The conversation about shopping agents usually starts with discovery: does the agent find the product, does it pick the right variant? In our field study that part was rarely the problem. The losses came later: after the product was in the cart, in the last few clicks before checkout.
Three sticking points from the field study
On 7 October 2026 we sent AI agents shopping in a real browser 40 times across five live stores. On the nine tasks that could be completed, the two strongest models reached checkout on 9 of 9; the two lighter models on 7 of 9. In other words, the lighter models failed to reach checkout in 22% of the tasks (2 of 9). It's a small sample; read it as a field observation, not a success rate (the full field study).
Three places where agents got stuck between the cart and checkout:
- A checkout button you can see but the agent can't read. On one store, the "Go to checkout" button in the cart drawer sat inside a third-party component built with a closed shadow DOM. The button was on screen, but it never appeared among the elements the agents could act on. Agents only reached checkout by guessing the URL.
- A cart URL the agent has to guess. On a store where the cart link was hard to find, some agents tried to guess the cart address and landed on not-found pages.
- Sign in or continue as a guest. Agents reached this screen and the run ended there. Without a guest option, an agent acting for a user would have to stop: it can't create an account or sign in on the user's behalf.
What the three have in common: none of them would stop a human shopper. A person sees the button and clicks it, spots the cart icon, notices the "guest" link on the sign-in screen. An agent can only use what the page offers it.
How the simulator marks the cart section
After each run, the agent simulator produces a 33-item friction checklist. The "Cart & account" section has five items: forced sign-up, a surprise login wall, pre-ticked boxes, items added to the cart automatically, and cart timers. These items are marked from the run log by rules, without an AI model. On e-commerce sites each item gets one of four states:
- Detected: the friction was flagged in this run, with evidence and the step where it happened.
- Clear: the agent reached that page, we looked, and the friction wasn't there.
- Unreached: the agent never got to that page, so this run couldn't check it. That does not mean "clear".
- Expert: an item a live run can't see; an expert checks it.
Payment and post-order items are not measured automatically by design, because the agent stops at checkout: we never place orders, pay or enter personal data. Those items open an expert task.
The "unreached" state matters most in the cart section. The forced sign-up item, for example, can only be checked once the agent reaches checkout or the "sign in or continue as a guest" screen; items added automatically only once the agent adds a product and opens the cart. If the agent didn't get that far, these items show as "unreached" and the real problem sits one step earlier (choosing a variant or adding to cart, for instance). Fixing that item first lets the next run actually check the cart items.
What to fix first
- Keep the link to checkout on the page itself. Even if a third-party cart component shows its own button, keep a standard link or button on the page that an agent can read.
- Make cart and checkout URLs predictable. Have a crawlable cart link in the main navigation; send common cart and checkout URLs to a meaningful page instead of a not-found page.
- Make guest checkout as visible as the sign-in button. Forced sign-up is where the agent's journey ends.
- Don't add items to the cart on your own, and don't pre-tick boxes. The agent has to notice and remove them, or it misses them and takes a wrong cart back to its user.
- Hold the cart for as long as the agent needs to ask its user. A cart that expires while waiting for approval means starting over.
Most of these also make life easier for human customers. The difference: a person works around the friction and carries on; an agent hits the same friction and goes to another store. The agent simulator shows step by step how these last few steps go on your store.