each asked live to ChatGPT, Gemini and Claude, give you a grid: question by model, with three states per cell. You appear, a competitor appears, or nobody is named clearly. The third state is the opportunity, because the model has no good answer yet.

Open questions are the only ones that count as a recommendation

The first maps we ran taught us a rule we now apply in code. A question is open when your brand name or domain is not in the text of the question. Only open questions count toward "recommended".

The reason is a measurement we got wrong on one of our own products, a wishlist app. The models named it in 24 of 36 answers. Read as a headline, that looks healthy. Split by question type, the brand appeared in 0 of the 12 answers to open questions and in all the ones where the buyer had typed its name. The models knew the product existed. They did not associate it with the situation of anyone who had not heard of it yet. Those are two different problems, and adding them together hid the one that mattered.

So the map labels the two separately. In a comparison or objection question the buyer already names you, and a mention there means "knows you". In a use case or context question a mention means "recommends you". The verdict is built on the second group only.

What "recommended in 0 of 12 open buyer questions" means

It means twelve open questions were asked live and none of the answers named you. Three things it does not mean:

• It is a count. Counts can be repeated and audited; a number out of 100 promises a precision that three models with different opinions do not have.
• It is one day's snapshot. Each cell is one run on one day. Models vary between runs, so a single map is a snapshot, and the right comparison is the same questions asked again after you publish something.
• It says nothing about product quality. It usually means the page that would answer the question does not exist, or exists and does not say the thing the buyer asked.

What it comes with is the part you act on: for each open question, which competitor was named and which questions nobody answered.

What to do the week after

Read the grid by column state, not by row.

Cells where nobody is named. Write for these first. The model has no confident answer, so a page that answers the question directly has a real chance of becoming the answer. Put it on the page that owns the topic, as a section with the question as its heading, rather than on a separate FAQ hub. We wrote up why in FAQ page or H2 sections: what AI models actually cite.

Cells where a competitor is named. This is content to build, and the competitor's name tells you what the model already trusts. If a rival shows up for "which [category] invoices in pesos", your pricing page needs to say, in plain words, that you invoice in pesos. If it shows up for an integration question, the integration page needs to exist and name the other product.

Cells where you appear. Leave them alone, and use them as your control group. If a re-measurement loses these cells too, the change is noise across the board, not a regression in the ones you worked on.

Then wait, and re-run the same questions. Not new questions, because a new set of questions is a new experiment. Two to four weeks is usually enough for the models that search the live web to have seen the page. Read the delta on open questions only.

If you want to see the grid for your own product instead of building it by hand, the free audit reads your site, generates the buyer questions, asks ChatGPT, Gemini and Claude each one live, and shows where you appear, where a competitor does, and where nobody does. The first run is free and asks for no card.