Stall library · Buy
A simpler AI agent stopped at a step a capable one got past
This finding means the purchase went through, but not for every kind of AI agent. A scripted walk, which behaves like the simpler agents that act on fixed rules, stopped at a step of the buy path. An AI agent reading the page then took over that step and got past it. The store passes, and the step is where agents with less judgement are likely to give up.
Why the agent stops
Shopping agents vary. Some read the page and reason about what to press; others follow the patterns most stores use and stop when a page does something different. A step that only the first kind can finish is a step where part of the agent traffic is lost.
The step itself is named in the finding, with what the scripted walk saw when it stopped and what the AI agent did instead. That is usually a control built in an unusual way: an option that is not a real form control, a button that stays disabled until something off to the side is pressed, or a popup with no clear way to close it.
How often we see this
Not stated here. A frequency means nothing without the cohort it was counted over, and this page is about one stall rather than one category. Where a category has enough stores measured at every stage, the share of them showing this stall is published on its benchmark page, with the methodology beside it.
See it for yourself
- Open the step named in the finding in a browser and do what the finding says the AI agent did.
- Inspect the control the scripted walk could not use and check whether it is a real form control, such as a select, a radio button or a submit button.
- Try the same step with the keyboard alone. A step that needs a pointer or a guess is the step to change.
What to change
- Build the control from the standard element for the job: a select or radio buttons for options, a submit button inside a form for add to cart.
- Give every control an accessible name that says what it does.
- Follow the fix the finding carries for the step it names, which is the same fix the stall on that step would carry.
What Ottom looks at
The journey agent runs a scripted walk first. When the script stops, an AI agent takes over that one step in the same browser, and code, not the agent, checks whether the step is done. This is raised when that check passes, so the finding records a step that was completed only with judgement a simpler agent would not have.
Questions
Is it a stall when only a capable AI agent gets through?
No. The purchase got to the payment step, so the buy stage passes and this finding is rated low. It is reported because the step it names is one a less capable agent would not finish, and benchmarks do not count it as a stall.
Why fix something an AI agent already got past?
Because the agents shopping on your customers' behalf are not all equally capable, and you cannot choose which one a customer uses. The fix is usually small, the same change that makes the step work with a keyboard, and it removes the step where the simpler ones stop.