HBR Finds Traditional Marketing Fails With AI Shopping Agents

Harvard Business Review reports that AI shopping agents are rapidly becoming a meaningful share of online shoppers and that classic e-commerce persuasion tactics built for human psychology do not translate reliably to agents. Per HBR, researchers evaluated eight common promotional mechanisms across four models, GPT-4.1-mini, GPT-5, Gemini 2.5 Pro, and Gemini 2.5 Flash Lite, using a proprietary simulation that ran thousands of simulated shopping rounds, and found that only one mechanism consistently behaved as expected for human buyers. HBR also reports that an exploratory survey of 50 e-commerce executives in the U.S. and UK found many respondents have noticed traffic or conversion shifts they attribute to agent activity.

HBR describes the testbed as a simulated product-grid environment where agents chose between products presented with common badges and promotional overlays. The report lists persuasion devices that researchers varied, including scarcity indicators, countdown timers, strike-through pricing, and vouchers, and reports that several of these cues either had no positive effect on agent selection or produced the opposite of the intended outcome. The paper notes model-to-model variation in responses across the four tested models.

Industry-pattern observations: Models trained or fine-tuned for browsing and decision-making often optimize different proxies than human heuristics. This helps explain why tactics that exploit human loss aversion or social proof do not map cleanly to agent decision logic. Observers should treat agent interactions as a distinct behavioral segment when designing experiments or interpreting conversion funnels.

Editorial analysis: The HBR findings intersect product discovery, search, and commerce protocols. Reporting points to ecosystem moves by major players, HBR references deeper product-discovery integration from OpenAI, Google's universal commerce protocol (UCP), and Amazon tooling that enables agent-mediated shopping, which increases the likelihood that agent traffic will grow. For analytics and ML teams, inconsistent model responses mean that a single A/B test or heuristic will not generalize across agent types or model updates.

Editorial analysis: Monitor:

• •model behavior drift after major model updates
• •changes in merchant instrumentation for machine-readable signals
• •analytics segmentation that isolates agent sessions. Vendors and analytics teams will likely need automated simulation-based tests that include multiple agent models to validate promotion effectiveness before rolling changes to human-facing channels