Marketing to Machines: How AI Shopping Agents Respond to Promotional Cues

Abstract

As AI agents increasingly make purchasing decisions on behalf of consumers, do standard promotional cues designed for human cognition retain their influence? We examine whether common promotional cues designed to influence human shoppers affect AI agents in comparable ways. We develop a controlled simulation environment that replicates common e-commerce promotional badges-including assurance signals, scarcity, strike-through pricing, countdown timers, social proof, vouchers, bundles, and ratingsand conduct repeated simulated choice experiments using four frontier AI models across four product categories. The results reveal pronounced heterogeneity both across models and across product categories within the same model. Among the badges tested, only ratings consistently increased selection probabilities across models in a manner aligned with established human reliance on quality signals. All other promotional effects varied by model and product category. In an exploratory follow-up, we examined whether these patterns could be explained by human behavioural mechanisms or instead reflected model-specific computational responses. Constructs derived from prospect theory, including loss aversion, probability weighting, and diminishing sensitivity, provided little systematic explanatory power. These findings provide controlled behavioural evidence on promotional effectiveness in agent-mediated shopping environments, show that AI agents cannot be treated as a homogeneous class of consumers, and suggest that marketing-to-machines cannot be inferred directly from established human behavioural theory. The study offers practical implications for platforms, merchants, and AI developers as algorithmic intermediaries play an expanding role in online purchasing decisions.