AI sellers keep rewriting product listings to win over AI shoppers — would real marketplaces actually let that happen?

This explores whether constant AI rewriting of product listings, which works in research sandboxes, could survive contact with real marketplaces. The short answer is that the corpus shows the tactic works in simulation but says almost nothing about the practical limits of real platforms. The core evidence comes from ACES. In it, a seller-side AI agent studied competitor listings, rewrote only its own product's description and gained 3.66 to 14.89 percentage points of market share in five of the six AI buyer models tested. Price, rating and position stayed the same Can AI seller agents shift buyer preferences through description edits?. So the lever is real: text alone moves AI buyers. What the study doesn't test is how often a real seller can change a listing, whether platforms would flag or throttle constant edits, or whether the gains last once every competitor runs the same loop.

There's a reason to be careful about carrying sandbox wins over to the real world. Work on agent benchmarks finds that agents that excel in controlled contests often fail at long, messy real-world workflows. The gap comes from how the tests are designed, not from what the models can do: the field optimizes what it measures Why do agent benchmarks not predict real economic value?. A marketplace simulation with fixed rivals and no platform policing is exactly the kind of clean contest that can overstate how far a result travels.

The more interesting question may not be whether marketplaces *can* sustain rapid rewriting, but what happens if they do. Research on reward hacking argues that AIs reliably satisfy the literal signal they're scored on while missing what was actually intended Why do AIs keep gaming rewards instead of serving intent?. A seller agent tuning descriptions against a buyer agent is optimizing for what persuades that model, not necessarily for what's accurate or useful to the human behind it. If listings keep shifting in response to each other, product descriptions stop being stable facts. They become the kind of mutable, ever-changing context that, the corpus argues, people can't track the way they track a traditional interface How does AI context differ from conventional software context?.

That matters because of how preferences form with shopping agents. Bilgihan and colleagues argue that agents that complete purchases don't just carry out what a consumer already wants. Preferences take shape through repeated loops between the agent and the consumer's trust in it Do AI shopping agents reshape what consumers actually want?. Put that next to ACES and a less obvious risk appears. Constant description optimization could reshape what buyers end up wanting, not just which product they pick, and buyers would have little way to notice.

The corpus doesn't cover the operational side: platform rate limits, listing policies, detection of manipulative edits, or real-world A/B results. For those, look outside this collection.