How ChatGPT, Gemini and Google choose which brands to name

AI assistants name a brand for two reasons: the model already links that brand to the question, or its search found pages that name the brand. No assistant publishes how it weighs the two. The best evidence so far suggests that memory carries famous brands, and that lesser-known B2B vendors depend more on third-party pages.

That evidence is mainly one University of Toronto preprint from October 2026, which tested ChatGPT, Gemini and Perplexity through their APIs on 143,249 answers. It found that 46% of ChatGPT’s mentions of brands the question didn’t name had no cited page naming them. The study is correlational, isn’t peer reviewed and used small API models, so treat it as a strong lead. Nobody can guarantee that an assistant will name you.

Where do the brand names in an AI answer come from?

The brand names in an AI answer come from the model’s memory, from pages its search returned, or from both. Memory is what the model absorbed in training. Search is the set of pages it pulled for this question. A citation can only come from search, but a name can come from either, so an assistant can name a brand and link to nothing.

Our guide to what AEO is covers how a page gets cited. Being named is a separate outcome, and it’s the one this article is about.

How much does a model’s memory decide?

A model’s memory seems to decide more for well-known brands in consumer categories than for niche B2B vendors. In the Toronto preprint’s consumer data (651 questions), 47% of ChatGPT’s citations went to websites the model named without searching. In its B2B data (55 questions about proposal software for architecture, engineering and construction firms), the figure was 32%.

Memory also appears to steer search. In the same preprint, ChatGPT wrote a brand into its own search queries in 67% of cases when the model already linked that brand to the question, and in 11% of cases otherwise. So a well-known brand can get a search built around it, while a rival never gets looked up.

Being known isn’t the same as being suggested. A 2026 preprint based on thesis research tested 112 startups from Product Hunt’s 2025 top 500, using 2,240 queries. ChatGPT (gpt-4o-mini) recognised 99.4% of them when asked by name, and surfaced 3.32% when asked a discovery-style question. Perplexity (Sonar) went from 94.3% to 8.29%. These were startups, so the gap may be smaller for larger brands.

Our view: a larger brand usually has a different problem from a startup. The model probably knows the name, so the first questions are which categories it links you to and whether it describes you correctly.

Why do other sites matter so much for being named?

Other sites seem to matter most for brands the model doesn’t already know. In the Toronto preprint’s small B2B dataset, a cited third-party page naming a vendor raised the odds that the answer named that vendor 39.7 times on ChatGPT and 17.7 times on Gemini. In the consumer data, the same figures were 4.2 and 6.3.

Read those figures with care. The B2B set has only 55 questions, the paper doesn’t say what its 39.7 figure was adjusted for, and a brand that was already likely to be named is also more likely to appear on the pages cited. The models also differed: for those B2B questions, the model behind ChatGPT named the tracked vendors for 2% to 16% of questions, and the model behind Gemini for 60% to 96%.

Your own pages still have a job. When ChatGPT had consulted several pages at equal rank in the same preprint, brand pages had 1.71 times the odds of being cited. A citation doesn’t put your name in the answer, though, as our AEO guide explains.

The type of question matters too. Semrush, which sells an AI visibility tool, looked at 3,981 domain appearances across 115 prompts in June 2026. Brands were named in 43.3% of appearances on comparative prompts and 18% on informational ones. It’s a modest sample, but it suggests a definition page and a “which tool should I buy” page will behave differently.

Does publishing your own “best” list get you named?

The evidence on self-published “best” lists conflicts, so treat the question as unsettled. These lists are clearly sourced often. Ahrefs, which sells SEO tools, found “best X” lists made up 43.8% of source links on 750 top-of-funnel ChatGPT prompts in December 2025. That shows what gets sourced, not that publishing a list helps you.

Two later studies point in opposite directions:

• Against. Lily Ray, an SEO consultant, checked 100 B2B “best software” queries on three dates from April to June 2026, using one tool. Of 323 self-promotional lists cited in Google’s AI Overviews, 224 (69%) came from brands the answer didn’t recommend. Recommended brands tended to be established, with more referring domains.
• For. Maeva Cifuentes used Scrunch data from five B2B SaaS clients, covering 93,968 answers from Q2 and Q3 2026 on ChatGPT, Copilot, Perplexity and AI Mode. Vendors were named in 45.4% of answers citing their own list, and in 16.3% on the same questions without it. She notes the link may run backwards: an assistant already planning to name a vendor may cite its page.

The Toronto preprint adds a third result. Lists were retrieved often, but at equal rank ChatGPT had 0.54 times the odds of citing one.

Our view: the studies cover different engines and samples, so they can both be right. A list on your own site doesn’t replace evidence from other sites, and a list that ranks you first without a stated method will be hard to defend to a reader. If you publish one, make it useful without your product in it.

How do ChatGPT, Gemini and Google differ?

ChatGPT, Gemini and Google’s AI features differ mainly in where they search and in how readily they turn a source into a name. In Semrush’s June 2026 study, ChatGPT tended to cite a domain without naming the brand, and Gemini tended to name a brand without citing it. The table below combines what each company or court has said with what studies have measured.

These numbers vary with how they’re collected. Ahrefs found 64% of Gemini answers cited nothing in its tracked questions in October 2026, while the Toronto preprint’s API runs of a Gemini model averaged about 10 pages per answer. Apps and APIs may behave differently, so don’t carry a figure from one to the other.

The engines also share few sources. Wellows, which sells AI visibility tools, found 79.6% of website-and-question pairs were cited by only one of five engines, from January to June 2026. It collected one run per engine, which may inflate the difference.

Our view: don’t write a version of each page for each engine. Act on the differences through access (crawlers, Bing and Google indexing) and through where evidence about you lives.

Why does the same question name different brands each time?

The same question names different brands because an AI answer behaves like a sample and not like a fixed ranking. In the Toronto preprint, two same-day ChatGPT runs shared 2.7% of their search queries and 15% of their cited pages. So “which brands does ChatGPT choose?” only has an answer as a share across many runs.

Names are steadier than links. Ahrefs, which sells a tracking product, asked five tools the same questions daily for 20 days. About 50% to 70% of cited websites appeared on only one day, against about one in five brand names. SparkToro found the same list almost never repeats, yet in a follow-up on headphones the leading brands appeared in 55% to 77% of 994 responses. That work wasn’t peer reviewed, and one co-author works at a tracking company.

What can’t a brand control here?

A brand can’t control what an assistant says, and no one can promise a mention, including us. It can’t retrain a model either. Our view is that memory is the slowest part to shift, because it changes on the model maker’s schedule. A 2026 survey of 45 studies, a preprint by a single author, found no technique with a stable, long-term effect on organic discovery across engines.

We also know of no test showing that earning a mention causes later answers to name you. The link between third-party pages and naming is a correlation. And if a rival really does suit the question better, no amount of coverage should change that for long.

Shortcuts carry their own cost. Microsoft’s security team reported 50 hidden “remember this company” prompts from 31 real companies over 60 days in February 2026, and it treats them as an attack. Google says that creating pages mainly to manipulate its AI responses breaks its spam policy.

How can you find out why an assistant names a rival?

You can find out why an assistant names a rival by reading the answers and their sources by hand. The sorting below is our own way of doing it, not a research method, and one answer proves little.

1. List 20 to 30 questions buyers ask. Start with the AI question generator, then correct them from sales calls. Include comparison questions.
2. Ask each question several times on different days, in new chats, in each assistant. SparkToro suggests 60 to 100 runs per prompt, which isn’t realistic by hand, so treat your results as a rough guide.
3. Record who’s named, how they’re described and which pages are cited. The AI answer analyser can pull this from a pasted answer.
4. Sort each mention of a rival. If no cited page names it, memory is the likely source. If a third-party page names it, that’s a page you can read and possibly earn a place on. If its own page is cited, compare that page with yours.
5. Check your own description with the AI answer fact-check, and use the query fan-out tool to see which narrower searches you might be missing.
6. Pick one gap, fix it honestly, and ask the same questions again alongside a few you didn’t work on.

Doing this by hand gets slow once you track many questions. Quotable, which we make, runs your questions on each engine on a schedule and tracks who’s named and cited. Ahrefs Brand Radar, Semrush’s AI Visibility Toolkit, Scrunch and Gumshoe also measure AI answers. None of these tools, ours included, can see inside a model, so none can tell you for certain why a brand was named.

How would you brief this to a leadership team?

A leadership team needs to hear three things: what’s known, what isn’t and what you’ll report. You could adapt this note:

“AI assistants name brands from what their models already know and from the pages their searches find. For a brand of our size, we should check which categories models link us to, how they describe us and which third-party pages rivals appear on. The evidence is mostly correlational and changes with each model, so nobody can promise a result. We’ll report how often we’re named per engine, as a count across many answers, next to questions we aren’t working on. We don’t yet have evidence linking these numbers to revenue, so we’ll read them alongside branded search and sign-ups.”