Your AI persona has never used your product.

It has read every review ever written, and it has never once been disappointed.

The persona in your research has never used your product.

It has read everything anyone has ever written about using it. Every review, every complaint thread, every comparison, every recommendation. And it has never opened the package, never been let down by it, never stood in an aisle deciding whether to buy it again.

For a while I didn't think that mattered much. I've come to think it's the whole thing.

Here's what changed my mind, and it started with something I noticed in focus groups long before any of this existed.

People love telling you what they don't like.

Put eight people in a room and ask what they think of something. The praise is polite, vague, and over in about a minute. The complaints are specific, detailed, occasionally unreasonable, and they will go on for as long as you let them. Somebody will tell you the packaging is annoying in a way that has clearly been bothering them for two years. That is the most valuable twenty minutes you will buy all quarter, and you don't have to ask a clever question to get it. You just have to be in the room with a person who has something to get off their chest.

But I don't think that's only about people being negative. There's a mechanism underneath it, and it explains what the machines are missing.

Think about walking offstage after speaking. Somebody you don't know says good job. You say thank you and keep walking. You don't turn around and ask them good job about what. You don't want the details.

Now imagine that same person says it could have been better. You stop. You want to know exactly what they meant, which part, what you should have done instead. You will stand there for ten minutes to find out.

Praise gets accepted. Criticism gets interrogated.

The specificity isn't in the negative feedback. It's in the follow-up. The useful material gets manufactured by pressing a real person on a real memory until they get precise about it.

There's nothing behind an AI's answer, because it has no real experience to draw on. Ask a persona what it meant and it will hand you a reason instantly, because producing reasons is the whole of what it does. But it isn't remembering anything. There's no afternoon it can go back to. So what it will never say is the thing a real person says when you ask again.

The gap isn't information. It's experience.

The machine has never done the thing. It has only ever gathered up what other people said about doing the thing. It can give you an answer. It has not been let down by your product in the second week.

A customer knows what your product is like after the novelty wears off. When the packaging that seemed fine in the store is annoying on a Tuesday morning. When the thing that sold them stops mattering and something they never thought to mention starts to.

That is the part they will tell you about, in detail and unprompted, if you are in the room and you ask a second time.

There's an obvious objection here, and it's getting truer every month, so I'd rather raise it than wait for someone else to.

AI is running customer service now. Which means it is having millions of conversations with genuinely unhappy people, hearing specific complaints, unprompted, at a volume no focus group could ever reach. If my argument is that machines never encounter real dissatisfaction, that argument is expiring.

I think it narrows the claim rather than ending it, for three reasons.

Support is also trying to solve a problem, while research is trying to understand one. Support asks: Did that work? Are you satisfied with the resolution? Research asks: What were you expecting? When did you first notice? What did you do instead? And the customer is in a different mode. Someone in a support queue wants the problem solved and wants to leave. Someone in a focus group will wander into things they didn't come to say. That wandering is often where the strategy-changing finding lives.

There's also a cost to automating that channel that I don't hear discussed.

A support rep who takes forty calls a week develops a hunch. Something changed this month. People keep mentioning the lid. That hunch used to travel, informally and unreliably, up to product and marketing, usually because somebody complained about it in a hallway.

Automate the channel and every transcript becomes searchable, which is a real gain. But nobody has the feeling anymore. You can query what was said, and you have lost the person who noticed.

I learned to trust that the slow way, on a project years ago.

I was doing audience research for a group of food brands. Canyon Bakehouse, Nancy's Yogurt, and a2 Milk.

The usual way to start a project like that is to invent a person. You pick an age range, a household income, a city. You add a stock photo of a woman holding a grocery bag. Sometimes you give her a name. Then everyone writes to her for the next two years.

We did the opposite. We looked at the people already there, not just what they said about bread or yogurt, but what else was in their lives. Who did they follow? What other brands appeared in the same feeds? What did they post about when nobody was selling them anything?

We also collected vocabulary. The words people used in their own posts, in everyday life, when they were talking to each other instead of answering a survey. That language shaped how the brands wrote.

I don't have it anymore, and I'm not going to reconstruct it from memory and call it data. But the principle transfers without the examples. Every category develops a dialect, and it's mostly written by marketers talking to other marketers. Customers almost never speak it. Write in the dialect and you sound like the category. Write in their words and you sound like somebody who has actually met them.

Two things surprised us, and they are the reason I still think about this project.

The first is that we learned more from who people did not follow than from who they did.

The absences were louder than the presences. Categories of brand you would have bet on, looking at the demographic profile, simply weren't there. You cannot get that from a list of what people like. You only see it if you walk in expecting something and then notice it missing.

The second was in the photos.

When people posted pictures involving these products, the picture was almost never of the product. It was of the experience around it. And it was overwhelmingly social. Friends together. People out doing something. The food was in the frame, not the subject of it.

Our marketing looked nothing like that. Product shots on clean counters. One person doing yoga. Tidy, well lit, and alone.

So we changed it. Bright, colorful groups of friends out in real situations, with the products in their lives rather than on a pedestal.

They had already shot the campaign. We just hadn't gone and looked at it.

If that conclusion sounds obvious now, it should. Social-first imagery became standard across the category. It wasn't then, and we didn't arrive at it by copying someone who was doing it well. We got there by looking at what people were already photographing.

An insight that later becomes conventional wisdom isn't a dated insight. It's evidence the method worked.

Which brings me back to the machines, because you can buy a version of that work now.

There are a dozen companies that will do it for a monthly fee. Point them at your brand, get a report by Friday. For a while I figured that made the story obsolete. Why explain how we did by hand what software can do in an afternoon?

Then I read the research testing these tools against real people, and it lined up with what I'd been noticing in rooms for years.

Two studies are worth knowing about, and I'd rather name them than say "studies show."

The first is Bisbee, Clinton, Dorff, Kenkel and Larson, published in Political Analysis in 2024. They asked ChatGPT to stand in for respondents to a large, long-running American survey, then compared what the machine said against what real people had actually said.

The averages came out fine. That's the part people quote when they're selling you this.

The relationships between things did not. When they ran the same analyses on the synthetic answers that researchers run on real ones, nearly half the results came back significantly different from the real data. Of those, about a third pointed in the opposite direction. Not weaker. Backwards.

They also found the synthetic answers varied far less than real people do, which makes you more confident rather than less. And running the same prompt in different months gave different results.

Stay with the backwards number for a second. If you had used that to decide what your brand should say, a third of the time you would have confidently said the opposite of the right thing, with tighter margins of error than the truth deserved.

The second study is from this year, in Artificial Intelligence Review. Corrêa and colleagues generated forty thousand personas across four languages and then looked at who the machine invents when you ask it to imagine a customer.

It invents an idealized one. The paper calls it narrative sanitization. The personas came out aspirational and overwhelmingly upbeat. They clustered around middle age and around technical professions. In one setup, English prompts to one model produced male personas 99.76 percent of the time. Non-binary identities came in under one percent. Whole categories of ordinary work simply didn't appear.

Read that as a marketer rather than as a researcher. If you ask a machine to imagine your customer, it hands you a cheerful, employed, middle-aged, mostly male person who likes things.

That isn't a customer. That's a stock photo with a paragraph attached.

Which is exactly what we were trying to get away from when we stopped inventing personas in the first place. And it is the same gap I started with. The most useful raw material in qualitative research is somebody's specific, earned dissatisfaction, and that is the one thing a system with no experience of disappointment cannot manufacture.

Look again, too, at what our two most useful findings actually were.

One was an absence. A set of brands that should have been in those feeds and wasn't. The other was in pictures rather than words, and the insight was about who else was in the frame.

AI can analyze both now. That's not the limitation.

The harder question is whether anyone, human or machine, knows to look for what's missing.

Pattern recognition is very good at telling us what's there. Research sometimes advances because someone expected to see something, didn't, and got curious about the gap.

There's one more problem, and this one is mine, not theirs. It applies to what I did too.

When you study the people who already follow a brand, you are studying the people you already reached. Everything you learn is downstream of marketing you already did. Doing that faster doesn't fix it. It just makes you extraordinarily precise about the world you already created.

So what would I tell a team now?

Use the fast tools. I'm not going back to five people with spreadsheets. They're good for narrowing a wide field, for early direction, for working out which three ideas deserve real money.

Just don't let one become your evidence.

Build them out of what people actually did, not out of a description of who you think they are. Ask any vendor what their data really is, because "audience interests" often means something much thinner than it sounds. Go find people you haven't already won. And before you spend real money, put yourself in a room with somebody who has used the thing and is willing to tell you what annoyed them.

That last one is the whole argument. The best tools in this category are getting much better at grounding synthetic customers in what real people said and did. But the model still hasn't had the experience it's explaining. It can reconstruct disappointment from evidence. It hasn't been disappointed.

Your customers are still out there, still writing it down, still turning it into content.

Somebody still has to go and ask them what they meant.

Sources

James Bisbee, Joshua D. Clinton, Cassy Dorff, Brenton Kenkel and Jennifer M. Larson, "Synthetic Replacements for Human Survey Data? The Perils of Large Language Models," Political Analysis, volume 32, issue 4, October 2024.

Nicholas Kluge Corrêa, Rafaela Weber Mallmann, David Kaczér, Florian Mai, Ana Ilievska and Julia Maria Mönig, "All too perfect: bias and aspiration in persona generation with LLMs," Artificial Intelligence Review, volume 59, article 187, 2026.