The product is in the cart. Now comes payment
A shopping agent finds the right item and puts it in the cart. It fills in the delivery address and reaches checkout. Depending on the shop, it can pick a payment method as a guest or has to bring the customer back earlier. That’s the point the Bonn agency i22 looked at in its study of 50 large online stores in Germany. The agent reached the choice of payment method on its own in 21 of the 48 stores with usable results, about 44%. Every run ended before an order was placed.
What interests me is the way the study makes us think about the handover. Sometimes the customer starts in the browser and the assistant helps later. Sometimes the customer asks an agent to find the product, prepare the order and see it through once approved. The cart can look identical in either case. What the shop has to know about the customer and what the agent is allowed to do can be quite different.
How did i22 test 50 online stores?
From late October to early November 2026, i22 ran three separate tests. It put 70 shopping questions to free ChatGPT, ChatGPT Plus and Google AI Mode, collecting 210 responses. It checked all 50 shops for machine readable product data and ways for agents to connect. Then it used a Claude browser agent in Chrome to try the guest checkout. The agent had no account with the shops and no saved payment details. i22 recorded 147 purchase attempts from one location and browser profile.
In 22 of the 48 stores with usable checkout results, a person had to take over earlier. In 14, the payment method was the obstacle, and in eight, the shop required an account. The attempt failed for other reasons in five stores, four of them because the order fell below the minimum value. Two shops could not be scored for checkout. The published results are aggregated. They describe one agent setup and a selected group of stores.
Does being found in AI search mean an agent can buy from you?
Almost a third of the stores never appeared in the answers to their category’s questions. Being mentioned made little difference to whether the browser agent reached the payment choice in this test. i22 also found three working routes designed for agents, all focused on discovery or advice. The measured routes did not support completing a purchase.
In Chapter 8 of my book, finding a suitcase is the first step. To buy it, the traveller needs a specific offer she can approve, with the right variant, total and delivery promise. Forter’s analysis of agent orders picks up further down that path, looking at orders on its own network. i22 asks where a particular browser agent stops on the way there. Those are different measurements and both belong in a retailer’s picture of how agents shop.
How I’d test my own checkout
I’d choose a real variant product in test stock and give an agent a precise task. Follow it through the choice of variant and into the offer, then record when it needs the customer. A handover for payment approval may be just what you want. The customer should pick up the right cart and see the terms that apply, then find out whether the order went through.
In a test environment you can go further with my acceptance guide for agent purchases. Look in the shop for the order and the stock change, and check the authorised amount independently with the payment provider. Then run the same task with a changed delivery date. You’ll see whether the agent asks again and whether the offer the customer approved is the one that ends up on the order.
Source and further reading
Read the original announcement ↗
