AI synthetic audiences promise customer research in an afternoon, but they can only reflect what you fed the model. This piece takes the position that synthetic panels are a fast first draft for cheap decisions, never a verdict, and that a business which cannot yet measure its real customers has no business simulating fake ones.
A synthetic customer cannot tell you what you failed to imagine. It can only remix what you already handed it.
The Take: A synthetic customer will tell you what you already believe, and it will do it in an afternoon. That speed is the danger. A synthetic panel is a first draft written by a mirror. Treat it as a cheap way to get moving, never as the verdict on a decision that costs real money.
What can a synthetic panel actually tell you?
BCG got a synthetic panel to predict real drink choices with 92% accuracy, but only after months of tuning against actual sales, for a product close to something buyers already knew.
That is the useful ceiling and the honest limit in one line. Fast, plausible and capable of real accuracy once corrected against reality. Useless as a substitute for the reality you corrected it against.
BCG reached 92% accuracy predicting real consumer choices with a synthetic panel, but only after fine-tuning the outputs against real results over time.
Read the conditions on that number, because they are the whole story. BCG itself says plainly that synthetic panels cannot replace traditional market research and get weaker the more genuinely new your product is. The accuracy was earned by grinding the model against reality, not by skipping it. The panel did not know anything. It was corrected until it stopped being wrong.
Why a simulator is not a flight
Think of a synthetic panel as a flight simulator. It is cheap to run, you can crash it a hundred times before lunch and a good one teaches you real things about how the controls respond. Serious pilots spend hours in one. Not one hour of that time counts as having flown.
The simulator only knows the weather you loaded into it. A synthetic customer is built from what the model already absorbed plus whatever you put in the brief. Ask it a question and it answers from that closed world. It cannot surprise you with the thing you did not think to ask, which is the exact thing real research exists to catch.
The panel is only as honest as the reality behind it. If you have no clean read on your real buyers, you are tuning a simulator against nothing.
Synthetic panel versus real research: what you are buying
Speed and cost are real wins. The right response to a tool that is fast and cheap is to use it for the fast, cheap decisions, then spend real time and real money on the ones that carry real risk. That is the whole discipline, and it is not complicated.
Should you simulate a customer you cannot yet measure?
No. This is where most of the market is about to hurt itself.
Australian businesses spent $28.9 billion on advertising in 2025, up 5.2% on the year, according to WPP Media. We have measured marketing across the Australian market on six dimensions, and the one that lands last, over and over, is data and measurement. Most businesses here cannot yet prove whether their real marketing is working.
The order of operations matters. A business that cannot measure its real customers is being sold a tool to simulate fake ones. That is not insight. It is the guess, automated and dressed up in a confident chart. Simulating a customer before you can measure a real one does not remove the uncertainty. It only hides it.
What do the rules already say?
The people who do research for a living have moved faster than the buyers. The 2025 ICC/ESOMAR International Code now makes disclosure mandatory when synthetic data or AI is used, so you cannot present a synthetic finding as equivalent to the real thing without saying so out loud.
The tools are here in Australia already. Enero-owned agency Orchard has launched AudienceIQ, which builds synthetic patient and clinician personas so pharma teams can test messaging in days rather than months. On the first pass, that is a genuine speed gain. As a replacement for talking to a real clinician, it is an expensive way to confirm your own assumptions.
What I would do about it
Fix the order.
Get a clean read on your real customers first. Analytics that fire, conversions you can trust, a survey or ten honest conversations. If you cannot measure reality, you are not ready to simulate it.
Use synthetic panels for the low-stakes, high-volume pass. Screen concepts, cut a long list, pressure-test a headline. Cheap decisions, cheap tool.
Validate every expensive decision against real buyers. Anything that moves budget, pricing or positioning gets tested on people who can actually hand you money.
Disclose it. If a synthetic panel shaped a call, say so, in line with the ESOMAR position above.
Keep tuning against reality. BCG's 92% came from correction, not from trust. Treat every synthetic output as a hypothesis with a use-by date.
Methodology. New Rebellion scores marketing across six dimensions using a mix of machine-observable signals and structured review, benchmarked against Australian businesses. Where the data behind a business is thin, we say so and read the result directionally rather than claiming a precision the evidence does not support. The full methodology is public: how we score marketing.
That habit is the entire point. A number is only as good as the reality you checked it against.
In my opinion, synthetic panels are one of the more useful research tools to arrive in a decade. They are also one of the easiest to misuse. They will make good teams faster and make guessing teams more confident about their guesses. The line between the two is whether you can already measure a real customer. A flight simulator will train you for years and never once actually fly, and it will never crash when you make the mistake that would have cost you the aircraft. Get your real numbers first. If you want a clear read on where your marketing actually stands before you simulate anything, start with a real measurement.
Frequently asked questions
Are synthetic panels accurate enough to replace market research?
No. BCG reached 92% accuracy only after tuning against real sales and only for a familiar product, and it states outright that synthetic panels cannot replace traditional research. Treat them as a first pass, not a verdict.
When should I use a synthetic panel?
On cheap, high-volume, low-risk decisions. Screening concepts, cutting a long list of ideas, sanity-checking a message before fieldwork. Keep the expensive calls on real buyers.
Do I have to disclose that I used AI in research?
Yes, if you follow the 2025 ICC/ESOMAR Code. Disclosure of synthetic or AI-generated data is now mandatory, so a finding cannot be passed off as equivalent to traditional research without a caveat.
What if I cannot measure my real customers yet?
Then fix that before you simulate anything. A synthetic panel built on no real measurement just automates your existing assumptions. Get analytics and real customer feedback working first.