“We’re B2B. Nobody’s finding us through a chatbot yet.”
Say that in a budget meeting and the AI-visibility line item quietly dies, right there, with everyone nodding.
The honest problem is that you don’t have anything to say back. I didn’t either. I brought AI-referral numbers to a review once and got “our customers are companies” and folded.
Then someone went and asked the people who sign the contracts.
Where Do B2B Buyers Actually Start Looking For A Vendor?
22% of them start inside a chatbot like ChatGPT. 16% start in a search engine. The largest single answer is still a recommendation from a peer, at 49% — but search and AI have already traded places.
The research comes from Profound, which sells AI-visibility measurement, and Listen Labs, which runs the interviews. They published it on 15 July 2026.
The sample is 100 CMOs, interviewed one at a time, across technology, financial services, healthcare, and professional services. The country and the fieldwork dates aren’t published.
What that 22% covers is the stretch before anyone types your name into anything. No brand in mind yet, no shortlist, no tab open. Miss that stretch and you don’t get evaluated badly — you don’t get evaluated.
The Executives Themselves Are In There Every Day
90% of them use a chatbot as part of daily work. 85% say their usage grew over the last six to twelve months. And 47% have moved past asking it things on impulse and wired it into a standing workflow.
That 47% is the number I’d underline. A workflow means the same question gets asked again next month, and the month after.
Which means the list of “the three companies you’d call in this category” is being assembled somewhere other than the conference floor.
“The senior people aren’t using this stuff yet” doesn’t hold for these hundred, at least.
Peer Recommendations Still Win, By 27 Points
49% against 22% is not a story about AI overtaking word of mouth, and reading it that way will get you in trouble internally.
Promise the board that AI visibility replaces the referral engine and you’ve written a cheque the data doesn’t cover.
The honest framing is narrower: this is the channel that fills the candidate list before anyone asks a human.
Because asking a peer “what do you think of these guys?” requires already having a name to put in the question. That name comes from somewhere, and a share of “somewhere” is now a model’s answer.
Measuring Only ChatGPT Leaves Out The Half That Matters
The tools split by job. Claude takes 44% for strategy work and long-form writing, ChatGPT 29% for content and copy, Gemini 22% inside Google workflows, Copilot 12% inside Microsoft ones.
So a visibility check that only queries ChatGPT skips the assistant these people open when they’re deciding strategy.
Our own tracking says the same thing from the other side. Asking the same five topics across four assistants, the number of sources each one cited came out at 106 for ChatGPT, 100 for Perplexity, 74 for Gemini, and 72 for Claude (our own measurement, n=20 queries, ChatGPT/Gemini/Perplexity/Claude, measured 2026-08-05).
Same questions, four different fields of view. Whichever one you measure is the one you’ll believe.
What You Have To Say Out Loud When You Quote This
The base is 100 people. One decent-sized meeting room.
There’s no claim of a sampling frame that represents a population. It’s interview research, and it reads as interview research.
Neither the country nor the fieldwork period is published, so you can’t date it or place it. That rules it out as evidence for where the Japanese or any other specific market stands right now.
It’s also self-reported. Not logs of what these executives opened, but their account of what they open.
And Profound, one of the two firms behind it, sells AI-visibility measurement. “Buying now starts in AI” is a conclusion that happens to be good for the product. Discount accordingly.
If you put this in a deck, write “n=100” and “dates not disclosed” next to the number, in the same breath. Leave them out and the first person who asks will take the whole slide down with them.
I’ve written before about what B2B buyers do with the names the model gives them. This one sits one step earlier: where they go to get names at all.
Write Down Three Buying Questions And Ask Four Models
If one thing survives from this study into your next meeting, make it the ordering, not the percentages. The starting point for a B2B vendor hunt moved above search, in a room of a hundred people who buy for a living.
That’s enough to justify checking whether you’re in the candidate set before anyone searches your name.
The thing to decide first is the questions, not the tooling. Write down three questions a buyer in your category would plausibly type — “compare the [X] tools for [industry] companies”, “what are the options for a mid-market company adopting [X]” — and put the same three to several assistants.
If your name never appears, that’s a touchpoint you’re losing upstream of any search. If it does appear, read how you’re described; that copy is usually more useful to your sales team than another positioning workshop.
My own read is that this study is background, not evidence. It earns you the right to go measure your category in your language. What you actually bring to the meeting is the answer those three questions give back.
So before the deck gets built, go ask them yourself.
