It all started with a finding I couldn’t explain. For our 2027 Consumer trends report, we had asked 4,000 consumers in the US and the UK to walk us through something they’d bought for the first time. Among the purchases where an AI assistant had been involved, the share going to a brand the shopper had heard of but never got round to buying rose from 24% to 36%. It was the largest single effect in the study. It replicated independently in both markets. It held across categories.
And I had no idea why it was happening.
That’s the limit of a survey, and it isn’t a flaw in survey design. A well-built questionnaire can establish that something is true, how often, and for whom. It cannot tell you what was going through someone’s mind at the moment it happened, because you’d have had to know to ask.
So alongside the survey we ran 60 AI-moderated interviews through Attest Explore, 30 audio interviews in the UK and 30 in the US. These weren’t the same individuals, but they came from the same audience, with the same demographic criteria and the same screening question applied. Run your qual through a separate agency with a different screen and and you can’t be sure the people you interviewed are anything like the people who answered your survey.
And because Explore sits in the same platform, there was no second project to set up and no separate budget to find. The interviews came out of the same credit pot as the survey, and ran the same week.
The answer was in the fourth interview
A man in the UK had seen a home drinks machine on Facebook. He liked it, and then he left it, because he already owned something similar and couldn’t work out whether a second one was worth the money. It sat there for a while.
Then he opened an assistant.
“I asked Gemini more questions about the machine, like what beers can I put in it, and then once I found out I was confident.”
That’s the whole mechanism, in one sentence, from someone who had no idea he was explaining anything.
He wasn’t introduced to the brand, or even recommended it. He had already decided he quite wanted it, and one unresolved question was parked in front of the purchase. The assistant answered it in a few seconds, and he bought.
Once I’d seen it there, I could see it everywhere in the transcripts. A woman who had heard of a skincare brand but never tried it, waiting to know whether it would suit her skin. A man who’d known about a software product for a while without knowing whether it did the one thing he needed:
“I heard of it but I didn’t know the features and the stuff that you were able to do on it until actually doing research.”
The survey proved the lift existed, and told us it wasn’t discovery — these were brands people already knew. The interviews showed it wasn’t persuasion either. It was a backlog of nearly-decided purchases being cleared. That reframed the whole chapter, and it changed what we told brands to do about it. You don’t respond to “AI converts awareness” the same way you respond to “AI answers the one question that was holding people back”. The second tells you what to go and fix.
Then they stopped me writing something wrong
The survey told us that 56% of shoppers whose purchase involved an assistant had ruled at least one product out on what it told them, without checking anywhere else. I had written a line about AI taking charge of what gets ruled out.
The interviews didn’t support it.
In most of them, the shopper arrived with the rejection already formed. The assistant agreed with a decision they’d made on their own:
“It wouldn’t matter how much Gemini described them to me. If I don’t want to buy from somewhere, I don’t.”
That is a meaningfully different claim, and it changed the copy. We now say the assistant is where rejections get finalised, rather than where they originate. It’s no less serious for a brand — a product dropped inside a conversation is just as invisible either way — but it’s what the evidence supports.
This is the part clients tend not to expect from qual. It isn’t only there to add colour to the data. It can tell you that the story you’ve built on a number is the wrong story, before you present it to the board.
And they answered a question I never asked
Our survey established that 46% of shoppers think an assistant’s product advice is “usually” or “always” impartial. It never asked how they’d reached that view, because it hadn’t occurred to me that it was a question.
The interviews answered it anyway, over and over, unprompted. People read neutrality off the shape of the answer:
“It wasn’t pushing one item on me. It compared them both. It gave me the cons and pros of both of them.”
That’s not a belief about who owns the technology, or how it’s funded. It’s an inference from format: a list, two sides, no obvious advocate. One man went further and described exactly why he hadn’t double-checked anything:
“It gave sources where they got the information from, so that I could check further if necessary.”
He didn’t check. Knowing he could was enough.
That matters commercially, and we’d never have found it in a survey. If assumed neutrality is inferred from the format of an answer rather than from any view about the platform, then advertising doesn’t have to change what an assistant recommends in order to damage it. It only has to make the answer look like it’s arguing for something. That’s now a whole section of the report, and it started as an aside in an interview about hair products.
Four ways to use this on your own study
Most of the clients I talk to struggle to see where interviews fit alongside surveys, beyond a few quotes for the front of a deck. Here’s where they earned their place on this project.
When you have a number you can’t explain. The most valuable thing qual did here was answer a “why?” question. If your tracker has moved and nobody can account for it, 30 conversations will usually tell you more than another wave of the same questions.
Before you write the survey, not after. Several of our answer options came from listening to how people described what they’d done. You can’t write a good closed question about a behavior you haven’t heard anyone describe.
To know what people had in mind when they answered. We asked whether an assistant had suggested products or services people hadn’t heard of, and 78% said yes. The interviews told us what they were picturing — often a new model from a brand they already used, which is completely different from brand discovery. That changed how we wrote the number up, and stopped a product-level finding being reported as a brand one.
To stress-test a conclusion before it goes out. Run a short round against the finding you’re least sure of. Ours moved a claim from “AI decides what gets ruled out” to something defensible, which is cheap insurance for a report that’s going to be read by research and insights professionals.
Sixty interviews, thirty per market planned, fielded in a few days alongside the survey rather than as a separate project months earlier. That’s the part AI-moderated interviews have changed. Qual used to mean recruiting a room and waiting six weeks for a debrief. When it runs at the same pace as the survey, you can use it to answer the questions the survey raises, while there’s still time to act on them.
