The most agreeable shopper ever assigned to this desk spent an afternoon trying to cancel a streaming subscription and ended up, instead, the proud owner of a discounted annual plan. It found the account page easily enough. It located something called “Manage membership.” What it never found was the cancel button, because the cancel button was not, in any way a machine can use, a button. It was a grey text link reading “click here,” crouched between two large, properly labelled buttons promising to “Keep my benefits.” The agent kept the benefits.
The shopper was an AI browsing agent — one of the new crop of assistants you point at the web and tell to get on with it. The desk pointed it at three ordinary errands: buy a specific replacement gasket for a six-cup moka pot, cancel that subscription, and find a phone number at which a human being might conceivably answer. The results were unscientific and unambiguous. And the failures looked familiar — almost file-for-file familiar — to anyone who has watched a screen reader attempt the same errands. The same invisible buttons. The same links that announce nothing. The same forms whose fields introduce themselves to no one. A shopping agent and a screen reader, it turns out, read a page the same way: the description, not the picture.
The page underneath the page
The resemblance is not a coincidence; it is the architecture. Beneath every web page’s picture — the gradients, the hero image, the big orange pill of a checkout button — the browser maintains a second version of the site: a tree of roles, names, states and relationships that says, in effect, “this is a button, this is its name, this is the main content, this is a field labelled ‘card number.’” Screen readers have consumed that description for as long as the web has had standards, which is since 1999; the devices themselves are older than the web. As the people building agents have begun pointing out, a browsing agent consumes the same description. To a machine, a gorgeous checkout button coded as a bare, nameless rectangle is not a button at all. It is décor.
Which is how the gasket expedition died: the agent browsed, compared, even found a sensible price, then stalled at a checkout whose size selector was a row of anonymous boxes and whose pay button announced itself as nothing. The phone-number hunt was more poetic. The agent worked through the help centre with great diligence and concluded the company had no phone number — which is precisely the conclusion the company spent considerable design money engineering humans to reach. And the cancellation maze, that calibrated obstacle course of retention offers and grey-on-grey links, caught the agent as reliably as it catches people. A human gives up; an agent loops. Either way, nobody cancels.
A necessary caveat: not every agent reads the tree. Some of the newest systems work from screenshots, peering at the picture roughly as a person would, and those can see an unlabelled orange pill perfectly well. But the agents being aimed at commerce today mostly read structure, because structure is fast and cheap — and even the screenshot-readers still inherit the mazes, the pop-ups and the CAPTCHAs built specifically to stop them.
Our trials were a handful of afternoons, not a benchmark. The closest thing to a census is a 2025 audit by Webply, an accessibility vendor that ran automated checks over 116 European business websites — 1,556 pages — and found all but two sites carrying at least one problem it classed as “agent-blocking”: 4,735 links and 607 buttons with no accessible name at all, and more than 3,000 links whose entire identity was “click here,” “read more,” “learn more” or “here.” On 46 of the 116 sites, unlabelled form fields sat on login, checkout or contact pages — the exact rooms where money changes hands. The caveats write themselves: a company that sells accessibility fixes has discovered an accessibility emergency, its checks were automated and catch only some barriers, and “agent-blocking” is its own filter. But the direction is hard to wave away, and it matches what we watched.
A blind user with a credit card
This is where the story stops being about a clumsy robot. For more than twenty-five years, accessibility has been argued for on grounds of rights, decency and, eventually, law — the EU’s European Accessibility Act now explicitly covers e-commerce and banking — and compliance has been grudging, minimal, eternally on the roadmap. Disabled people, roughly one in four of the UK population by the government’s own survey, were filed as an edge case. Then the industry decided the future of shopping is agents: software that browses, compares and buys on your behalf. And an agent, structurally speaking, is a blind user with a credit card. The moment the buyer is a machine, the unlabelled button stops being an accommodation request and becomes an abandoned cart. Accessibility is about to be rebranded as conversion-rate optimization, and at last it will get budget.
There is a vindication here with an uncomfortable edge. Everything the accessibility community said for decades was true; it simply wasn’t priced. The fix is coming not because it was right but because it is profitable, and the people who asked first may be forgiven for not applauding the timing. The sharper question is what kind of fix arrives. A business that needs machines to buy, but has never much minded humans struggling, has an obvious shortcut: skip the page entirely and build the machines their own entrance — structured product feeds, agent-only checkout lanes, clean pipes running underneath the funhouse. The curb cut was the great parable of accessibility: a ramp built for wheelchair users that ended up serving everyone with a pram or a suitcase. This time the ramp may be built for the suitcase. The machines will get their smooth entrance, and the people who requested one a quarter-century ago will be told theirs is on the roadmap.