Browser automation usually depends on CSS selectors, XPath, or test IDs.

But browsers already build another representation of the page: the accessibility tree.

So I am trying something weird: automate websites using only that tree.

No CSS selectors. No XPath. No test IDs.

Instead of looking for `#checkout-button`, the tool looks for “button named Checkout.”

In this talk, I will build and break that idea on real interfaces, including bad ARIA, fake buttons, dynamic content, and other cases where the browser's semantic view falls apart.

The goal is not to prove this is better automation.

The goal is to see how far we can push the browser before it stops understanding the page.