How an assistant decides which brand to name

When an assistant recommends three brands in a category, the recommendation came from one of two places, and the difference decides what you can do about it. Both are a departure from how search rewarded the page that ranked.

The first is retrieval. Google describes its own AI features as retrieval augmented generation, “a technique used to improve quality, accuracy, and freshness of AI responses by relying on core Search ranking systems to retrieve relevant, up-to-date web pages”, combined with query fan-out, where one question becomes many searches. On that path your page is fetched, read and possibly cited, and the levers are the familiar ones: be crawlable, be findable, be quotable. Google’s guide to optimising for generative AI features is explicit that there are no AI-specific ranking requirements and no special markup to add.

The second is the model’s own memory of your brand, and it is the one most measurement misses.

The model may not have read anything

A Berkeley team at EMNLP 2024 took 1,147 manufacturer pages across 50 product categories and asked which factors decided how models ranked the products. Their finding: models “vary significantly in prioritizing product name, document content, and context position”, and for some, the page barely registered. “GPT-4 Turbo and Llama 3 are heavily influenced by their latent knowledge of product names”, they write, and “GPT-4 Turbo is also minimally influenced by retrieved documents” (Pfrommer et al., arXiv:2406.03589).

A 2026 comparison of assistants and web search found the same thing from the other direction. Across ranking-style consumer queries, 16% of the entities that appeared in a generated ranking “did not occur in any retrieved snippet” (Chen et al., arXiv:2601.16858). The model named brands it had not just looked up. You cannot earn a citation into a list that came from the model’s priors; you can only become one of the brands it holds priors about, which is slower work and mostly happens off your own domain.

Why this splits your to-do list

If your absence is a retrieval problem, it is addressable this quarter: the page exists, it is just not being pulled into answers, and there is evidence about what changes that.

If your absence is a memory problem, no amount of on-page work fixes it directly. What moves it is being described, consistently and in many places, by sources the model ingested: retailers, reviews, comparisons, reference sites, forums.

The two also fail differently. Retrieval failures are stable and reproducible. Memory failures are erratic: the brand appears for one phrasing and vanishes for a near-identical one, which is why a single spot check tells you almost nothing about either. The same team that studied non-determinism in supposedly deterministic settings found accuracy varying by up to 15% across ten identical runs (Atil et al., arXiv:2408.04667), so how you sample matters more than what you sample.

The practical starting point is to separate the two in your own data. When your brand appears, was a page of yours cited? When it does not appear, was a competitor cited, or did the model simply name names? Those two columns turn one vague number into two different pieces of work.