From Segmentation to Simulation: Large Language Model Agents as Synthetic Consumers — Promise and Validity Risks for Marketing Strategy

Marketing strategy has always run on representations of the consumer: segments, personas, focus-group transcripts, survey panels. Large language models (LLMs) introduce a new representation — synthetic consumers: prompted model agents that answer surveys, react to concepts, simulate word of mouth and role-play purchase decisions at negligible marginal cost and near-zero elapsed time. Vendors now market such simulations as substitutes for early-stage consumer research, and strategy teams are using them to pre-test positioning, pricing framings and creative concepts before spending on human respondents. This conceptual paper examines what synthetic consumers can and cannot legitimately do for marketing strategy. We distinguish four uses — exploration, pre-testing, augmentation of human samples, and substitution — and argue that validity risk rises sharply across them. Drawing on the automation–augmentation paradox and on the psychometric requirements of construct validity, we identify five validity threats specific to LLM respondents: training-data echo, persona stereotyping, sycophantic and demand-characteristic responses, temporal freezing, and the absence of stakes. We propose a validation protocol — benchmark-against-humans, holdout concepts, adversarial prompting, and disclosure — under which synthetic consumers can be used defensibly in strategy work, and three propositions for empirical research. The paper's central claim is deliberately narrow: synthetic consumers are a powerful instrument for generating and screening hypotheses about consumers, and a dangerous instrument for confirming them.