The Buyer's Guide to Synthetic Research · Article 1
A three-part series for research and insight buyers · 2025
Synthetic data: permission to be sceptical
The research industry has a new miracle ingredient. Your scepticism is not a bug. It's the most useful tool you own. Here's how to point it in the right direction.
The research industry has a new miracle ingredient. Your scepticism is not a bug. It's the most useful tool you own. Here's how to point it in the right direction.
Every few years, market research announces its own reinvention. Online panels were going to ruin everything, then became everything. Mobile-first was a fad until it was table stakes. Now it's synthetic data (artificially generated respondents and datasets), and the industry has split, with theatrical predictability, into evangelists and doom-mongers.
If you buy research, both camps are wasting your time. The evangelists are selling you something. The doom-mongers are protecting something. Neither position helps you answer the only question that matters: is this fit for my purpose?
What synthetic data actually is
Two quite different things travel under one name, which is where most confusion starts. The first is statistical synthetic data: artificial records generated from a model trained on real respondent data, preserving its patterns without copying individuals. Think of it as a statistically faithful extension of data you already trust. The second is LLM-generated synthetic respondents: AI personas that simulate how a defined type of person might answer questions, built with no real fieldwork at all.
The first is an established statistical technique with decades of lineage. The UK Information Commissioner's Office lists synthetic data among its recommended privacy-enhancing technologies. The second is genuinely new, genuinely interesting, and genuinely unproven for most decision-grade uses. A supplier who doesn't immediately distinguish between the two is either confused or hoping you are.
Why your nervousness is rational
The honest evidence base is mixed. A Harvard Business School working paper found that LLM-derived estimates of consumer willingness-to-pay were sometimes realistic, and sometimes flatly wrong-signed, which is the polite academic way of saying the machine confidently pointed in the wrong direction. Stanford's widely cited generative agent study achieved 85% accuracy replicating individuals' survey responses. Impressive, until you notice it required a two-hour interview with each real human first. The best synthetic results, in other words, are built on a foundation of real ones.
So be nervous. But be precisely nervous. The risk isn't that synthetic data exists; it's that it gets used for the wrong jobs, invisibly, without your knowledge or consent. Which brings us to the real buyer's agenda: knowing which jobs are which, and forcing disclosure. That's the subject of the next two articles in this series.
What to do this quarter
Three practical moves cost you almost nothing. First, add one line to every research contract and RFP: all synthetic or AI-generated elements must be disclosed and labelled in deliverables. Regulation is heading this way anyway, with the EU AI Act's transparency obligations for AI-generated content already in force for many uses. Second, ask your incumbent agencies what their synthetic data policy actually is; the quality of the answer is itself a finding. Third, nominate one person on your team to get literate in the vocabulary (seed data, calibration, validation), because the suppliers already have.
The uncomfortable truth for buyers who'd rather sit this out: your competitors' insight teams are learning to buy synthetic well (faster exploration, cheaper pre-testing, bigger effective samples). The opportunity cost of blanket refusal is now real and growing each quarter. Scepticism is a compass, not a bunker. Point it well and it will steer you to the small set of synthetic applications that are genuinely fit for your purpose, and away, calmly and defensibly, from everything else.