Visa finds AI shopping agents succeed only about 57% of the time in controlled tests
Twenty-two agents bought a TV, a laptop and a hair appliance; Gemini Aragon 4 and Meta Making Wonder topped 82%, Claude 76%, and an OpenAI agent 64%.
Why it matters
Visa's controlled shopping study is one of the clearest public scorecards yet for agentic commerce: even frontier-backed agents failed nearly half the time when shoppers withheld some requirements. The gap between Gemini and Meta leaders at 82% and weaker agents near 50% shows model choice still dominates product discovery. Liability for rogue purchases and payment rails remain unresolved. Watch whether merchants adopt Visa and Mastercard agentic-commerce tools after these numbers.
Key facts
- 22 agents, 3 product types
- Overall success ~57%
- Gemini Aragon 4 & Meta Making Wonder 82%
- Claude 76%; OpenAI agent 64%
- Study stresses product discovery, not checkout