How AI decides which brands to recommend

Last updated: August 15, 2026

ChatGPT does not recommend brands from memory. When a shopper asks it what to buy, it searches the web, opens a handful of pages, reads them, and writes its answer from what those pages say. So the brands it names are, overwhelmingly, the brands those pages name.

What actually happens when a shopper asks

Imagine someone types “what’s the best niacinamide serum for oily skin?” into ChatGPT. Behind the answer, roughly this happens:

• It runs a web search, much like you would.
• It opens a small number of results — often a magazine roundup, a Reddit thread, a retailer page, a brand’s own product page.
• It reads them and writes a summary answer.
• It lists those pages underneath as its sources, which is why you see little link chips below an AI answer.

This is the whole mechanism, and almost everything else follows from it. It also explains the thing founders find most confusing: you can have a great product, a fast site and real customers, and still never be named — because none of that is what the assistant read.

Why it isn’t about the model “knowing” you

There are two different ways an AI can produce a brand name. It can recall something from its training data, or it can read a page right now and repeat what that page says. For shopping questions, the second dominates — which is good news, because training data is frozen and unreachable, while the live web is neither.

It also means the answer moves. A roundup article that adds your product this month can change what shoppers are told next month. That is the lever this whole category is built on.

Which pages AI tends to read

Studies of AI citations keep landing in the same place: independent, discussion-heavy and comparison-shaped pages punch far above brand-owned pages. Ahrefs’ analysis of the most-cited domains in ChatGPT found Reddit alone taking about 16.7% of citations, with Wikipedia and Reddit together accounting for more than a quarter of US ChatGPT citations.

The pattern for products specifically is roundups and listicles — the “10 best X” article — plus review platforms and community threads. Which is to say: the pages deciding whether you get recommended are mostly pages you do not own.

What that means practically

• Your own site still matters, because AI does read brand pages and does cite them. It is just rarely the only thing it reads.
• Being in other people’s articles matters more than it used to. A single frequently-cited roundup can carry more weight than a page of your own.
• Community discussion counts. Reddit threads are read and cited heavily in this category.

What if AI names my brand and a competitor together?

Very common — a comparison article, or an answer that recommends three products including yours. Brands are not competing for one slot per answer: both get named, and both count as a mention. That is also why mention rates across a set of brands don’t add up to 100%. What separates you is how often you appear across many answers, not whether you edged someone out of one.

So what do you actually do?

Work backwards from the pages. If you know which pages an assistant reads before recommending someone in your category, you know exactly where to go. That list is what a citation gap report is — the pages AI cites when it recommends a competitor but has never cited for you.

The honest caveat: we can see which pages an assistant cited and who it cited them for. We do not read the page’s own text, so we never claim your brand is missing from a page — only that AI has not yet used it in your favour.

If you want to check where you stand before any of this, here is how to test it by hand.