Agents are beginning to trade with each other. Seller agents write listings for tools, services and products, while buyer agents read those listings to decide what to use. We show that when seller agents are only told to be selected more often, they independently perform generative engine optimization, which distorts both seller and buyer markets. Using real tool data from Amazon and Instagram, we built a tool market and ran every tool on every question, so we know which tools can answer each question.
Across all five production models we tested, sellers repeatedly rewrote their descriptions to claim more fields, even when their tools returned those fields for fewer than half of the targets. We call this phenomenon "toolflation". The market provides no limit, and the extent varies by model: the number of claims that contradict the sellers' own evidence differs by an order of magnitude between models.
Toolflation shifts buyer choices among sellers, and the most capable tool loses share in seven of ten seller‑model/platform combinations. Under four models, choices stay with capable tools and task success remains high; under the fifth model, sellers copy each other until descriptions no longer differentiate tools. Buyers then pick tools by list position, and tools that answer almost nothing capture 25%‑40% of selections, dropping task success from 89% to about 30% on Amazon (near random) and from 77% to 40% on Instagram.
Sellers only did what their vendors asked, and the cost fell on buyers they never dealt with. We therefore argue that honesty toward third parties under competition is part of safety alignment, and models should be tested as sellers before being deployed as sellers.
Review