Why ChatGPT names some brands and not others

A model has no opinion about your product. When someone asks it for a recommendation, it does two things: it recalls what it read during training, and, when it can browse, it fetches a few pages and reads them again. The names it gives are the names that appeared clearly, consistently and recently in what it read. That is the entire mechanism, and it explains most of what founders find confusing.

Recall: what it already believes

Training data is a snapshot. If your brand was mentioned often, in the right context, on pages the model was trained on, it can name you from memory. This is why old, well-known brands show up even when their product has fallen behind. It is also why a brand launched last year often does not show up at all, whatever its quality.

You cannot change the snapshot. You can change what the model reads when it browses, and you can change the next snapshot.

Browsing: what it reads right now

When browsing is on, the model runs a search, opens a handful of results, and answers from them. Which pages? Mostly the ones that rank for the question, and the ones on sites the model has learned to trust for that kind of question: review sites, comparison articles, community threads, documentation.

This is the door that is open to you. If the page that ranks for "best X for small teams" names you, the model names you. If it does not, the model does not, however good you are.

The phrase test

Open a stored answer that names a rival and look at the sentence. Nine times out of ten it is a light paraphrase of a sentence on a page somewhere: the rival's comparison page, a review site, a Reddit reply. The model is quoting. That sentence is your brief, because the page that beats it is the page that gets quoted next.

What the model trusts

A few patterns come up in every set of answers we have looked at.

• Pages that answer one question directly, with the answer in the first paragraph, get quoted more than pages that build up to it.
• Comparison and "best of" pages on sites that are not the vendor get quoted more than vendor pages, because the model has learned they are less biased.
• Community threads where several people agree get quoted for "what do people actually use" questions.
• Structured facts, like a price in a table or in schema, get quoted as facts. Prose about pricing gets skipped.
• Recent pages win ties. A 2026 comparison beats a 2023 one.

What that means for you

Three jobs, and they are the same three jobs the search engines wanted all along, done more carefully.

1. Write the page that answers each question your buyers ask, directly, with the answer at the top and the facts in a table.
2. Get onto the review and comparison pages the engines already cite for your questions. Not any pages: those pages.
3. Make sure a model can read your site at all: no blocked crawlers, a sitemap, schema on the pages that carry facts.

The uncomfortable part

None of this is a trick, and none of it is fast on its own. It is a page a day and a mention a week, for months. That is why we built a crew to do it rather than a dashboard to watch it. The sources feature lists the pages the engines cite for your questions; the content engine writes the pages that answer them.