AI agents went shopping for TVs and laptops—and messed up nearly 50% of the time, study shows

Visa’s test of 22 AI shopping agents found an overall purchase success rate of about 57%; Google Gemini’s Aragon 4 and Meta’s Wonder tied at 82%, while Claude scored 76% and an OpenAI agent 64%. Visa says agent-driven shopping remains early, with product discovery, secure payments and unresolved liability for rogue purchases or fraud among the obstacles to broader adoption.

Analysis

Agentic shopping is a capability signal, not yet a revenue signal. The key bottleneck is not simply model quality: agents must clarify ambiguous intent, navigate retailer-specific catalogs, and complete a trustworthy transaction. Better models may improve conversion, but higher inference cost could consume some of the value unless retailers or platforms can monetize incremental basket size or reduce service costs. This makes near-term upside for GOOG and META primarily strategic optionality; the study does not establish consumer adoption, purchase volume, or attractive unit economics.

For V and MA, the more durable opportunity is to become the authentication, tokenization, and dispute-management layer for delegated purchases. But agent interfaces could also concentrate customer access in AI platforms, increasing pressure on networks to prove their value beyond routing and potentially shifting economics toward identity and fraud controls. Retailers that own product data, inventory, and fulfillment retain leverage; weak catalog quality or inconsistent return terms will limit agent conversion regardless of model advances.

Near term (days), this is likely thematic rather than earnings-moving. Over 1–3 months, watch for merchant integrations and disclosed transaction pilots; over 6–18 months, adoption and fraud/dispute outcomes will determine whether payments economics expand or merely migrate. Contrarian point: the low task success rate may be a stronger warning about commerce workflow and liability than a temporary model deficit. Do not extrapolate model rankings into durable share gains. Thesis improves with repeat-purchase conversion and measurable incremental volume; it weakens if pilots stall, fraud disputes rise, or platforms keep transactions inside proprietary ecosystems.

Key Decisions for Investors

• Do not chase GOOG or META on this study alone. Treat both as watch-list beneficiaries; seek evidence that agent shopping increases paid usage or commerce-linked revenue without inference costs overwhelming the contribution.
• Keep V and MA as conditional medium-term beneficiaries, not an immediate thematic buy. Track merchant acceptance, agent-specific tokenization/authentication launches, and whether these tools preserve network routing rather than enable closed-loop platform payments.
• Set an alert for disclosed pilot metrics: completed-purchase conversion, repeat usage, incremental spend, fraud/chargeback rates, and who bears liability. These are the missing data needed to underwrite revenue and margin impact.
• Falsify the payments-upside thesis if major agent platforms route meaningful volume through proprietary payment arrangements, or if rising disputes and liability costs offset transaction growth; reassess any positive read-through if merchant integrations remain announcements without operating metrics.