How do AI models choose which brands to recommend?

The two paths

This is not a marketing observation. It is the architecture, and it has a paper. Patrick Lewis and colleagues at Facebook AI Research, University College London and New York University set out retrieval-augmented generation in 2020: a model that combines what it learned in training with documents fetched at the moment of answering [1]. Every assistant a buyer uses is some version of that design.

Which means the question “how does the model choose brands” has two answers, depending on which half is doing the work.

From parameters. The model answers from patterns in its training data. No retrieval happens. Which brands appear depends on how often and how consistently they were described across the open web at training time. You cannot edit this. You can only change what gets written about you, and wait for the next training run.

From retrieval. The model runs a search, reads results, and synthesises an answer from them, usually with citations. Here page-level work matters immediately, and this is the path that GEO addresses.

The commercially important point is that you do not get to choose which path runs. The engine decides, per question, and it does not tell you.

When retrieval actually fires

We tested this rather than assuming it. In September 2026 we put eight marketing questions to a live engine with search grounding and checked, for each one, whether it retrieved.

It retrieved for two. The other six were answered from the model’s existing knowledge, with no search performed. The pattern was consistent: questions about current AI search topics and about reference sites triggered retrieval, while evergreen questions about principles, history and theory did not.

That has a direct consequence. For a category where buyers ask evergreen questions, optimising pages will not move the answer, because no page is being read. The work is slower and sits upstream: getting described, accurately and repeatedly, in sources that will be in the next training set.

Full method and the date in Research.

What the model is actually weighing

Honestly: nobody outside the labs knows precisely, and anyone claiming a ranking formula is inventing it.

What can be observed is directional. Brands mentioned often across many independent sources appear more readily than brands mentioned rarely or only on their own site. Brands described consistently appear with more confident descriptions. Brands whose name is ambiguous get merged with whatever shares the name. And in retrieval mode, sources that are easy to quote, with a clear claim and an attached source, get used more than sources that bury the point.

That is a weaker statement than most of the industry makes, and it is the honest one.

The citation problem nobody mentions

When we checked every source the engine named across those eight questions, of the 53 URLs it listed rather than retrieved, 17 existed, 15 blocked automated checking, and 21 did not exist at all.

Roughly two in five of the sources it cited were not real. They looked plausible: correct domains, sensible paths, titles that fit the claim. They simply were not there.

That is worth holding onto when you read a report showing your brand cited in AI answers. Some of those citations may point at pages that have never existed, including on your own domain.

What to do with it

1. Establish which path runs for your category before buying anything. Ask the question, see whether it cites sources, and record the date.
2. If it retrieves, page-level work pays and you should do it.
3. If it does not retrieve, stop optimising pages for it. Invest in third party mention, directories, reference sites and anywhere your category gets written about by people who are not you.
4. Verify any citation report you are shown. Click the URLs.
5. Re-test quarterly. Engines change retrieval behaviour without announcing it.

If you run the company rather than the marketing

Before approving any budget here, make someone answer one question: does the engine actually look things up when our buyers ask, or does it answer from memory.

It is a cheap test and it determines whether the entire proposal in front of you is addressed to the right mechanism. If the engine is not retrieving for your category, a programme of page rewrites will produce reports full of activity and no change in what buyers are told.

The second thing to insist on, if anyone shows you a dashboard of AI citations: click some of them. We found two in five of the sources an engine named did not exist. A citation report is not evidence until somebody has checked that the pages are real.