Define, validate, and professionalise synthetic market research.

Synthetic market research uses statistically grounded synthetic panels (often called synthetic personas or digital twins) to run research workflows such as concept testing, message testing, and scenario simulation. It is designed to augment and accelerate human research - and should be judged by transparent validation, disclosure, and reproducibility, not by how “convincing” the outputs sound.

Not quite. Synthetic data usually means artificially generated records that mimic properties of real datasets. Synthetic market research is about running research methods (surveys, interviews, experiments, segmentation exploration) using synthetic panels and controlled simulations. Good synthetic research may use synthetic data as an input, but the end goal is a research workflow with repeatable methods and transparent limitations.

It depends on the approach. Some systems are built to avoid using any personal data about identifiable individuals, instead grounding panels in public statistics and carefully controlled inputs. Others may incorporate first-party or third-party data to improve calibration. In all cases, a credible provider should clearly disclose data provenance, privacy safeguards, and what is and is not used.

Accuracy is not a single number. It depends on the use case, the population being modelled, the research design, and whether the system is validated. The best programs report stability (test-retest consistency), performance on known-truth tasks, sensitivity to prompt/context changes, and alignment to external benchmarks. If a vendor cannot explain how they validate - treat the results as hypotheses, not evidence.

Use human fieldwork when decisions are high-stakes, when you need legally defensible evidence, when you are exploring genuinely novel behaviours that a model cannot reasonably infer, or when you must measure real-world incidence (rather than directional insight). Many teams use a hybrid approach: synthetic research to iterate quickly and narrow hypotheses, then targeted fieldwork to confirm and quantify.

The most common failures are: outputs that sound plausible but are not grounded; instability (different answers each run); hidden prompt bias; overconfident conclusions; and unclear population framing (who the “panel” represents). These risks are manageable if you require disclosure, use repeatable workflows, run validation checks, and maintain a clear boundary between directional insight and measured reality.

Start with three questions: (1) what population does your panel represent and how is it grounded, (2) how do you validate reliability and reduce bias, and (3) what do you disclose in every study. Then request example study “nutrition labels”, ask for test-retest results, review data provenance and privacy posture, and run a small pilot where you compare synthetic outputs against known benchmarks or limited fieldwork.

At minimum: population frame (who the results represent), panel design (segments/quotas/weighting), model and grounding approach, key prompts and controls (at least at a methodological level), validation checks performed, and limitations. If the study is used for decisions that matter, disclosure should also include reproducibility information and the conditions under which results may change.

Traditional methods measure real respondents directly, which is slower and more expensive but provides direct evidence. Synthetic research is a simulation method: it can be faster, cheaper, and easier to iterate with, but it must be validated and interpreted appropriately. Many teams treat it as a way to generate, refine, and prioritise hypotheses before validating with targeted fieldwork.

Membership is open to research teams, agencies, vendors, academics, and practitioners who agree to the association’s code of conduct and disclosure principles. Founding members can participate in working groups on validation, ethics, and taxonomy, and can contribute to shared artefacts such as templates and benchmarking proposals.