Research · October 2026

How often do AI shopping agents fail at checkout?

We sent AI shopping agents to buy something on 73 real online stores, from search to the payment page. More than a quarter of cloud-agent checkouts never got there, and none of the stores could see why.

28%of cloud-agent checkouts never reached the payment page

73real online stores, 102 checkout runs

16 → 4steps wasted on one broken checkout, before and after a warning from earlier runs

100 checkout runs, one square each

Reached the payment page (72)Never got there (28)

Cloud (headless) browser-use agent with Claude, 73 US stores. Two further runs are left out because they measured our test email, not the store (see caveats).

How we tested

The agent: browser-use with Claude, in a cloud (headless) browser, the way most agent companies run today.
The task: find a product, choose options, add it to the cart, check out as a guest and fill in shipping with test details. Success means reaching the payment page. The agent never entered payment details or placed an order.
The stores: 73 US online stores across apparel, home, beauty, food, shoes and more, on Shopify, Salesforce Commerce, WooCommerce, BigCommerce, Magento and custom platforms. Tested in three waves in October 2026, one to two runs per store.
Checking: a separate model judged every run from its recording, and we reviewed disagreements by hand.

What we found

Bot checks were the most common blocker. In our last wave, all 7 failures were bot walls on the very first page: the agent never saw a product.
Many bot walls only block cloud browsers. When we reran those 7 stores in a regular, visible browser, 5 got through. For agents running in the cloud, those stores are closed. For an agent in a person's own browser, they mostly aren't.
Broken checkout steps fail silently. A Continue button that greys out and does nothing, a bot check that appears halfway through checkout, a page that suddenly asks for an account. The agent retries, gives up, and the store's analytics count a visitor who left.
Agents repeat the same failure. Agents usually start each attempt from scratch. On one store with a broken email field, our agent spent 16 steps fighting it. With a one-line warning from earlier runs, it stopped in 4 and told its user.
Agents will report problems if they're allowed to. We put a visible "Report a problem" form on a test store. Without permission, 0 of 3 agents used it (all 3 mentioned it to their user instead). With one line in their instructions allowing reports, 3 of 3 did, and none included personal details.

The 7 stores whose bot wall stopped our cloud agent

Cloud (headless) browser

Regular, visible browser

Got throughBlocked

Same agent, same task, wave 3 (October 2026).

Steps spent on one broken checkout field

Starting from scratch

With a one-line warning

With the warning, the agent stopped early and told its user. Small sample: 2 runs each way.

Agents that sent a problem report to the store

Not told they could

One line allowing it

On a test store with a visible "Report a problem" form. None of the reports included personal details.

Real shopping agents hit it too

Our numbers come from one test agent, so we gave two consumer AI shopping agents the same task on the store where every one of our runs stalled at a silent Continue button. We're not naming the agents or the store.

Same bug, 4 of 5 runs. Over two days, both agents hit the same dead Continue button. It doesn't happen every time, which makes it easy for the store to miss in its own testing.
They got through, slowly. Each found a workaround (reloading, rebuilding the cart, a fresh browser session), but the first runs took 16 to 18 minutes for one order.
Memory didn't fix it. On day two, one agent remembered the bug and still hit it twice. Workarounds didn’t carry over: a fix that worked in one run failed in the next, including when we handed it to the agent as a tip.

Five real-agent checkouts on the same store

Day 1 · Agent AHit the dead buttonReloaded past it. 16 minutes.

Day 1 · Agent BHit the dead buttonRebuilt the cart by another route. About 18 minutes.

Day 2 · Agent ANo bug this timeAbout 5 minutes, helped by remembering day 1.

Day 2 · Agent BHit it twiceDay 1's fix failed. Needed a fresh browser session.

Day 2 · Agent B, with a tipTip’s fix failedFound a new way past it. 10 minutes.

Both agents reached the payment page every time; no orders were placed. Times are from message timestamps and include waiting on our replies.

What this means

For online stores: some of your "abandoned" sessions are AI agents shopping for real customers, stopped by a bot check or a broken step you can't see in analytics. Even when a capable agent gets through, a broken step can stretch one order past fifteen minutes. Agents can tell you exactly where, in their own words.

For agent builders: a large share of failed runs happen on steps other agents have already failed on. Knowing about them before a run saves steps and tokens, and lets the agent tell its user early.

About Herald

Herald is a feedback layer between AI agents and websites. Agents report where a site broke for them, businesses learn what to fix, and other agents get warned before they hit the same problem. It works with any MCP client, browser-use, or a site's own script tag.

Connect your agent Get a free report on your site

Caveats: small samples per store (one to two runs), one agent framework and model, and results from October 2026; sites change. The 28% applies to cloud and headless agents. The real-agent results are five checkouts run by hand on one store over two days; both agents could see their earlier runs. One store rejected our original @example.com test email, which looked like a broken step; we left those two runs out. Questions about the data: hello@useherald.dev.