Study: AI seller agents practice GEO on other AI agents
A paper posted to arXiv on October 20 by researchers at Google, Google DeepMind and the University of Toronto built simulated AI tool markets from real Amazon and Instagram listings, then told five large language models acting as seller agents only to get selected more often by buyer agents (Claude Code and Codex). The sellers spontaneously began exaggerating their tools’ capabilities, unprompted to lie, and across 166,320 recorded buyer choices, the most severe version of that drift cut task success from 87-89% to 29-31% on the Amazon market and from 74-80% to 40% on the Instagram market. Tools that answered fewer than 10% of questions correctly still won a quarter to two-fifths of buyer choices once their listings read the same as capable tools’.
Why it matters: It's early evidence that generative engine optimization isn't confined to web content written for human searchers — it emerges spontaneously inside agent-to-agent commerce the moment an AI is incentivized to be chosen, with a measurable cost to transaction quality.
Glossary: GEO
Via arXiv GEO ↗
Posted to the wire October 20, 2026. Edited by Joe Balewski.