You have seen the numbers. Some enormous percentage of B2B buyers now research with AI. Conversions from AI referrals up by some triple-digit multiple. A specific share of buyers who switched vendors because a chatbot told them to. They appear in pitch decks, in board updates, and in the opening paragraph of nearly every article about this discipline, including articles selling exactly the service the number justifies.
I run an AI search agency. Those statistics are extremely good for my business, and I do not quote them on this site. This post explains why, and gives you a way to read any claim in this field, including the ones I would benefit from you believing.
The circle
Start with a small experiment. Take any striking AI search statistic and try to trace it to its origin. Not to the article you read it in, to the study that produced it.
Most of the time the trail runs like this: a marketing blog cites an industry roundup, which cites another roundup, which cites a vendor’s report, which is a survey the vendor ran of its own audience or a study of data from its own product, published as content marketing. Somewhere in there the qualifier gets dropped, the sample gets forgotten, and a survey of a self-selected list becomes “73% of B2B buyers.”
This is not fraud. Vendor research is a legitimate genre, and some of it is careful. The problem is structural: the companies with the strongest incentive to produce these numbers are the ones selling the remedy, and they are also the only ones with easy access to the data. So the field’s evidence base is disproportionately produced by interested parties, then laundered into common knowledge by repetition.
Once you see the circle you cannot unsee it. Two articles quoting “the same finding” are often quoting the same original press release.
Four tests
You do not need research training to filter this. Four questions do most of the work.
Who paid for it, and what do they sell? Not disqualifying, but it sets your prior. A visibility platform reporting that visibility is urgent is not lying, it is answering a question it chose because it liked the likely answer.
Who was actually asked, and how many? A survey of a vendor’s newsletter list measures that list, not “B2B buyers.” Ask for the sample size and how respondents were recruited. If neither is published, the number is decoration.
What was the exact question? “Have you ever used an AI tool during a purchase process” and “did an AI tool change which vendor you chose” produce wildly different numbers and get reported with the same headline. Enormous adoption figures usually rest on the loosest possible definition of use.
Is it a measurement or a forecast? Predictions about search volume years out get quoted as if they were observations. They are opinions with a decimal point.
Run those four against the next AI search statistic you see. Most collapse on the second.
What I actually believe, and why it is not a number
Here is the awkward part for a post like this. Having disqualified most of the available evidence, I still run this work full time and recommend that B2B SaaS companies take it seriously. On what basis?
Not on the stats. On three things that do not require anyone’s survey.
You can verify the mechanism yourself in an afternoon. Open the engines, type the questions your buyers ask, and read what comes back. If your category’s questions get answered with a confident shortlist that does not include you, that is not a projection, it is your current situation, and it is measurable on a schedule.
You can watch it in your own analytics, once you fix the attribution. AI referral traffic is undercounted by default, and the honest read is that whatever you see is a floor.
And you can ask your own buyers. Sales calls where the prospect arrives with an opinion they did not get from you are evidence in your own building, gathered from your own market, with a sample that actually matters to you.
That is a smaller, duller evidence base than a headline percentage. It is also yours, current, and unfalsifiable by a vendor’s next press release.
The peer-reviewed exception
There is genuine academic work in this space, including controlled experiments testing which content changes affect inclusion in generated answers. It is worth more than the survey pile, and I will not summarize its findings here as a number, because the honest caveats matter more than the headline: the engines it tested have since changed, and results from a research setting do not transfer cleanly to your category.
Read the actual papers if the question matters to your decision. Do not read a marketing blog’s summary of a paper, which is how most of these findings enter circulation wearing a confidence they did not start with.
What I got wrong
I used to quote these figures. My early pitch deck opened with a big adoption percentage, because that is what everyone does and it worked, in the sense that nobody challenged it.
Then a prospect did. A former analyst asked, politely, where the number came from. I did not know. I had taken it from an article that took it from somewhere else, and when I went looking afterward, the trail ended in a vendor survey with an undisclosed sample. I had been repeating something I could not stand behind, to people making budget decisions.
What I noticed next was more useful than the embarrassment. The deck worked better without it. Opening with a real finding from the prospect’s own category, three engines, their actual buyer question, the competitor named ahead of them, does something no industry statistic can: it is about them, and they can check it while you are talking. That became the audit I run before every engagement, and the rule on this site has been the same ever since. If I cannot verify it, I do not print it.
Why this matters beyond good manners
There is a practical reason to care, and it is the same mechanism this whole discipline runs on. AI engines assemble answers from what sources agree on, so a claim repeated across enough pages becomes what the engine tells your buyers, whether or not anyone checked it. Corroboration is the currency, and unsourced statistics corroborate beautifully. The field is currently teaching the machines its own marketing.
You cannot fix that. You can decline to add to it, and you can refuse to make decisions on it. When someone quotes you a number about AI search, including me, ask where it came from. If the answer is a link to another article, you have learned something about the number and something about the person quoting it.
If you would rather see what is actually true in your category than what is true in a press release, that is the method, and the service is the version where I run it for you.