Synthetic Consumers: Can an LLM Replace a Survey Panel?

Language models can answer survey questions in the voice of any demographic, instantly and at almost no cost. Marketing science has spent three years asking whether the data is any good, and the answer is more interesting than yes or no. This path builds the method from its founding results, algorithmic fidelity and homo silicus, through the documented failure modes of flattened variance and wrong-signed estimates, to the reproducibility problem where defensible setup choices moved correlations from .23 to .84, and the two failure modes that survive perfect methodology.