This visitor is an AI agent, and many agents read a site much as a screen reader does: through its structure and labels, by what the markup says rather than how it looks.
Agents now arrive ready to buy
And increasingly, that visitor has come to do business. The newest agentic browsers go well beyond summarising a page for whoever sent them. Ask one to find the cheapest direct flight and book it, and it will work through your site all the way to payment with no human at the screen. When someone delegates a purchase like that, the agent arrives at your checkout with the money.
What the research shows
All of this depends on the agent getting through, and the early evidence gives cause for concern. A study accepted to CHI 2026 gave an AI agent sixty everyday web tasks. Under normal conditions it completed 78% of them. Restricted to keyboard-only navigation, a fair stand-in for how screen-reader users and assistive technology move through a page, success fell to 42%. The obstacles that turn a disabled person away are, almost to the letter, the obstacles that leave an agent two steps short of paying you.
Accessibility as a revenue question
The agentic web adds a commercial reason to the case for accessibility. A div styled to look like a button, or a form the keyboard can't tab through, becomes a locked door with a paying customer on the other side.
Accessibility has often been treated as a compliance cost. It now shows up as revenue too: every inaccessible checkout is a sale an agent couldn't complete, and a customer who couldn't complete it either.
What this means for testing
So the question I'm adding to my test plans goes beyond "can a person use this?" to "can an agent get through it?"
Clean semantic markup and a working keyboard path, the unglamorous discipline we've always known mattered, turn out to be exactly what this next wave of users needs - human or machine.
The fundamentals stay the same. The audience has grown.
By Simon Royall, Test Analyst at Mando Group