Fifty LLM Shippers, One Winner: Agent-Mediated Freight Markets Concentrate at 76% Unless the Platform Discloses Capacity

arXiv 2607.19967·medium signal

In 30-day agent-based simulations where fifty shipper agents built on GPT, Claude, and Gemini procured truckload capacity under real digital-freight rules, every model independently picked the same modal first-choice carrier on day one, drawing up to 76% of requests — and concentration rose steeply once displayed candidate lists exceeded about ten carriers. Which carrier won varied wildly between sampled markets, and showing true quality instead of estimated ratings changed nothing. The one intervention that worked was disclosing each carrier's remaining daily capacity: concentration fell by a third and shipper surplus doubled, while vendor diversification, list randomization, and popularity display did nothing. Platform information design, not model choice, is the control surface.