Visa finds AI shopping agents pass 57% of purchase runs

Overall, 57.2% of runs passed. Another 23.1% ended in a purchase that missed. The last 19.7% split two ways: runs with no decision, and customers who left with nothing although a product fitted, usually after the assistant said none did.

Gemini 4 Argon and Muse Wonder share the top score, 82.2%, though Visa says no model clearly leads, because the margins of error overlap with the next eleven. Price tracked results: the cheapest five models per run managed 50%, the dearest five 67%.

The way each one fails differs. In 38% of its runs GPT-6.1 Terra walked away from a sale it could have made, with a suitable product available. Gemini 3.1 Pro leaned the other way, and 51% of its runs ended in a purchase that missed.

The assistants people reach for are cheap to try. Google AI Pro was $19.99 a month and Claude Pro $20 at our 6 November check, though Visa’s cost figure is per run and does not map to those plans.

For a seller the two failures would land differently. The model that walks away loses a sale that was there to be had, and the model that buys wrongly delivers an order that missed a requirement or came before that requirement surfaced.