In Silico Advertising Tests

Abstract

This research examines whether large language model agents can produce analytically useful evidence for advertising testing when treated as synthetic respondents rather than as substitutes for human consumers. We report an in-silico comparative post-test of two television advertisements in the French telecommunications market, conducted on a panel of 800 profile-conditioned consumer agents. The study evaluates the synthetic corpus through internal validity criteria: whether the questionnaire covers the main dimensions of advertising response, whether evaluative dimensions form coherent relations, whether the two stimuli are discriminated across preference labels, scores, lexical patterns and semantic embeddings, and whether responses vary intelligibly across social profiles.The results indicate that the synthetic panel produces structured and interpretable evidence. The contribution is not that the analysis discovers a contrast invisible in the advertisements themselves: Bouygues and Orange are clearly built around different persuasive registers. Rather, the contribution is methodological. The synthetic-interview pipeline consistently recovers and differentiates these registers across preference labels, scores, lexical patterns and semantic comparisons. Bouygues is associated mainly with creative, humorous and memorable qualities,while Orange is associated with reliability, informativeness, purchase orientation and overall preference. This structure is visible quantitatively in the main six dimensional coding scheme. Among all 800 synthetic agents, Bouygues leads on creativity (83.6% versus 14.8% for Orange), whereas Orange leads on informationseeking or consideration (76.8% versus 14.8% for Bouygues) and overall preference (54.8% versus 20.6% for Bouygues), with the remaining responses classified as mixed or neutral. The analysis remains deliberately limited to synthetic respondents.Claims about similarity to human preferences or market behaviour would require external comparison with representative human data or market-level indicators.