The Agent Does Not Browse: ChatGPT-Referred Basket Structure Compared to Traditional Channels

Consumers ask AI assistants what to buy; retailers receive the result as organic LLM referral traffic (oLLM). It converts at rising rates, yet its order value is low and falling, and the orders behind it are unknown. On a product-level panel of 677 retailers and 202.9 million orders (January 2025 to July 2026) we compare ChatGPT-referred orders with Organic Search orders at the same store on the same day. An oLLM order is worth 5.8% less, and the entire gap, and more, is quantity: 6.7% fewer distinct products and 10.4% fewer units, at a 5.2% higher price per unit and no discount at any level we measure. Three quarters of oLLM orders contain one product, the deficit sits at the first and second product, and a second item, where added, is a small extra. The quantity gap deepens month by month with a weaker, statistically uncertain price component. No other channel combines this deficit with full within-category prices and product-page entry. The pattern fits a channel that hands over a decided shopper and a product link, not a visitor and a storefront. For retailers, the contest on the AI shelf is to be the one product the order completes around.

Working Paper