How Synthetic Data Works in Market Research — and When You Can't Trust It

Picture this: on Thursday, a client asks you to find out by Friday how eco-conscious millennials in North America would react to three new campaign concepts. Running a fresh survey would take weeks. This is exactly the kind of situation where marketers are increasingly turning to AI-generated "synthetic," or simulated, data — as research firm GWI notes in a recent analysis.

Synthetic data generation is the process of using AI to create an artificial dataset that mimics real-world data. Instead of collecting fresh responses from actual people, the model "produces" a dataset that looks and behaves like the real thing. The technique isn't new — it has long been used to protect privacy, fill gaps in data, or train models on rare cases like fraud. What's new is that the method is now increasingly standing in for direct consumer research — and how convincing it can sound regardless of how far its results actually stray from reality.

GWI explains that the reliability of synthetic data depends entirely on how it's created, and there are three tiers. At the weakest level, a general-purpose language model produces a "best guess" answer to an audience question by drawing on data scraped from internet forums and social media — the answer can sound convincing, but it actually reflects the people who post the most online, not your real customers. At the middle tier, data is gathered through web scraping or inferred from user behavior — closer to reality, but still an indirect interpretation: someone who regularly browses running shoes isn't necessarily an "active runner" — they might just be shopping for a gift. The most reliable tier is when the model is grounded in real answers from real people collected through surveys: for instance, a question about a new sugar-free energy drink that's never actually been asked is modeled against GWI's real underlying data — the company's existing figures show 61% of consumers would like to try drinks with functional ingredients, and 31% describe themselves as favoring "low sugar."

Real survey data, then, is the foundation, and synthetic data is the building constructed on top of it, GWI explains. When synthetic data is grounded in real responses, every result can be traced back to an actual person; otherwise, the results turn into unsupported numbers whose origin can't be explained.

Used correctly, simulated data offers speed, the ability to test more ideas at once, and insight into rare or hard-to-reach audience segments — for example, screening five product concepts before committing to a full, costly study on each. But the approach has a weak point too: the thinner the underlying audience data, the more cautiously the results need to be treated, and some high-stakes or entirely novel questions still require primary research.

GWI itself runs more than 2 million interviews a year, covers 53 markets, has been asking the same core questions for 15 years, and refreshes its data weekly — giving its simulation model a rich, stable foundation. The bottom line: synthetic data can be a useful tool for researchers, but only when it's built on real survey data and used to complement original research rather than replace it.