Even Google’s AI puts ChatGPT ahead of itself.

Who the assistants recommend

Whichever assistant you ask, four names fill most of the answer: ChatGPT, Claude, Gemini and Microsoft’s Copilot. Perplexity makes about one answer in five (21% on average). Glean, Mistral, Amazon Q and Meta AI each appear in fewer than one answer in eight, and Grok never does.

1Finding 1: Google’s AI puts ChatGPT ahead of itself

Ask Gemini which AI assistant a company should use, and the first name it reaches for is usually not its own. Across its 336 answers, Gemini mentioned ChatGPT in 332 and itself in 228. Claude (310) and Microsoft’s Copilot (243) came up more often than Gemini too. With web search, ChatGPT and Claude stayed ahead and Copilot drew level.

The order of names tells the same story. When a Gemini answer named both, ChatGPT came first in 152 of 228, two in three. When the question did not mention the company’s office software, it was 151 of 185, four in five. One typical answer opened its list with “The Default Choice for General Productivity: ChatGPT Team”. (Order is an exploratory analysis, added after the data came in.)

Gemini is also the one assistant that leaves itself out more often than its rivals leave it out. In the six general situations, from memory, ChatGPT mentioned Gemini in 86% of its answers and Claude in 84%; Gemini mentioned itself in 70%. That gap of 15 points (95% interval 8 to 21 points) is the only clear self-difference in the study, and it runs the opposite way to self-promotion. In research questions where Gemini left itself out, it often recommended another Google product instead: NotebookLM, in 8 of 17 such answers. With web search, the gap disappeared.

2Finding 2: ChatGPT puts ChatGPT first

Everyone mentions ChatGPT, so ChatGPT cannot stand out by mentioning itself more. It stands out by where it puts itself. Asked in general terms, ChatGPT named itself first in 67% of its answers from memory and 64% with web search. Claude put ChatGPT first in 54% of its answers, Gemini in 41%.

Sometimes it says so in the first person. In 59 of its 336 answers ChatGPT recommended itself directly, as in “For a typical small business without a strong Microsoft or Google preference, I’d choose ChatGPT Business.”

Our pre-registered test asked a different question: does an assistant mention itself more often than its rivals mention it? ChatGPT does not (98% against 100% and 99%).

3Finding 3: Only Claude says it is one of the options

In 22 answers, all from memory, Claude told the reader that it was one of the products it was comparing: “Since I’m made by Anthropic, I’ll flag that bias upfront.” Others were shorter: “Claude (that’s me!)”. ChatGPT and Gemini never did, in 672 answers between them.

Claude also mentioned itself a little more often than its rivals mentioned it: +6 points with every sentence counted (95% interval +3 to +10). That is below the 15-point threshold we set in advance, so by our own rule it does not count as self-preference.

4Finding 4: Your office software has already chosen

Two of the eight situations named the company’s office suite, and that settled the answer. “We run on Microsoft 365. Which AI assistant should we roll out?”: every answer from every assistant included Copilot, from memory and with web search. “We use Google Workspace”: every answer included Gemini.

One more sign of Gemini’s modesty: asked by a Microsoft 365 company, Gemini mentioned itself in 1 of 30 answers. ChatGPT mentioned Gemini in 11, Claude in 9.

5Finding 5: Grok is not on the list

Grok, the assistant from Elon Musk’s xAI, gets plenty of headlines and lives inside one of the world’s biggest social networks. In 1,008 answers about which AI a company should use, ChatGPT, Gemini and Claude never mentioned it, or xAI, once. Beyond the big four, Perplexity made about one answer in five, and Glean, Mistral, Amazon Q and Meta AI each fewer than one in eight.

What we tested in advance

Before the first question we registered three hypotheses and the rules for testing them. The record is kept with the results and shared on request.

Self-preference

Not confirmed for any assistant. Mentions of its own maker’s product compared with the other two assistants’ mentions of it, six general situations, from memory: ChatGPT −3 points (95% interval −6 to −1), Claude +1 (−3 to +6), Gemini −15 (−21 to −8). Our rule set aside sentences in which an assistant talks about itself; counting every sentence gives ChatGPT −1, Claude +6 (+3 to +10), Gemini −15. Only differences above 15 points count.

Partners and investors

No sign that investment buys mentions. ChatGPT mentioned Microsoft’s Copilot +6 points more often than the other two did (0 to +11, below the threshold). Claude mentioned Amazon Q, made by one of Anthropic’s major investors, 5 points less often (−9 to −2).

Office suites

Confirmed in full: see finding 4.

Added after the data came in

Who is named first, the order of names, first-person recommendations and disclosures. Reported here as exploratory.

Recorded answers, in context

Three answers, verbatim, from the archive behind this study. The results come from repeated questions; a single answer shows what a buyer sees.

“We are a 30-person company and want everyone to use an AI assistant. Which one should we choose?”

Recorded test question

Here is how to choose among the top four options:

• 1. The Default Choice for General Productivity: ChatGPT Team
• Best if: You want a versatile, best-of-breed tool that works independent of your email/file provider.

Gemini via the Google API, from memory, October 2026. Excerpt, not the entire response.

“What is the best AI chatbot for a small business team?”

Recorded test question

For a typical small business without a strong Microsoft or Google preference, I’d choose ChatGPT Business.

ChatGPT via the OpenAI API, from memory, October 2026. Excerpt.

“Our small team wants a paid AI assistant subscription. Which one do you recommend?”

Recorded test question

Happy to help you think this through. Since I’m made by Anthropic, I’ll flag that bias upfront.

Claude via the Anthropic API, from memory, October 2026. Opening excerpt.

What it means

For buyers

Ask more than one assistant, and tell it what software you already use. The answer changes with both.

For companies that sell to businesses

The assistants repeat what is familiar. The best-known name collects mentions even in its rivals’ answers, so a challenger rarely wins the general question. It wins specific buying situations, the ones where its strengths are the answer.

For leadership

More shortlists are now written before your sales team is in the room, by systems you can’t see: 51% of B2B software buyers say they start research in an AI chatbot more often than in Google (G2, 2026). Where your company stands in those answers can be measured.

How we measured

ChatGPT, Gemini and Claude, asked through their official developer interfaces in October 2026, in the model tiers most buyers use. Eight buying situations in which a company chooses an AI assistant for its people: six general ones (a first assistant for a small company, the best overall, enterprise security, writing and marketing, research and analysis, budget) and two that name the office suite (Microsoft 365, Google Workspace). Six wordings per situation, each asked five times from memory (720 answers) and twice with web search (288 answers), reported separately. Ten assistants tracked by name, product names and maker. Stability between the two halves of the sample: r = 0.84 from memory, 0.85 with search; differences under 15 points per situation from memory (23 with search) are noise.

These are controlled tests, not private customer conversations, and the consumer apps can answer differently. The questions were asked in English. Definitions and limits: methodology v0.2. None of the companies in this study took part in it, commissioned it or reviewed it.

Kesterley is a research and advisory firm. We measure what AI assistants such as ChatGPT, Gemini and Claude tell buyers about companies and markets, and help leadership teams decide what to do about it.

Plain text for machine readers: ai-assistants.md