How AI Assistants Select Brands to Cite

An AI assistant answering a commercial question names a handful of brands, without ranking positions and without any account of what it discarded. This report sets out the four stages at which that set is decided and what each one rewards. It describes mechanisms, not measured outcomes.

The selection path

A citation is the end of a path with four stages: retrieval, candidate assembly, verification, and rendering. A brand can be eliminated at any of them. Visibility in conventional search establishes eligibility for the first stage and nothing beyond it.

Retrieval. The assistant reformulates the user’s question into one or more queries and fetches documents, either from a live index or from an internal corpus. Reformulation matters more than it does in search. A user who asks a long, conditional question has that question decomposed into narrower ones. Documents that answer the narrow sub-question are retrieved; documents that answer the original question broadly may not be.

Candidate assembly. The retrieved documents are read for entities that satisfy the constraints in the question. A brand mentioned in a retrieved document but not described in it — a bare name in a list, a logo, a link with no surrounding text — supplies nothing for the model to assemble. The unit that survives this stage is a claim about a named entity, not the name itself.

Verification. Where the assistant can check a claim against a second source, it prefers claims that check out. Corroboration across independent documents is the practical form this takes. A specification stated only on the vendor’s own site is weaker than the same specification repeated in a directory, a review, and a technical write-up.

Rendering. The assistant composes an answer of finite length. A brand that needs a sentence of qualification is dropped for one that can be placed in a clause. Category precision is what makes a brand cheap to render.

What each stage rewards

The four stages reward different properties, and none of them is in tension with another.

Retrieval rewards documents written at the granularity of a question. A page that covers one narrow topic completely is retrieved for that topic; a page that covers a whole domain shallowly competes with everything.

Candidate assembly rewards explicit attribution. Statements of the form “X does Y for Z” survive extraction. Marketing copy that describes a benefit without naming the entity performing it does not, because the entity and the claim become separable in the retrieved fragment.

Verification rewards independent repetition. This is a property of a brand’s whole footprint rather than of any one page, and it is built by third parties describing the brand in their own words, in sources the model is likely to hold.

Rendering rewards category clarity. A brand that occupies a nameable category is included in answers about that category. A brand whose category must first be explained is included less often, because the explanation costs words the assistant will not spend.

Two failure modes

The undescribed brand. The brand is well known, appears widely, and is rarely cited. Its name occurs in many documents but with no accompanying claims. Assistants have nothing to say about it, so they say nothing.

The self-described brand. All substantive statements about the brand originate from the brand. Retrieval succeeds and assembly succeeds, but verification finds no second source. Assistants hedge, attribute the claim explicitly to the company, or omit it.

Neither failure mode is visible in conventional metrics. Traffic, rankings, and impressions can hold steady while citation rates stay low, because the two systems reward different things.

Volatility and its limits

Citation sets are less stable than search rankings. The same question asked twice can return different brands, because sampling, session context, and retrieval variance all bear on the outcome. One consequence governs any analysis of this: a single observed answer is not evidence of position. Claims about assistant visibility require repeated observation across phrasings, and a figure drawn from a handful of prompts is anecdote.

Implications for analysis

Assessing how a brand stands with assistants means assessing four separate things: whether documents about it are retrievable at question granularity; whether those documents carry extractable claims tied to its name; whether independent sources repeat those claims; and whether its category can be stated in a few words. These are properties of a brand’s public record rather than of any single website, and they change slowly.

Trellner reports keep the four as distinct headings rather than collapsing them into a score. A brand can be strong on retrieval and weak on verification, and the remedy for each is different.