How ChatGPT, Gemini, Claude and Perplexity Decide Which Brands to Mention

AI engines decide which brands to mention by combining what the model already learned during training with what it finds in a live web search, then writing an answer from that material. Outside studies suggest the brands that show up most clearly and consistently in it are the ones named most often, but no engine publishes a ranking formula for brand mentions. What the vendors do publish is how their models are trained, which crawlers feed their search features and what blocks them, and that is enough to explain most of what brand owners see.

I will take the two routes one at a time, then look at why answers change between runs, and finish with what is documented, what is studied and what is still unknown.

What are the two routes a brand takes into an AI answer?

Every brand name in an AI answer arrives by one of two routes, often both at once.

• Route one: training. A large language model is trained on a huge collection of text. Whatever it read about your brand before its knowledge cutoff is baked into the model, without links or dates attached.
• Route two: retrieval. When you ask a question, the engine may run a web search, read a few pages and write the answer from them. This is called retrieval-augmented generation, and it is how AI search features show sources.

Route one: what does the model already know?

Training is the slow route. Anthropic’s documentation explains the idea well: a model has a reliable knowledge cutoff, the date through which its knowledge is most complete, and a broader training data cutoff. Anything after that is not in the model at all.

The gap is real even for the newest models. As of October 2026, Anthropic lists June 2026 as the training data cutoff for its current Claude models, and OpenAI lists knowledge cutoffs in April and May 2026 for its current API models. Google’s Gemini 2.5 Pro shipped with a knowledge cutoff of January 2025 and a June 2025 update date. Labs release new models every few months, but each one stops learning at a fixed date.

For brands this has two consequences. A brand that launched or rebranded after a model’s cutoff cannot be in that model’s memory, so it depends entirely on retrieval. And a brand that was described thinly or inconsistently in the years before the cutoff will be remembered thinly or inconsistently. You cannot edit a trained model. You can only change what the next one reads. Our guide on how a new brand gets into AI answers covers the launch case.

Route two: what does the engine look up live?

Retrieval is the fast route, and every major engine now does it. Each one documents its crawlers and sources a little differently.

ChatGPT. OpenAI says ChatGPT search draws on third-party search providers plus content provided directly by news partners. Microsoft announced in 2023 that Bing would be the default search experience in ChatGPT. OpenAI’s own crawler for search is OAI-SearchBot, and its documentation is direct: sites that opt out of OAI-SearchBot will not be shown in ChatGPT search answers. GPTBot, by contrast, is the training crawler. OpenAI’s publisher FAQ says any public website can appear in ChatGPT search as long as it does not block OAI-SearchBot. Our guide on where ChatGPT gets its information goes into each crawler.

Perplexity. Perplexity is built around retrieval and shows its sources alongside its answers. Its docs say PerplexityBot surfaces and links websites in Perplexity’s results and is not used to train foundation models.

Claude. Anthropic made web search available on all Claude plans worldwide in May 2025. Its help center says blocking Claude-SearchBot or Claude-User can stop Claude from indexing or retrieving a site’s content for user queries, while ClaudeBot is the training crawler.

Gemini, AI Overviews and AI Mode. Google’s AI features in Search sit on top of Google’s normal index. Google states there are no additional requirements or special optimizations to appear in AI Overviews or AI Mode; a page has to be indexed and eligible for a snippet. Google also says these features may use “query fan-out”, running several related searches across subtopics to build one answer. Separately, Google-Extended (a robots.txt control token, not a separate crawler) governs training and grounding for Gemini apps, but Google says it does not affect inclusion in Google Search.

Copilot. Microsoft says Copilot has about 150 million monthly active users across work and consumer versions. Microsoft’s public material in our sources does not detail how Copilot picks pages, so we will not guess; since Bing is Microsoft’s search engine, being indexed in Bing is a sensible baseline.

The common thread: retrieval can only use pages that are public, crawlable by the right bot, indexed and relevant to the query. A mention on a page no engine has indexed does not exist for this route.

What happens between the search and the answer?

This is where documented facts thin out, so I will mark what follows as informed reasoning plus published studies, not vendor documentation.

Once an engine has a set of pages, the model reads them and writes an answer. If a question is “best payroll apps for small restaurants” and four of the six retrieved pages name the same three apps, those three are likely to appear. A brand named on one page, in passing, is less likely to make it. A brand described on several pages with the right topic context (restaurants, tip pooling, small teams) has more chances to be picked.

Studies support parts of this picture. Because of query fan-out, Google’s AI answers often cite pages that do not rank for the query itself: by early 2026, only 37.9 percent of AI Overview citations came from the top 10 results. BrightEdge found that commercial wording such as “where to buy” produced 4 to 8 times more brand mentions than informational questions, and that 44 percent of prompts it tested named no brand at all. How a person phrases the question changes which brands get named at all.

Why does the same question get different brands each time?

Ask the same question twice and you will often get a different list. This is normal, and it is measurable.

In January 2026, SparkToro and Gumshoe had 600 volunteers run 12 prompts through ChatGPT, Claude and Google’s AI 2,961 times. ChatGPT and Google’s AI had less than a 1 in 100 chance of giving the same brand list in any two runs, and the same order roughly once in 1,000. Even category leaders in a follow-up headphone test appeared in only 55 to 77 percent of answers.

Google’s features move too. Ahrefs found an AI Overview has a 70 percent chance of changing between observations, though the core meaning stays similar. And AI Mode and AI Overviews, for the same query, shared only 13.7 percent of cited URLs.

The practical lesson: a single screenshot proves nothing, good or bad. Visibility is a rate. Our guide on measuring AI visibility shows how to build a prompt set and track your mention rate over repeated runs.

Why does consensus across many sources matter?

If answers are partly random, what makes a brand show up more often? Here the evidence is consistent across studies, even if none proves cause.

Ahrefs found branded web mentions correlated with visibility in ChatGPT (0.664), AI Mode (0.709) and AI Overviews (0.656). The same research found visibility moves together: brand outputs correlated 0.749 between AI Overviews and ChatGPT, and 0.769 between AI Mode and ChatGPT. Brands that are visible on one engine tend to be visible on the others.

Our reading is simple. Each run of an AI answer samples from what the engine knows and what it retrieves. The more independent, indexed pages describe your brand in the same context, the more of those samples contain you. One strong article helps on the days it is retrieved. Many consistent mentions help across training, retrieval and every engine at once. That is the case for building brand mentions rather than chasing a single placement.

What is documented, what is studied and what is unknown?

Anyone who claims to know the last row in detail, or promises a ranking in ChatGPT by a date, is guessing or selling.

What does this mean for your brand?

Put the two routes together and a few practical rules follow. To be in the next model’s memory, your brand needs to be described clearly on many pages before that model’s cutoff. To be in the answers engines retrieve now, those pages need to be public, indexed and crawlable by the search bots, and your own site should be too. To be picked from the retrieved pages, the words around your name need to match the questions buyers ask: your category, your use case, your market.

None of this needs a trick, and none of it can be guaranteed. It is the same work for every engine, which is convenient, because the evidence says visibility on one tends to travel to the others. If your mention rate is near zero today, start with the diagnosis in our guide on why AI does not mention your brand.

What to do next

• Check your robots.txt for OAI-SearchBot, PerplexityBot and Claude-SearchBot, so retrieval can reach your site.
• Run ten buyer questions three times each across engines and record your mention rate.
• Get a free AI visibility check to see where you stand today across ChatGPT, Gemini, Perplexity and AI Overviews.
• If AI already chooses you some of the time, growing how consistently and how high ChatGPT names you is the step The Blue Ocean GPT works on.