There is a number floating around the market research world that deserves more attention than it gets. When researchers were surveyed about AI in their work, 97 percent said they use it somewhere in their process. Same profession, same survey territory, a different question: do you trust synthetic users, AI-generated stand-ins for real research participants, to substitute for the real thing? 8 percent said yes. 64 percent were opposed outright.
Read those numbers together and the story gets interesting. This is not a population of Luddites. These are people who use language models every working day, for transcription, for summarization, for cleaning up discussion guides. They know exactly what the technology can do, and that is precisely why they refuse to let it play the part of a human being. The people closest to the machine trust it least for this one job.
The pitch is seductive. Instead of recruiting real customers, you prompt a model to be one. "You are Sarah, 42, a marketing director at a mid-size logistics company, anxious about proving ROI." Then you interview Sarah. Sarah answers instantly, at 2 a.m. if you like, never cancels, never asks for an incentive payment, and never says anything that makes the room uncomfortable.
The computing world has already named the trap. Researchers writing in ACM territory called it the Synthetic Persona Fallacy: the belief that a model which has read a great deal about people like your customer can therefore substitute for your customer. It cannot, and the reason is structural, not a matter of waiting for a better model. (The glossary entry has the short version.)
Sarah cannot deviate from the average. A language model generates the most probable response given everything written about people vaguely like Sarah. But the entire commercial value of research lives in the deviation, the place where your buyers differ from the plausible average. When we researched a test-prep market, the average said parents buy on scores and credentials. The real buyers, read in their own forums, were buying whatever made them feel like good parents. That gap between the average and the actual is where a sales page doubled its revenue. Sarah would have recited the average, confidently, with excellent formatting.
Sarah cannot have stakes. She has never spent her own money, lain awake over a decision, or felt embarrassed at a school pickup. Real buyer speech is shaped by consequence, and consequence is exactly what a simulation lacks. The forum post written at 11 p.m. by a parent who just paid for the third tutoring program that did not work carries information no prompt can conjure, because the information came from somewhere.
Sarah cannot be checked. This is the one that should end the argument. When a real buyer says something surprising, you can verify it: find the same theme in other threads, in reviews, in a rival's comment section. Evidence accumulates or it does not. When Sarah says something surprising, there is nothing behind the sentence to check. She is not reporting an experience. She is completing a pattern. An interview you cannot verify is not research, it is theater with a transcript.
A synthetic persona cannot be wrong in any way you could catch. That is not reliability. That is the absence of the thing research is for.
There is a fourth problem worth naming, because anyone who has spent time with these models knows it in their hands: Sarah wants the interview to go well. Language models are trained to be helpful and agreeable, and an agreeable research subject is a compromised one. Ask Sarah a leading question and she will follow the lead. Show her your new positioning and she will find something to like about it. Real buyers do the opposite. They shrug, change the subject, and fixate on a detail you thought was minor, and that friction is the data. A subject incapable of disappointing you is incapable of informing you.
Because real buyers are inconvenient. They ramble, contradict themselves, take days to recruit, and say things that complicate the strategy deck. Sarah is fast, cheap, and agreeable. She tells you a flattering version of what you already believed, which is a pleasant experience that feels like validation and costs almost nothing.
And there is a quieter reason, the one Girard would point at. Synthetic research is spreading the way most tools spread in marketing: mimetically. Companies adopt it because other companies adopted it, and the case studies are about speed and cost, almost never about a decision that turned out to be correct because a synthetic user predicted something real buyers confirmed. The 8 percent number suggests the practitioners have noticed.
We read real buyers. For every Mimetic Intelligence Snapshot, a human reads 200 or more primary sources from your specific market: the Reddit threads, the forum arguments, the reviews where the anger is specific, the community discussions where your buyers talk to each other with nobody selling to them. Nobody prompted those people. They wrote because they had something to say, and the writing carries the fingerprints of actual stakes.
I will be straight about where software fits, because pretending otherwise would be its own kind of slop: tools help us collect and organize sources. They are good at that. The reading, the judgment about what matters, and the call about what it means for your business are done by a person, and the findings are checkable because every claim traces back to real humans saying real things in public. That is the standard the 92 percent are holding out for, and I think they are holding the right line.
If your strategy is currently resting on an AI-generated avatar, I am not saying you learned nothing. I am saying you learned what the average sounds like, and your competitors who prompted the same model learned the same average the same afternoon. Your revenue lives somewhere more specific.
The free Mimetic Intelligence Snapshot is built by a human from 200+ primary sources in your market. No synthetic anything. Who your buyers copy, what rivalries drive them, where desire is heading, one concrete move. 5 business days.
Get Your Free Report