AI Talked as Many Buyers Out of a Purchase as Into One

Semrush asked 2,338 US adults in July what artificial intelligence had done to their shopping. Two numbers came back almost identical. 57.5% said a chatbot had talked them out of buying something. 57.51% said a chatbot had talked them into buying something.

That’s a coin flip, and the study carries a margin of error of two points either way, so treat the two as the same number rather than a photo finish. Every pitch for answer-engine optimization you have received this year sold you the second figure. The first one was in the same dataset the whole time.

The assistant works as a filter. It removes buyers at roughly the rate it delivers them, and only one of those two jobs has been priced into anybody’s retainer.

The deciding input is on somebody else’s website

The mechanism sits one layer down, in a question about reviews.

74.15% of respondents said they would be at least somewhat less likely to buy a product if a chatbot told them its reviews were mixed or negative. Among people who actually use AI regularly, that rises to 85.31%. Nothing else in the survey moves behaviour that hard.

So the model reads a verdict aloud from wherever it found one. It invents nothing.

Where it found one is measurable. Muck Rack analysed more than 25 million links cited by ChatGPT, Claude and Gemini across 17 industries and found 84% of citations went to earned media: journalism, third-party write-ups, review sites, community threads. Paid and advertorial placements accounted for 0.3%. The pattern has held across three of their reports since July 2025.

Greg Galant, Muck Rack’s co-founder and CEO, read the run of reports plainly:

“Three editions in, the data keeps telling the same story: earned media is what AI trusts. For communications teams, that is validation that the work they do to earn coverage in the right outlets has real consequences beyond traditional metrics.”

He’s describing the upside. The Semrush numbers describe what the same mechanism does when the coverage is unflattering.

Put those two findings next to each other and the sentence that decides a deal is one you didn’t write, on a site you don’t control, summarised by a model you can’t audit. We’ve measured the shelf life of the good version of that sentence before, and the median AI citation decays inside a month.

Why this is not the old reputation problem

Review management has existed for fifteen years. What changed is who reads them.

A buyer skimming a review page applies their own discount. They notice that the angriest review is three years old, that the reviewer wanted something the product never claimed to do, that the four-star average rests on 900 ratings and the two-star complaint has eleven. That discounting happens automatically and nobody has to be taught it.

A chatbot compresses the same page into one clause. Answer engines quote people rather than brands, and a person’s two-star complaint compresses cleanly. “Reviews are mixed.” The three-year-old complaint and the 900 ratings arrive as a single sentence with no age, no volume and no denominator attached. Then 74% of the people reading that clause step back.

The compression is the story. Not the sentiment.

What the survey does not say

Two limits, and they matter before anyone reallocates a budget on this.

The study is self-reported. People are describing what they believe a chatbot did to a decision they made, which isn’t the same as a measured attribution, and nobody in this dataset was watched making a purchase. Recall about why you did not buy something is the least reliable kind there is.

It’s also a consumer survey of US adults rather than a B2B buying-committee study. A 2,338-person sample with a two-point margin is solid for what it measures. It isn’t evidence about how a six-person software evaluation runs, and anyone selling you a B2B conclusion from this number is stretching it.

What survives both caveats is the asymmetry in what marketers are being sold. The upside of AI discovery has been measured, published and priced into retainers. The downside was measured in the same survey and nobody built a service around it.

What to do about it this month

Three steps, in order, none of which needs a new vendor.

1. Ask the assistants what they say about you. Open ChatGPT, Claude and Gemini and ask each one whether it would recommend your product, then ask what the reviews say. You’re not looking for a score. You’re looking for which source it names. That source is your actual storefront.
2. Fix the cited source, not your own site. If the model quotes a two-year-old G2 thread or a Reddit comment, that page is doing more work than your homepage. Your own domain is roughly one in six of what gets cited. Spending the quarter on product-page copy while the model reads somebody else’s page is the misallocation this data exposes.
3. Ask your agency for the blocked number. Every AI visibility tracker on the market reports the mentions. Ask what share of those mentions carried a caveat about your reviews, pricing or support. If the answer is that they don’t track it, that half of the coin flip is unmeasured, and it’s the half that costs you deals.

The one-line version: buyers are asking a machine whether you are any good, and the machine is answering from a page you’ve never edited.