Text-Only AI Versions of Websites Strip Out the Actions Agents Need
WebAIM found 95.9% of top home pages fail WCAG 2, and the failures that hide actions from AI agents are the most common ones. Shopify is the only platform with a declared tool surface live at scale.
Updated
The brief
- WebAIM's 2026 evaluation found 95.9% of the top million home pages fail WCAG 2, up from 94.8% in 2025, with errors averaging 56.1 per page, up 10.1% year over year.
- Shopify switched on WebMCP tools — catalog search, cart, checkout and policy lookup — for every Liquid storefront on August 5, 2026, live by default with nothing to install.
- A CHI 2026 study found Claude Sonnet 4.5's task success fell from 78.3% under default conditions to 41.7% keyboard-only and 28.3% with the viewport magnified to 150%.
WebAIM's 2026 evaluation of the top million home pages found 95.9% failing WCAG 2, up from 94.8% in 2025 — and three of the six most common failures are the exact errors that make actions invisible to AI agents. The finding comes from an analysis published on No Hacks arguing that the markdown mirrors and readiness scores now served to AI systems do not merely fail to carry actions, they strip them out.
"A markdown mirror of a web page is prose. Whatever a visitor could have pressed is gone by the time the machine receives it," the author writes.
Seven months after arguing in February 2026 that serving markdown to AI agents solves the reading problem and not the doing problem, the author reports that one platform, Shopify, has touched the doing problem. The markdown mirrors, the readiness scores, and GEO — the discipline for getting cited in AI answers — have not.
Structured data is the machine-only surface websites already have
Structured data is content written for machines and invisible to people. JSON-LD alone runs on 55.6% of the websites W3Techs measures, as of September 2026. It survives any simplification exercise because a person was never its audience.
The markdown mirrors and readiness scores released since February belong to the same family, according to the analysis. Both answer what a page is. Neither answers what can be done with it. The readiness scanner checks whether WebMCP tools are registered on the page and says nothing about whether they work.
Semantic HTML is the floor, a declared tool surface is the ceiling
A website has two structural routes to expose actions to a machine, and both work. The floor is semantic HTML and accessibility features: a native button says what it is, a label names a field, a form has a target and a method. The ceiling is a declared tool surface — WebMCP is the current proposal, where the page registers functions an agent may call without inferring anything from markup.
Ideally a site builds the ceiling. But WebMCP is still a proposed standard, so the floor is what matters. And the floor is broken for most websites.
WebAIM's 2026 numbers quantify it. Errors averaged 56.1 per page, up 10.1% in a year, reversing six years of small improvements. Pages using ARIA averaged 59.1 errors against 42 for pages without it — though WebAIM notes those pages were also more complex.\n The failures that make actions disappear are everywhere. Form inputs with no label appeared on 51% of home pages, empty links on 46.3%, and empty buttons on 30.6%. "A button with no name is a button an agent cannot tell apart from the one next to it," the analysis states.
On those pages, the accessibility tree an agent works through is missing the same things a screen reader is missing. A study accepted to CHI 2026 ran Anthropic's Claude Sonnet 4.5 as a computer-use agent on 60 everyday tasks and measured success falling from 78.3% under default conditions to 41.7% keyboard-only and 28.3% with the viewport magnified to 150%. It did not test broken markup; it tested the agent working the way people who rely on assistive technology work.
Agents that cannot confirm success will repeat the action
On a project the author worked on, an AI agent set to submit web forms was broken by two things: invalid HTML, and the absence of programmatic success or error feedback. "The agent would try to submit again, because it did not know it had submitted successfully."
The form worked. A person would have seen the confirmation and stopped. The agent could not, because the confirmation was rendered for eyes, so it repeated the request. "Every duplicate order, duplicate ticket and duplicate signup out of that pattern is the website's missing feedback, not the agent's mistake."
Shopify installed the actions, not the merchants
Shopify switched on WebMCP tools for every storefront built on Liquid, its theme language, on August 5, 2026: catalog search, cart, checkout and policy lookup, live by default, nothing to install. In August it was the only declared tool surface the author knew of running at scale — and the merchants did not build it. Three storefronts checked the next day all loaded the same adapter script from Shopify's CDN, because the platform wrote it once and serves it everywhere.
The tools read the same catalog, cart and checkout the human storefront does, so they could not drift, and they cost the merchant nothing. On August 6, the read path worked — a catalog search returned products and prices — while the buying path died on an internal error. The author has not re-run it since.
Shopify told its earnings call on August 5 that AI-driven traffic and orders had tripled year over year. That claim, the analysis notes, describes humans arriving from AI answers and buying the way humans always have — so whether a machine has actually called those tools remains an open question.
GEO optimizes the half that cannot act
GEO — generative engine optimisation — focuses on getting cited or recommended in LLM answers, and the author concedes citation is what pays today: "It is how people arrive; it is measurable enough to sell, and telling a business to ignore it would be wrong."
But GEO has nothing to do with action, and what it does touch — how describable a page is — is what SEO already touched. The author's sharpest point: GEO is sold as SEO for more powerful systems, yet what makes those systems more powerful is that they can act. "So the discipline named for the new systems is the one that does not address what makes them new."
Every company that can cite a business today is working on a version of itself that can also do more, and on the protocols to enable it with other vendors in the space, according to the analysis.
What marketers should do now
The author's framework: a website has three layers — visual, structure, and content — and the actions live in the structure. Remove the visual layer and a machine can still use what remains. Remove the structure, which is what a text-only version does, and the machine can only read about it. The recommendations:
- Sort what you serve machines into describable and doable. Expect a full first pile and an empty second one. The sorting takes a few hours.
- Find out who installed whatever is in the second pile. A platform probably put it there, and the terms sit in documentation you may never have opened.
- Check what your form says back. If the only sign it worked is a green box, an AI will likely struggle with it.
- Fix the floor before buying a ceiling. Run WAVE, the checker WebAIM's evaluation is built on, against the page with your most important form, and count the unlabelled inputs, empty links and empty buttons.
The closing argument lands hard on the current text-only trend: "A text-only version with the actions taken out is a brochure. And it's served to the user most eager to act."
Based on nohacks.co
Filed under geo, accessibility, webmcp, ai-agents, structured-data
More from Priya Raman
Show full bio
Correspondent covering industry trends and analytics at Marketing Herald.
41 articles