Why do AI agents struggle with websites?

Websites assume a human is watching

A person reads a page as a whole: they notice the greyed-out button, the red border, the toast that flashed for two seconds, the fact that nothing happened after a click. An agent sees a serialised snapshot and has to infer state from it.

• State changes with no URL change, so the agent cannot tell if progress happened.
• Errors expressed as colour, icons or animation instead of text.
• Controls that look like buttons but are unlabelled divs.
• Interfaces that depend on hover, drag, or precise scroll position.
• Content that only appears after a lazy load the agent never triggers.

The long tail is where it breaks

First steps usually work. Failures cluster late in the journey — the address form, the delivery selector, the consent modal, the confirmation screen. That is also where failure is most expensive, because the agent was one step from converting.

It is not a fixed target

Two agents from the same vendor can behave differently after a model update, with nothing changed on your side. A journey that passed last month can fail today. That is why a one-off audit ages badly and repeated testing does not.

What to do about it

Run real agents against your real journeys, find the exact step they lose, fix that step, then keep the check running. PlateTale does that end to end.