How ChatGPT, Gemini and Perplexity choose the brands they recommend

Two places an answer comes from

An assistant's answer draws on two kinds of knowledge. The first is what the model learned during training, from a very large snapshot of text. It knows well-established brands this way, but its knowledge stops at a cut-off date and it cannot tell you where a given belief came from.

The second is live retrieval. For many questions, especially about products, prices and recent events, assistants such as ChatGPT search, Perplexity, Gemini and Google AI Overviews run a web search, read a set of pages and write the answer from them, often showing those pages as sources.

When the assistant searches, the sources decide

When retrieval is involved, the brands in the answer are largely the brands present in the pages retrieved. That makes the question very concrete: for the query your buyer types, which pages does the assistant read, and are you in them?

Those pages are rarely only vendor sites. Comparison articles, review platforms, community discussions, directories and specialist media appear often, because they discuss several options at once, which is exactly what a recommendation question needs.

Signals that plausibly help

The exact weighting is not public, so treat this as informed practice rather than a formula. Brands that assistants name consistently tend to have:

• A clear, consistent description of what they do, repeated across their own site and third-party sites.
• Specific pages that match the question: a use case, an industry, a comparison, a pricing explanation.
• Mentions on independent sources in their category, not only on their own domain.
• Pages that are accessible to crawlers and fast to parse, with the key facts in text rather than images.

Why the answer changes from one run to the next

The same question asked twice can produce different names. Models generate text with some randomness, retrieval results shift, and the assistant may phrase or interpret the question differently each time. Answers also differ by country, language and whether the user is signed in.

This is why a single screenshot proves little. The useful measure is how often you are named across repeated runs of the same question, on each assistant separately.