Agentic Shopping Gets Manipulated Just Like Human Buying

AI agents acting as shoppers succumb to marketing

Ethan Mollick highlighted new research focused on "agentic shopping": a scenario where autonomous AI agents make online purchases on behalf of users. The main question of the study was whether the behavior of these agents can be consistently predicted or even targeted and influenced by classic marketing techniques.

Researchers found that the preferences of agents are surprisingly fragile and vulnerable. Small changes in the environment were able to significantly shift the AI's decision-making.

A new front is emerging for marketers

The study shows that the behavior of AI assistants when selecting products is not a purely objective and rational calculation. Things like the order in which the agent views pages, or the way its long-term memory is configured, led to unpredictable changes in its final selection.

This means that Search Engine Optimization (SEO) and Conversion Rate Optimization (CRO) will soon gain a new layer: optimization for agents. Whoever figures out how to structure data to influence the agent's "thinking" while browsing the web will gain a massive advantage.

The fragility of autonomous assistants' convictions

The most interesting finding is precisely that even minor noise in the context can derail the logic of a language model. If the order of viewed items (priming and context window) affects the purchase, it means that agents suffer from cognitive biases similar to humans. The only difference is that their "biases" stem from the technical properties of attention mechanisms in the transformer architecture.

Who will win the war for agents' attention

Success in e-commerce in the coming years may not depend on how you convince a human customer. The decisive factor will be what data, and in what order, you slip to their purchasing agent. The emergence of new agencies specializing purely in manipulating the context window of shopping AIs will be the clearest signal.

Lilith's verdict

The problem with agents isn't that they're too smart. It's that the attention architecture makes them just as easily manipulated as humans in a discount aisle.