The Buyer's Guide to Synthetic Research · Article 2
A three-part series for research and insight buyers · 2025
Fit for which purpose? A field map
Synthetic solutions aren't good or bad. They're fit or unfit, and fitness depends entirely on the job. A working map for buyers.
Synthetic solutions aren't good or bad. They're fit or unfit, and fitness depends entirely on the job. A working map for buyers.
Ask a bad supplier what synthetic data is for and you'll get one word: everything. Ask a good one and you'll get a question back: what's the decision this research needs to support? Here is the map we'd draw on the whiteboard.
Where synthetic earns its keep today
Sample augmentation. You ran a robust study, but the subgroup you most care about came back at n=75. Statistical synthetic augmentation, modelled on your real data, can extend that base for analysis. Kantar, hardly a hype merchant, frames synthetic data as exactly this: bounded augmentation of human research, not a substitute for it.
Privacy protection. Sharing sensitive datasets across borders or with third parties is a compliance headache. Synthetic versions preserve analytical value while breaking the link to real individuals. The use case regulators actively encourage.
Early exploration and pre-testing. Before spending six figures on fieldwork, synthetic personas can pressure-test your hypotheses, expose a leading question, or kill a concept that was never going to fly. Ipsos has gone as far as partnering with Stanford University to build validated "digital twins", a signal of where the serious end of the industry thinks this is going.
Hard-to-reach audiences at discovery stage. Simulating a first conversation with an audience you can barely recruit is cheaper than pretending you'll recruit them properly next quarter.
Where it isn't fit (yet)
Tracking and benchmarks: synthetic trend data has no anchor in real-world change; it moves when the model moves. Decision-grade evidence for major bets (launches, pricing, M&A) where being confidently wrong is catastrophic. Emotional and behavioural insight, because language models simulate how people write, not how they feel or behave. And rare events, which models systematically smooth away, often precisely the signals worth paying for.
A worked example
Consider a retail bank planning a proposition test with an over-65 vulnerable-customer segment. Recruiting that audience properly is slow, expensive, and ethically delicate. A sensible synthetic play: use LLM personas to pre-test the stimulus and discussion guide in week one, killing the two confusing concepts early; run real fieldwork with a smaller, better-targeted sample in weeks three to six; then apply statistical augmentation to the real data to enable subgroup cuts the budget couldn't otherwise support. Synthetic at the edges, humans at the centre, disclosure throughout. That's what fit for purpose looks like in practice.
The buyer's rule of thumb
Synthetic data is currently an accelerant and an extender, not a replacement. Use it to explore faster, augment defensibly, and protect privacy. Insist on real humans wherever the decision is expensive to reverse. And require every synthetic element to be labelled in every deliverable. If a supplier resists that, you've learned everything you need to know. Fitness for purpose isn't a verdict you deliver once; it's a discipline you apply per project, per decision, per dataset. The final article gives you the ten questions that make this enforceable.