Research · Study 03 · Paper under review

Can AI shopping agents buy from online stores?

Where stores let them in, AI agents can shop, but there's no trusted way for them to pay.

37%of the stores we tested stopped the agent before payment
413 of 427stores gave all three big models the same outcome
195 of 197checkouts set up through UCP; each one stopped before payment

One-page PDFRequest the data

Getting in

37% of the stores we tested stopped the agent before it reached payment (175 of 469 stores).

Stores that stopped the agent, by store popularity

Changing the model

Changing the AI model did not help an agent get in. Stores treated each agent largely the same: 413 of 427 stores gave Claude, GPT and Gemini the same result.

Share of stores that stopped each model

Paying

At Shopify stores that offer UCP, the agentic checkout route, the agent could set up a checkout almost every time (195 of 197), but every checkout stopped before payment. 181 came back requires_escalation, which UCP tells the agent to hand to the buyer. The other 14 came back ready_for_complete, but completing a checkout needs Shopify's permission to call complete_checkout, which the script did not have and never tried.

How far the agents got, through the website and through UCP

Add-ons

Of the 162 runs with a paid add-on on the payment page, 47 (29%) ended with it in the cart, such as package protection, returns coverage, a protection plan, or a membership. We ran a follow-up test on a checkout we built, with a $3.98 add-on already ticked. GPT and Gemini paid for it every time, and Claude removed it every time.

What happened when an add-on was offered

Three runs

Stitched from the screenshot saved at every step. Captions show what the agent did next.

Claude at MAC CosmeticsWatched the screen like a person. Reached the payment step in 24 actions, about a minute.
Gemini at the same storeRead the page as text, with its reasoning under each step. 10 actions.
Claude at Brilliant EarthA Cloudflare "Verify you are human" check after the first click. Agents were not allowed to solve CAPTCHAs, so the run ended.

Method and limits

Stores. We picked 500 popular US online stores, including large national retailers and smaller well-known brands. Just under half run on Shopify (222) and the rest on other e-commerce platforms (278), like Salesforce Commerce Cloud or their own custom builds. Our selection also includes all 55 of the NRF Top 100 retailers that sell online. When measuring results, we split results into 4 tiers ranked by popularity, using the Tranco list (Le Pochat et al. 2019).

No purchases were completed: no agent completed submitting payment, so this measures how often agents reached the payment step rather than completed purchases; the test-checkout follow-up covers the card step only.

Undeclared automation, independent agent: the agents drove a standard Chrome browser through automation software and did not identify themselves, and we tested with no platform partnership. Agents that run inside a person's own browser or sign their requests were not tested, and routes where approved agents can pay exist but none is open to an unaffiliated agent.