Marketing to Machines — AI models & e-commerce promotional cues

Summary:

This working paper looks at a simple question: if an AI shopping agent is choosing what to buy, do the usual online sales tricks still work?

The authors tested common e-commerce badges like low-stock warnings, crossed-out prices, countdown timers, vouchers, bundles, review counts, and star ratings to see how AI models react.

It was written by Jafar Sabbah and Oguz A. Acar from City St George’s, University of London and King’s College London.

They built their own controlled shopping simulator called Agentyx.

In each test, an AI model saw a fake store page with 8 products on screen and had to pick one. The team tested 4 AI models across 4 product categories, with about 1,000 repeated choice tasks per product per model, for roughly 16,000 choices total.

They randomly changed product position, price, and which promo badge appeared, then used statistical models to estimate which cues raised or lowered the chance of being picked. They also ran a follow-up test to see whether ideas from human behavioural theory, especially prospect theory, could explain the patterns.

The biggest result is that AI models did not act like one single kind of shopper. Responses varied a lot by model, and also by product category within the same model.

• Star ratings were the only cue that consistently helped across all models and all categories.
• Social proof was the next most reliable cue, but not universal.
• Price mattered a lot too: all models were less likely to choose higher-priced items.
• Most models also liked the top-left position on the page, though GPT-5 did not show that same pattern.
• Beyond that, the usual promo badges were hit and miss. Assurance badges helped often, but not always.
• Scarcity, strike-through pricing, timers, vouchers, and bundles worked for some models and some products, but failed for others.
• In some cases the effect even flipped, like bundling hurting GPT-5 in the phone category.
• When the authors tested stronger versions of the badges, like deeper discounts or shorter timers, those intensity changes usually did not matter much.

The clearest dose effect was for ratings: higher star scores often helped, and review counts sometimes helped too.

The follow-up theory test found little evidence that these patterns can be explained by the same human-style mechanisms marketers often rely on. The prospect-theory-based interactions were mostly weak or not robust after multiple-testing correction. That means marketers should be careful about assuming AI agents think like people just because both are making choices.

This is a controlled simulation, not a live shopping platform, so it strips away many real-world details like brands, detailed specs, recommendation systems, and multi-step shopping workflows.

The storefront was deliberately simple, with a fixed grid and generic products, which helps isolate effects but may limit real-world fit. The authors used a proprietary simulator they built themselves, so replication depends on the detail they disclosed.

The paper is also a working paper, not peer-reviewed journal publication. Still, the method is clearly described, the sample of simulated choices is large.