How assistants decide which brands to name

Everything an assistant says comes from one of two places: what it absorbed during training, and what it retrieved at the moment you asked. The distinction matters enormously for brands, because you can influence the second within weeks and the first barely at all.

Training gives the model its default picture of your category — which brands exist, roughly what they do, how they relate. That picture is frozen at the knowledge cutoff and updates only when a new model ships.

Retrieval is live. When an assistant searches, it reads pages as they are today, which is why a correction published this month can appear in an answer this month. Nearly all practical AEO works through this channel.

Retrieval does not match keywords. It converts your question into an embedding — a numeric representation of its meaning — and finds passages whose embeddings sit close to it. A page can be retrieved for a question sharing none of its words.

This is why writing naturally beats repeating an exact phrase. The system is comparing meaning, and a paragraph that answers the question clearly scores well whether or not it echoes the wording.

It also explains why the passage, not the page, is the unit that matters. A long article covering five subjects has no single passage that strongly matches any one question; five focused sections do.

Retrieval is deliberately generous — it returns more candidates than the answer will use. A second pass then reorders them by how well they actually answer the question, and only the top few reach the model.

Passages that answer directly and early tend to survive this cut. Passages that build context first tend not to, because the reranker is judging the passage on its own, without the paragraphs that were going to lead up to the point.

The model can only consider so much text at once, and that budget covers the conversation so far plus everything retrieved. In practice only a handful of passages inform any single answer.

That scarcity is the real reason AI answers name two or three brands rather than ten. It is not an editorial choice — there is physically no room for more, and the passages that made it in decide the shortlist.

Assistants frequently turn one question into several searches of their own. Asking for the best option in a category can quietly become separate retrievals for pricing, reviews, alternatives and common complaints.

This is why a brand can rank well for the obvious question and still go unnamed: the engine went looking for something adjacent and found a competitor there. Covering the sub-questions around a topic matters as much as covering the headline one.

Every stage of the mechanism points the same way, which is unusual and useful.

• Be retrievable — if crawlers cannot reach a page it cannot enter the pipeline at all.
• Write self-contained passages — retrieval and reranking both judge the passage alone.
• Answer early — the reranker sees the opening, not where you were heading.
• Cover the adjacent questions — fan-out means the search that decides the answer may not be the one that was asked.
• Fix the sources you do not own — retrieval draws on the whole web, not just your site.