How AI decides what to recommend
AI assistants surface a brand in one of two ways: parametrically, recalling it from training data, or through retrieval, searching the live web during the answer. Retrieval can be influenced within weeks through better pages and citations. Parametric recall moves only over training cycles, and needs broad, consistent third-party coverage.
Updated 19 July 2026
Two mechanisms, two strategies
AI assistants answer buyer questions in two ways, and the difference decides what you should actually do.
Sometimes they answer from what the model absorbed during training — its parametric memory. Sometimes they browse the web live and answer from what they find — retrieval. Most real answers are a mixture, but one usually dominates for a given brand and category.
Knowing which one is surfacing you is the difference between a plan that works this quarter and one that cannot.
If you surface through retrieval
You can move quickly. The assistant is reading the live web, so better pages and more citations from sources it trusts change your standing within weeks.
The work: publish the definitive answer pages for your category — buying guides, comparisons, clear FAQs — and earn mentions on the review sites, roundups and trade sources that assistants lean on when they search.
You can tell you are in this position when answers about you carry citations, and when the cited domains are ones you could plausibly get named on.
If you surface through trained memory
The model already knows you, and you cannot edit its memory this quarter. No amount of publishing changes what has already been trained.
The durable play is to become widely and consistently cited across the authoritative sources in your space. That wins retrieval immediately and, over training cycles, works its way into the model's memory too. It is the slower path and the more defensible one, because a reputation distributed across thousands of sources is not something a competitor can outspend quickly.
You can tell you are in this position when assistants describe you confidently with no citations attached — the knowledge is coming from inside the model.
Which is more common
In our own measurements, most brand mentions arrive parametrically, with no citation at all. This is the single most under-appreciated fact in the field, because it means a strategy consisting entirely of on-site optimisation is working on the smaller half of the problem.
It also explains a pattern that otherwise looks unfair: a brand with an excellent website losing to a competitor with a poor one. The competitor is better known to the model, and the website was never the thing being consulted.
What both paths share
Whichever mechanism dominates, the fundamentals are the same:
- Be a clearly-defined entity — consistent name, description, location and profiles, so a model can resolve who you are without ambiguity.
- Publish content that genuinely answers what buyers ask, rather than describing your company.
- Get named on the sources AI trusts, which is the lever that serves both mechanisms at once.
- Be readable — if a crawler cannot parse your site, nothing else matters.
The point of measuring is to tell you which lever is yours to pull first. You can scan your site free for the readable half, and the index shows how the brands we measure compare on both.