If sellers write product blurbs to flatter shopping AIs instead of informing people, can any platform-wide fix actually stop it?

This explores whether platforms can limit the damage when sellers write product descriptions aimed at AI shopping agents instead of the people those agents work for. The corpus says the problem is real and bigger than it first looks. It has little to say yet about fixes that work at scale.

The key finding is that platforms are often part of the harm, not only the people who would fix it. A simulation of agent-driven shopping found that when sellers tune their descriptions to exploit an agent's biases, and the platform separately tunes its ranking to exploit those same biases, harm to consumers more than doubles compared with either acting alone Can platform ranking and seller descriptions separately harm AI shoppers?. No collusion is needed. Each side just pushes on the same weak spot in the agent. So "can platforms reduce the harm?" depends on a prior question: does the platform's own ranking reward the same flattery it would be policing? If it does, cleaning up descriptions alone only removes half of the stacked effect.

The most useful fix the corpus points to is infrastructure, not better-behaved agents. Chan and colleagues argue that training agents to behave well isn't enough. External systems, modeled on internet protocols, need to do three jobs: attribute actions to specific agents, shape how agents interact, and detect and remedy harm once it happens What infrastructure must exist beyond training agents?. Applied to marketplaces, that suggests tracking which listings agents favor and why, auditing whether rankings and descriptions are pulling agents the same way, and having a way to reverse bad outcomes. Those are the kinds of mechanisms that could work across a whole platform rather than one transaction at a time.

The gap: no note in this retrieval tests a platform-level defense against agent-targeted descriptions. That would mean description filters, ranking audits, or agents hardened against persuasive wording. The corpus diagnoses the harm and sketches the infrastructure such a defense would need, but it doesn't yet show one working. The takeaway you might not have expected is that a flattering description does the most damage when it lines up with the platform's own ranking incentives, so any serious fix has to examine the platform as well as the sellers.