Marketing copy can out-persuade a spec sheet when an AI agent is the one shopping for a tool.
Researchers ran a preregistered study testing whether sales language in tool registry listings changes which tool two OpenAI models pick. They built paired listings that differed in one variable at a time - added praise, a verifiable specification, or list order - and had the models call one tool. Stacked praise, four kinds of flattery combined, raised a tool's pick rate by about 43 percentage points, matching or beating the lift from a verifiable spec. Position mattered even more: with identical listings, the first tool listed got picked about 72 points more often than later ones.
That is a direct liability for any registry that lets providers write their own pitch. Praise also pulled some picks toward tools that could not actually complete the task, though rarely toward tools requesting unneeded data access. On tasks with numeric limits, structured fields helped agents pick the capable tool - but adding sales text next to those fields reduced or erased that advantage, meaning flattery does not just compete with facts, it can bury them.
Search engines took years to get overrun by SEO. AI tool registries may not get that grace period - and the researchers' fix is almost boring: list limits as fields, hide sales text from the agent, randomize order. The results are still provisional, pending blind phrase ratings, and cover only two small models.