Distinguishes three types of synthetic data in marketing research and shifts focus from whether LLMs work to when they are appropriate, extending prior work.
The paper Synthetic Data in Marketing Research: How to Evaluate and When to Trust by Oded Netzer and Rajan Sambandam distinguishes three types of synthetic data: ungrounded LLM responses, segment-level personas, and individual-level digital twins. It argues that the critical question is not whether synthetic respondents work, but when they are appropriate for specific decisions. The authors note that reported accuracy varies widely because aggregate measures can mask a lack of respondent-level differentiation.
Summary
This paper addresses a growing practical problem in marketing research: the increasing use of synthetic data, often generated with LLMs, without a clear framework for deciding how much weight such data should receive. Rather than treating synthetic data as a simple substitute for human-collected data, the paper distinguishes three functionally different roles that synthetic data can play—such as stand-ins for observed consumer behavior, probes or simulations of latent consumer states, and diagnostic or stress-test material for models and strategies. This distinction matters because the appropriate evidentiary standard depends on the role: synthetic data used to estimate a market quantity require different validation than synthetic personas used to generate hypotheses or synthetic counterfactuals used to test model robustness.
The key contribution is a shift from asking whether LLMs can produce plausible marketing data to asking when such data are appropriate for a given decision. The paper argues that trust in synthetic data should be conditional on factors such as construct validity, distributional fidelity, stability across prompts or model versions, sensitivity to specification choices, and alignment between the synthetic data-generating process and the intended decision task. It also emphasizes provenance, documentation, and evaluation procedures, framing synthetic data as a tool whose usefulness depends on transparent use rather than an unqualified replacement for empirical evidence. This reframes prior work on LLM-based marketing research from model-centric benchmarking toward a governance- and decision-oriented framework.
The paper matters because it gives researchers and practitioners a more disciplined way to integrate synthetic data into marketing analytics. It helps separate legitimate uses—exploration, augmentation, robustness checking, and hypothesis generation—from overclaiming, especially in high-stakes settings such as forecasting, segmentation, creative testing, or strategy selection. For technically literate audiences, the value lies in providing a language and set of criteria for auditing synthetic data pipelines, calibrating confidence in LLM-generated outputs, and preventing the mistaken assumption that plausible-looking synthetic data automatically carry the same inferential weight as real consumer data.