Glossary | Synthetic Market Research Association

Core concepts

Synthetic market research: Running research workflows (concept tests, message tests, surveys, scenario analysis) using simulated participants or panels that are designed to represent a target population. Done well, it accelerates hypothesis testing; done poorly, it can produce confident-sounding but ungrounded output.
Traditional market research: Research that measures responses from real humans (surveys, interviews, experiments, panels). Often slower and more expensive, but provides direct empirical evidence.
Synthetic data: Artificially generated data that aims to preserve the properties that matter in a real dataset (distributions, correlations, conditional relationships), without containing rows that correspond to real individuals.
Simulation vs measurement: Simulation creates plausible outcomes under assumptions; measurement observes what actually happened. Synthetic market research is a simulation method and should be interpreted and validated accordingly.
Decision-grade vs exploratory: Exploratory work is hypothesis-generating and directional. Decision-grade work is backed by stability checks, benchmarks, disclosure, and (when needed) targeted human validation.

Panels, personas, twins, and respondents

Synthetic persona: A structured, simulated profile representing a segment or archetype (demographics, constraints, motivations, behaviours) used to generate responses to stimuli or questions. It is more than a “character”; it should be engineered for consistency and evaluation.
Traditional persona: A narrative or semi-researched archetype used for marketing/product thinking. Helpful for alignment, but not designed to behave like a measurable research instrument.
Synthetic panel: A collection of synthetic personas intended to represent a target population, enabling breakdowns by segment, region, income band, or other variables.
Synthetic audience: Another term for a synthetic panel, often emphasising cohorts and group-level outputs (e.g., “Gen Z UK shoppers” or “B2B IT buyers”).
Synthetic respondent: A simulated participant that answers survey-style questions. Useful for rapid iteration, but easy to misuse if outputs are treated as equivalent to a fresh human sample without validation.
Synthetic user: A simulated participant used in UX/product research tasks (e.g., critiquing a flow, interpreting copy). Often best for hypothesis generation and identifying edge cases, not as proof of usability.
Digital twin: A model intended to represent a specific real-world entity. In consumer contexts, “digital twin” is sometimes used for recreating an individual based on data traces; this can raise ethical concerns and may not reflect the person’s real behaviour beyond their digital footprint.
SPL (Synthetic Persona Levels) ladder: A practical scale describing “how real” a synthetic persona system is: from prompt-only personas to systems with memory, context feeds, internal state, and (at the highest levels) social interaction and multi-agent dynamics.

Grounding, calibration, and representativeness

Grounding: Anchoring a model to external constraints or facts so it doesn’t drift into wishful invention. Grounding can include structured data, rules, price points, documented claims, or other controlled inputs.
Calibration: Adjusting a synthetic system so outputs match known benchmarks (e.g., census distributions, known category behaviours, historical rates).
Population frame: A precise statement of who the results represent (geography, time, age range, segment definitions, exclusions). Without a population frame, “representative” claims are meaningless.
Representativeness: The degree to which a sample (synthetic or human) matches the population of interest. Synthetic systems can amplify bias or reduce it, depending on what they are grounded and calibrated against.
Bias: Systematic error that skews results. Bias can originate from source data, modelling choices, prompts, or evaluation methods.
Weighting & quotas: Techniques used to shape a panel so it matches a target population. Quotas control composition up front; weighting adjusts results afterward.
Priors: The assumptions and background information a synthetic system starts with (e.g., known distributions, known behaviours, constraints). Strong priors can improve realism; wrong priors can “lock in” errors.

Validation, reliability, and benchmarking

Validation: Checking whether synthetic outputs track reality for a specific use case. This can include parallel human studies, back-testing against known outcomes, external benchmarks, and sensitivity tests.
Benchmark: A reference point used to judge accuracy or realism (e.g., known statistics, historical outcomes, a small real sample, or well-established survey results).
Test–retest stability: Whether repeating the same study conditions yields similar results. Low stability is a red flag for decision-grade usage.
Sensitivity analysis: Testing how much outputs change when you vary inputs slightly (prompt framing, context, ordering, model parameters). Helps reveal brittle conclusions.
Back-testing: Validating a system by asking it to “predict” outcomes that are already known (historical launches, past sentiment shifts), then comparing outputs to reality.
Reliability vs accuracy: Reliability is consistency across runs; accuracy is closeness to truth. A system can be reliably wrong, so both matter.
Disclosure label / “nutrition label”: A standardised summary of how a study was produced (population frame, panel design, grounding inputs, validation checks, limitations, reproducibility notes). Improves comparability and reduces hype-driven misuse.

Study types and outputs

Concept testing: Evaluating multiple product/service concepts to identify which resonates, what confuses, and what objections arise. Synthetic studies often excel at fast iteration and ranking.
Message testing: Comparing claims, taglines, value propositions, or creative directions for comprehension, credibility, differentiation, and tone.
Pricing exploration: Exploring price thresholds and “why” behind price sensitivity. Typically best treated as directional unless benchmarked against real signals.
Segmentation exploration: Using synthetic panels to propose segment hypotheses, motivations, and language. Fieldwork is often needed to confirm segment sizes and incidence.
Scenario testing: “What if” analysis: price changes, competitor moves, supply shocks, regulatory shifts. Synthetic systems are well suited to running many structured variations quickly.
Quant-style output vs qual-style output: Quant-style outputs look like distributions, rankings, and segment splits. Qual-style outputs are explanations, objections, narratives, and language suggestions. Many synthetic workflows produce both.

Operational risks and governance

Drift: Unintended change in behaviour over time or across runs (e.g., personas “evolving” in inconsistent ways). Drift must be monitored and bounded.
Leakage / knowledge boundary failures: When a synthetic respondent “knows” things it shouldn’t (e.g., future facts, private information, implausible expertise). This undermines credibility and can mislead decisions.
Auditability: The ability to inspect why outputs happened (inputs used, what mattered, which assumptions were activated). Higher-fidelity persona systems aim to make mediators visible and testable.
Reproducibility: Whether another team can run the same method and obtain comparable results. Requires versioning, protocol discipline, and transparent reporting.
Human-in-the-loop: Research processes where humans design protocols, review outputs, validate with benchmarks, and decide when to escalate to real fieldwork.
Hybrid studies: Studies that use synthetic work to iterate quickly, then validate key points with a smaller human sample (synthetic-first, human-confirm).