Jeetendra Gangele is Cofounder and CTO of BluePill AI, leading engineering for AI consumer twins and synthetic consumer research.
Ask a general-purpose chatbot what a busy mother would think of your new cereal concept, and you will get a confident, well-written answer. Ask it about your weakest concept, and the answer will likely be identical. That’s a problem.
Synthetic consumers, also called AI consumer panels, are moving quickly into consumer research. Instead of six weeks and six figures to field a study with human respondents, brands query models built to respond the way real consumers would.
The basic value proposition is sound. My concern is narrower: fluency is mistaken for accuracy. If the output sounds like a thoughtful consumer, how do you know it actually reflects one? Executives need a way to make that distinction.
I’ve found the academic record is more useful here than most vendor decks. A 2023 paper in Political Analysis showed that language models conditioned on demographic profiles could reproduce human survey patterns. Additionally, a 2024 review in Psychology and Marketing found the same models approximated some consumer responses while failing to reproduce well-documented effects such as the endowment effect and mental accounting.
The Market Research Society’s 2024 Delphi report separates statistically synthesized data from LLM-generated participants, and warns the industry against assuming the second inherits the track record of the first.
The evidence is mixed. Four questions separate a research instrument from a writing exercise.
1. How many real humans is this built from, and what did each contribute?
A base language model knows consumers from the outside. While it has read a great deal about busy mothers, it has met none of them. It returns the average of everything written on the subject, and averages do not buy products.
Systems need to be grounded in primary human data: long-form interviews with real, consenting people, used to build one digital twin per person. A real consumer might buy organic cereal for her children and a donut for herself. A twin built from her interview keeps that tension, while a persona built from a prompt may not. This is because the prompt contains only what the buyer already assumed.
If the system is simply instructed to act like your target consumer, you are testing how well a model can play a role—not how well it can represent a real consumer.
2. What has been validated, and against what?
Most validation in this category is measured against human panels. Run the same study with real people and with twins, then compare the results. That provides a concrete measure of how closely the synthetic panel reproduces human responses. A 2025 study of 57 personal care surveys reported synthetic ratings reaching 90% of human test-retest reliability on purchase intent.
The numbers are narrower than they sound. Agreement with a human panel demonstrates that twins reproduce what people say. It does not demonstrate that they predict what people do. The two are not the same thing, and the research industry has known it for decades. Jamieson and Bass established in 1989 that stated purchase intent substantially overstates actual trial, by margins that vary heavily by category.
So ask two questions:
1. What is the agreement with human panels?
2. What is the evidence against real market outcomes?
Most of this field has considerably more of the first than the second. Anyone claiming stronger predictive power should be able to show it: for example, concepts predicted without revealing the expected result, followed by a comparison with actual sales or trials. Ask for the study design, the metric, the benchmark and the failure cases. A demo is not a validation.
3. Where does it break?
Twins built from interviews inherit stated preferences and articulated reasoning. That makes them strong on deliberate decisions, what psychologists call “System 2” thinking. It makes them weaker on impulsive behavior that drives a large share of packaged goods purchases. A last-second candy purchase at the checkout counter is not in the interview, because it was never a decision the person could narrate.
That is a limitation of the method. It is an open problem for the whole category, and anyone describing it as solved is likely just trying to sell something.
Twins also go stale. One built before a price shock represents a consumer who no longer exists, making it important to refresh them against new human research.
None of this makes the technology unusable. Instead, it helps define where it is usable. Concept testing, packaging evaluation, claims work and exploratory qualitative research are well served today. High-stakes decisions and anything requiring observed behavior still need people.
4. Are we asking the most valuable question?
Most of the current debate reduces to whether synthetic respondents are as accurate as humans. It’s a fair question, but I don’t believe it’s the most important.
The larger opportunity is testing what conventional research cannot reach. Evaluate a holiday product in December rather than asking someone in July to predict their December self. Run the same concept under high inflation and low inflation. Run it against three different competitive sets. External context shapes purchase behavior as much as the product does, and a traditional study captures one moment of it.
Use synthetic panels to vary context before committing to human research. Take a chocolate concept and test it in a July shopping basket, a December gifting basket and against a much cheaper private-label rival. If appeal stays steady but purchase intent falls, you have isolated a potential context effect rather than simply a problem with the concept. That kind of scenario testing can help you decide which questions are worth taking to real consumers.
That is the bigger value of synthetic panels: they make context easier to test as a variable. But that only works if the underlying consumer representations are grounded in real human data. Otherwise, you are simply generating plausible answers at scale.
Conclusion
Market research earned its place in the boardroom by being accountable to evidence. This technology deserves the same treatment, applied by the people buying it.
Show how the synthetic consumers were built, what they were validated against and where that validation stops. Show the failure cases, too. Then ask those questions before the category settles on its standard, because changing an established standard is much harder than setting one.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
