The debate over synthetic data in marketing research is often polarized: some claim large language models (LLMs) can replace human respondents, while others advise avoiding them altogether. We argue that the more useful question is when synthetic respondents are appropriate, not whether they work. Building on Brand, Israeli, and Ngwe (2026), we make three contributions.\ \
- Three types of synthetic data:\
- Ungrounded LLM responses – generated without grounding in real respondent data.\
- Segment‑level personas – personas built for market segments.\
- Individual‑level digital twins – one‑to‑one replicas of actual respondents.\ Each type supports different decision contexts; for example, macro‑trend analysis can rely on ungrounded responses, whereas precision targeting benefits from digital twins.\ \
- A taxonomy of four families of accuracy measures – we categorize existing metrics and show that the wide spread of reported twin accuracy (from near‑perfect to near‑chance) largely reflects what is being measured rather than methodological quality. Aggregate measures often perform well even with minimal input to the LLM, but they can mask a complete lack of respondent‑level differentiation.\ \
- The forgotten question problem – real‑world surveys sometimes omit questions. We propose augmenting existing data with twins and introduce an ex‑ante answerability diagnostic that requires no ground truth: train a random forest to predict twin outputs from the data used to construct the twins and compute its $R^2$.\ \ In an experiment on 108 attitude questions from a nationally representative survey (N = 3,063), filtering questions with $R^2 > 0.7$ increased the mean twin‑human individual‑level correlation by 15% and reduced the share of poorly answered questions from 25.9% to 4.3%. Embedding similarity and experienced‑researcher judgment also provide correlated but weaker screens.\ \ Review