Can You Understand Your Customer Without Talking to One?
Early in my career, I spent a lot of time behind one-way mirrors.
When we launched Cargo at Condé Nast, and later when we made significant changes to InStyle’s architecture or design, we’d run focus groups. Real people, a conference room, a facilitator, a plate of mediocre sandwiches, and a bunch of us on the other side of the glass watching and taking notes and trying not to overreact.
These weren’t casual exercises. We’d spend weeks, sometimes months, preparing. Working with the agency to get the discussion guide and prompts exactly right. Flying to different cities to reach different demographics. Then after the sessions were done, waiting weeks more for the agency to synthesize everything into a report.
And I always wondered about something nobody seemed to want to talk about: what kind of person actually takes the time to show up to an in-person focus group? They were, by definition, a very specific subset of our audience. But their opinions would shape what we did with the product for millions of readers who would never set foot in that room.
We’d bring out chocolates and sweets midway through to keep the energy up. And within the first ten minutes, two or three loud, opinionated participants would take over regardless. They’d state something with total confidence and the rest of the room would start nodding. The facilitator, no matter how skilled, could only do so much.
I remember one Cargo session where participants referred to our editorial features as “ads.” Not sponsored content. Not advertorials. The actual articles. Or someone would comment on the price of a product we’d featured and suddenly the entire conversation would pivot to cost sensitivity, even when price wasn’t remotely the question we were trying to answer.
Imperfect. But real.
The Room Without People
Companies are now using AI to build synthetic consumers. AI-generated personas, fed real behavioral data, that simulate how actual shoppers think and respond. You can test products, campaigns, pricing strategies. Run what amounts to a focus group without a single human in the room.
Bain and Company found that synthetic consumer research can deliver comparable insights in half the time at a third of the cost. Stanford and Google created AI agents based on interviews with over a thousand real people and found the AI replicated their survey responses with 85 percent accuracy. Target has talked publicly about using synthetic audiences to predict how customer segments will respond to campaigns before they launch.
No months of preparation. No flights to different cities. No waiting weeks for a synthesized report. No one calling your articles ads.
This is genuinely exciting. For teams without the budget or time for traditional research, it’s a real unlock. And for anyone who’s ever watched a single opinionated participant derail a carefully designed study, the promise of clean, scalable insight is hard to resist.
The Yes-Man Problem
But I keep coming back to one thing.
The old problem with focus groups was noise. Too many voices, too much groupthink, too easy for one person to distort the room.
The new problem might be the mirror image. Large language models are, at a fundamental level, agreeable. They synthesize and reflect back. Researchers are already finding that synthetic personas skew positive, provide feedback that’s vague and favorable, and lack the irrational, contradictory, deeply human quality that makes real consumer insight valuable.
And they have their own version of the selection bias problem. Synthetic personas are built on the data that exists, which means they represent the people who left reviews, took surveys, engaged online. The quiet customer who just buys your product and never tells anyone what they think? Neither method captures that person.
If the old focus group had the loudmouth problem, the AI focus group has the yes-man problem.
The Other 15 Percent
Here’s the thing I can’t let go of.
When I sat behind that glass, even with all the frustrations, something would occasionally happen that no data set could have predicted. Someone would say something sideways, almost as an aside, that completely reframed how I thought about what we were making. The person who’d been quiet for twenty minutes and then offered something that made everyone behind the mirror look at each other.
Those moments didn’t come from the questions we asked. They came from the mess of putting actual humans in a room and letting the conversation wander somewhere unexpected.
AI doesn’t wander.
85 percent accuracy is remarkable. I mean that. But the other 15 percent is where the weird stuff lives. The contradiction. The thing that doesn’t fit the model. The aside at minute 37 that changes everything.
That 15 percent is where the best editorial decisions I ever made came from. Not from data. From proximity. From being close enough to the reader, the subject, the culture, that I could feel something shifting before the numbers confirmed it.
I’m optimistic about where synthetic research is going. I think it makes almost everything about understanding your customer faster and more accessible. I just don’t want us to get so good at the 85 percent that we forget the 15 percent is where the magic usually happens.