How AI Search Decides Which Brands to Name

An assistant asked to recommend a brand does not consult a ranked list. It turns the question into searches, reads a small number of retrieved pages, and writes an answer from what it found. Three steps, each of which can exclude you, and none of which can be bought.

A note on confidence before the detail: the internals of these systems are proprietary and change often. What follows is inference from documented behaviour and from running a great many prompts, not a specification. We would rather say that plainly than present a diagram as fact.

Step one: the question becomes several searches

"Which retinol is safe for sensitive skin" is not run as a single query. The assistant decomposes it — sensitive skin retinol, retinol irritation, gentle retinol alternatives, encapsulated retinol — and searches for several of them at once.

This is the first place brands are lost, and it is invisible from the outside. If your content is optimised for the phrase a marketer would choose rather than the sub-questions a worried person actually has, you are absent from the retrieval before any judgement about your brand has been made. Nobody rejected you. You were never in the room.

The practical implication is that covering a question thoroughly beats matching a keyword precisely. A page that genuinely addresses irritation, tolerance, formulation and alternatives will surface across several of those decomposed searches. A page built around one keyword will surface for one, if that.

Step two: a handful of pages get read

Retrieval returns far more than the model uses. Something then selects a small subset — often a dozen or fewer — to actually read.

What appears to survive that filter is not what a marketer would predict. Pages that answer the specific question directly do well. Pages heavy on brand language and light on substance do badly, because there is nothing in them to extract. Recency matters in categories that move. And a page that requires interaction to reveal its content — an accordion, a tab, a script-rendered block — risks being read as empty.

This is where structured data earns its place. It does not persuade a model to like you. It removes ambiguity about what a page is, who wrote it, when, and what it claims, which makes the page cheaper to use and easier to attribute.

Step three: the answer gets written

The model now composes a recommendation from what it read. Two things govern which brands survive into the final text.

Agreement. A brand described in compatible terms across several independent sources is safe to name. A brand described enthusiastically by exactly one source, its own website, is not. This is not a rule someone wrote; it falls out of how a model trained to be accurate behaves when sources disagree.

Attributability. Where citations are shown, the model needs a source it can point at. A claim it cannot attribute is a claim it is safer to leave out. This quietly advantages brands whose facts appear in places that look like references — trade press, retailer pages, review platforms — over brands whose facts exist only in marketing copy.

Why a large budget does not help

Put those three steps together and the explanation for the most common complaint we hear becomes obvious. A brand can spend heavily on paid media, dominate every feed its customers use, and still be entirely absent from the answer to the question it most wants to own.

Advertising buys attention. It does not put text on the open web for a crawler to retrieve, it does not create independent sources describing the product in compatible terms, and it does not make a claim attributable. The mechanism simply has no input for spend. We saw this pattern clearly in our analysis of ingredient-led beauty brands: the brands winning citations were frequently not the brands winning share of voice.

The uncomfortable corollary

If corroboration is the mechanism, then the fastest route to being named is not on your website at all. It is having more people write about you accurately, which is slower, less controllable and less measurable than anything a marketing team is used to buying.

That is genuinely awkward, and it is why so much of the advice in this category quietly avoids it in favour of technical checklists. The checklists matter — they are the part we work through first — but they get you considered rather than chosen.

What this means in practice

Write for the sub-questions rather than the head term, because retrieval happens at that level. Make sure the substance is in the HTML rather than behind an interaction. Say what you claim in language other people could repeat, because they will need to repeat it for it to count. And accept that the highest-leverage work is getting independent sources to describe you consistently, which looks like public relations and is now also search.

The definitional version of all this is in our working definition of generative engine optimisation, and the honest limits of what any of it can promise are set out there too.