“This Only Pays Back For Companies People Already Know”
That sentence arrived halfway through a meeting about next year’s numbers, and it stopped the room, mostly because it sounds true. I said nothing. It wasn’t that there was no answer; I just didn’t have a figure in my hand at the time, and an argument without a figure loses to an argument without a figure that was said with more confidence.
Somebody in your building has said the same thing, I’d guess — that all this AI work does is chew through the same demand your branded search was already collecting, so the money moves sideways and nothing new comes in. It matters which way that goes. If the models only ever hand people back to companies they already trust, then measuring what the models say about you is a maintenance chore. If the models put unfamiliar names in front of people who then buy, it’s an acquisition channel with no analytics attached.
Somebody went and asked shoppers in the US and the UK which of those two worlds they’re living in.
Does An AI Recommendation Ever Land On A Brand Nobody Knows?
It does. Across a survey of more than 1,000 online shoppers in the US and the UK, one in five said they had bought a product from a brand they had never heard of before, because an AI recommended it (Rithum with Retail Dive, 2026).
That’s the number I wanted in the meeting.
Underneath it sits the usage figure that makes it plausible: among shoppers aged 44 and under, over 80% had used a large language model while shopping in the past three months. A large language model — an LLM — is a system trained on an enormous pile of text that answers questions in sentences rather than links; ChatGPT is the one everybody has open.
Narrow to households earning $100,000-$150,000 and the figure climbs to 84%. So the people with money to spend are, if anything, further into this than everyone else, which is an awkward shape for a wait-and-see plan.
Where Do Buyers Go To Check After The AI Names Someone?
Search engines, review sites, and people they know. The retailer’s own website gets 5%.
Read that 5% on its own and you’d conclude your site barely matters any more, which is a comforting reading if your site is a mess and a terrifying one if you’ve just rebuilt it.
But look at where the other 95% go. The wording of the reviews, the running order of the comparison posts, the sentence a trade publication uses to describe what you sell — every one of those surfaces is something you can put work into, and none of them are your homepage.
I wrote about the checking behavior itself a while back: shoppers touch an average of 2.4 platforms after a model names somebody, so the answer isn’t the end of the trip. How much of your name is sitting out on that 95% side?
One more thing from the same survey, and it cuts against the tidy version of this: higher-income, higher-intent shoppers were more likely to buy without checking anywhere at all. So the verification journey isn’t a corridor everybody walks down — some of them go straight from the answer to the checkout, which means the answer itself is doing the whole job for that group.
What Happens When The AI Gets Your Product Wrong?
58% said their trust in the brand goes down when an AI gives them incorrect product information. 16% said they’d drop the purchase.
Sixteen percent reads small next to fifty-eight. It’s the one I’d put on the slide.
Out of a hundred people who were going to consider you, sixteen leave over a sentence you didn’t write, about a product you do know, delivered by something that sounds authoritative and has no obligation to be right. It’s a stranger reading your company description aloud to every prospect, from a script nobody has checked.
I’m not sure the 16% generalizes cleanly — self-reported intent to abandon a purchase is exactly the kind of thing people overstate when a survey asks them (see the caveats below, which apply to this figure hardest of all). But the direction survives even if the level doesn’t.
Which is the argument for measuring what the models say about you before you decide how much it’s worth: you can’t price a risk you haven’t counted.
The Four Caveats That Go With This Number
Paste the 20% into a deck without these and somebody will take your legs out from under you in the Q&A.
- Rithum sells channel management and advertising for e-commerce. “Get ready for AI-driven discovery” is the conclusion that sells the product, so read the direction of travel with that in mind
- The respondents are US and UK consumers. This is not a measurement of Japanese-speaking shoppers, and nothing here promises they behave the same way
- It’s a self-reported survey, not purchase logs. People are describing what they think they did
- The publisher states no limitations. The fieldwork period, the sampling method and the question wording are all unpublished
The fourth is the one that’ll come back at you, because “where did this number come from?” is the first question any decent analyst asks.
And still: 20% is not a rounding error you can wave off. Different country, different survey design, and people who didn’t know the name are buying anyway.
Point The Target At People Who Don’t Know You Yet
If I take one thing out of this, it’s about where the goalposts sit. Measure AI visibility by whether branded search holds steady and that 20% never shows up in your reporting at all — the whole point of it is that those buyers had no name to type.
The second thing, and I’d rank it below the first, is checking that what the models say about your products is actually true. The 58% and the 16% are the bill for leaving that alone.
Since 95% of the checking happens somewhere other than your own site, most of what needs fixing lives outside it too. Read the reviews and the comparison pages as often as you read your own analytics and you’ll know which of the two you’re actually behind on.
Try counting how your name comes out of a model’s mouth first. The rest is easier to argue about once there’s a number on the table.
Sources
- Rithum & Industry Dive (Retail Dive), “Consumer Trust & Discovery Report 2026: The new discovery engine”, 2026 (survey of more than 1,000 online shoppers in the US and UK. One in five bought a product from a brand they had never heard of because an AI recommended it; over 80% of shoppers aged 44 and under used a large language model while shopping in the past three months, rising to 84% among households earning $100,000-$150,000. After an AI recommendation, buyers verify via search engines, reviews and acquaintances, with only 5% going to the retailer’s own site, while higher-income and higher-intent shoppers are more likely to buy without verifying. 58% say incorrect AI-supplied product information lowers their trust in the brand and 16% would abandon the purchase. Self-reported survey data, US and UK only; the publisher discloses no limitations, and the fieldwork period, sampling method and question wording are not published. Rithum sells e-commerce channel management and advertising.)
