How does AI choose which brands to recommend?

When an assistant names three brands in an answer, it hasn't consulted a ranking of brands. It has done something more ordinary: combined what it already knew with what it just read. Knowing which of those two produced a given answer tells you what, if anything, you can do about it.

Source one: what the model learned

Every model is trained on a large snapshot of text, and from it learns associations: this category, these names. A brand that has been written about widely for years, in news, forums, reviews and documentation, is a strong association. A brand that launched last year barely registers.

Answers from memory have recognisable traits. They cite nothing. They lean towards long-established names. They can be out of date: an old price, a product you renamed, a feature you dropped. In our own sample of eight ChatGPT answers, the one question answered with no citations at all, "AEO vs SEO", was a definition, the kind of question a model is confident answering from memory (the sample).

You can't edit the training data. You can influence the next snapshot, slowly, by being written about in many places, but that's a years-long effect, not a campaign.

Source two: what it reads while answering

For questions that need current information, which includes nearly every buying question, the major assistants search first. This is often called retrieval, or grounding.

• ChatGPT runs its own searches, reads pages, and cites them; our guide to how ChatGPT search picks sources covers the details.
• Google's AI Overviews and AI Mode are built on Google Search. Google says both "may use a 'query fan-out' technique", running several related searches to build one response.
• Perplexity searches on nearly every question and shows numbered sources beside the answer (how Perplexity picks citations).

The searches an assistant runs are its fan-out searches. The pages they return are what it reads, and the brands on those pages are the brands it's likely to name. This is why the answer to "which tool should I use?" so often mirrors the top roundups for that category.

What makes a page get used

Retrieval has two gates. First, the page has to be found: it must be crawlable by the assistant's search crawler and rank for the searches the assistant runs. Second, once read, it has to be useful for writing the answer. Pages that state the answer plainly, near the top, with specifics, are easier to use than pages that bury it.

There's some research on the second gate. A 2023 paper that named the field "generative engine optimization" tested rewriting methods on a benchmark of queries and reported that they "can boost visibility by up to 40% in generative engine responses", with different methods working in different subject areas. Treat that as a direction, not a promise: it was a research benchmark, not today's ChatGPT.

Why the same question gets different brands

Both sources add randomness. The model samples its words, so two answers to the same question differ. The searches it writes differ between asks, so the pages it reads differ too. The top two or three names in a category tend to be stable; the names below them rotate. That's why the useful measure is how often you're named across many answers, not whether you appear in one.

How the two sources combine

A rough guide to which source dominates:

The last row matters more than people expect. When an assistant looks a brand up by name, it often reads that brand's own site and help docs. For Prepco, our made-up exam-prep brand, that means its pricing page and its course pages are, in effect, what assistants say about it. Make sure they're current and say the important things in plain words. For software that means docs and pricing (AEO for B2B SaaS); for shops, product pages and feeds (AEO for ecommerce).

So what can you change?

Mostly the second source, and fairly quickly. Make sure the assistants' crawlers can read you, put direct answers on your own pages, and get onto the third-party pages they cite for your category. Our guide to why ChatGPT recommends your competitor turns this into a diagnosis you can run on one question at a time.