The Text-Only Version Of Your Website Strips Out The Wrong Layer

In February 2026 I wrote that serving markdown to AI agents solves the reading problem and not the doing problem. Seven months on, one platform, Shopify, has touched the doing problem, and the markdown mirrors, the readiness scores and GEO, the discipline for getting cited in AI answers, have not.

The text-only versions now being served to AI 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.

Contents

• A website no human will ever open needs no heavy JavaScript and no visual layer
• Structured data is the machine-only surface websites already have
• Semantic HTML is the floor for exposing actions and a declared tool surface is the ceiling
• Fixing the floor is the work, and it is measured every year
• An agent that cannot tell whether it worked will do it twice
• The actions machines can call today were installed by a platform
• GEO optimises the half that cannot act
• The visual layer can wait until the structure and the actions work

A website no human will ever open needs no heavy JavaScript and no visual layer

Heavy JavaScript is absolutely unnecessary for a website that is not some kind of web app, and the visual layer is not essential for machines either, especially if the structural layer is healthy. So put up a website that you know for certain no person will ever open, as an exercise, and both go. Layout, navigation, the design system, the image treatments and the hierarchy that tells an eye where to land all go with them, and nothing is lost.

Structured data is the machine-only surface websites already have

Structured data is content written for machines and invisible to people. JSON-LD alone is on 55.6% of the websites W3Techs measures, as of September 2026. Structured data survives the exercise because a person was never its audience.

The markdown mirrors and the readiness scores released since February are newer members of the same family. Both answer the same question, which is what this page is. The mirrors answer nothing about 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 for exposing actions and a declared tool surface is the ceiling

A website has two routes to structurally expose its actions to a machine, semantic HTML and a declared tool surface, and both work. Back in the exercise, the website no person will ever open, the machine turns up to do something rather than read something. It wants to cancel a subscription or buy the thing, and there is no screen and nobody to click.

• The floor is semantic HTML and accessibility features. Native elements say what a thing is and what it does. A <button> is a button, a <label> names a field, a <form> has a target and a method. Any website with a form either has this or should have it already, for reasons that predate agents by decades.
• The ceiling is a declared tool surface. WebMCP is the current proposal: the page registers the functions an agent may call, and the agent calls them without inferring anything from the markup.

Ideally you build the ceiling. The ceiling is still a proposed standard, so the floor is what matters, and the floor is broken for most websites.

Fixing the floor is the work, and it is measured every year

WebAIM's 2026 evaluation of the top million home pages found 95.9% failing WCAG 2, up from 94.8% in 2025, reversing six years of small improvements. Errors averaged 56.1 per page, up 10.1% in a year. Pages using ARIA, the attributes for adding accessibility semantics to markup, averaged 59.1 errors against 42 for pages without it. WebAIM is careful to say those pages were also more complex. The markup that reaches for extra semantics is still the markup carrying more errors.

Three of the six most common failures are the actions going missing. Form inputs with no label were 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. Every one of those is a fix on the page every visitor already uses. A text-only version leaves all of them where they are and adds a file.

On those pages the HTML is not semantic, so the accessibility tree an agent works through (how AI agents see your website covers the mechanism) is missing the same things a screen reader is missing. An unlabelled input is gone from both. A study accepted to CHI 2026 ran Anthropic's Claude Sonnet 4.5 as a computer-use agent on 60 everyday tasks and measured its 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.

An agent that cannot tell whether it worked will do it twice

On a project I worked on, an AI agent was set to submit web forms, and two things broke it: invalid HTML (obviously), and the absence of any 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. The invalid HTML is a floor problem. The missing feedback is one a form can have on a perfectly valid page.

After a machine does something, it has to be told what happened, in a form it can read. Nothing in a text-only version says anything about that, and neither does the readiness score. That is the list the exercise produces: expose what can be done, make it callable, report what happened.

The actions machines can call today were installed by a platform

Shopify switched on WebMCP tools for every storefront built on Liquid, its own 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 place I knew of with a declared tool surface live at scale, and the merchants did not build it. I checked three storefronts the next day and all three loaded the same adapter script from Shopify's CDN, because the platform wrote it once and serves it everywhere. The tool descriptions and results are instructions to the machine caller, the checkout tool moved the browser to the checkout page and told it not to navigate again, and a search result told it which tool to reach for next, and no merchant had to read a word of them.

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, and the buying path died on an internal error. I have not re-run it since. Shopify told its earnings call on August 5 that AI-driven traffic and orders had tripled year over year, and that claim describes humans arriving from AI answers and buying the way humans always have, so whether a machine has called those tools is still an open question. Parts of that list, expose, make callable, report, are being satisfied on somebody else's schedule, for people who were not asked.

GEO optimises the half that cannot act

We need to talk about what GEO is, and what it isn't.

GEO, generative engine optimisation, focuses on getting cited, or recommended, in that LLM answer, and it is absolutely correct that citation is what pays today, so there is no need to argue it. It is how people arrive, it is measurable enough to sell, and telling a business to ignore it would be wrong.

GEO has absolutely nothing to do with the action. What it does touch, how describable a page is, is what SEO already touched.

GEO is sold as SEO for more powerful systems. 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.

The defence is that action is a bet on a future that has not arrived. But every single company that can cite you today is working on a version of itself that also does more (the agentic browser landscape lists them), while working with the other vendors in the space on the protocols to do it. It is wrong to stop there.

The visual layer can wait until the structure and the actions work

A website has three layers: the visual one, the structure, and the content. The actions live in the structure. Take away the visual layer and keep the other two and a machine can use what is left. Take away the structure as well and keep only the content, which is what a text-only version does, and the machine can only read about it. Machine-First Architecture, the way I have been arguing websites should be built, means the structure and the actions have to work on their own, with or without the visual layer on top. The visual layer stays, for the people, and it is the last thing to worry about.

1. Sort what you serve machines into describable and doable. Expect a full first pile and an empty second one. The sorting takes an few hours.
2. Find out who installed whatever is in the second pile. If anything is there, a platform probably put it there, and the terms sit in documentation you may never have opened. Read it.
3. Check what your form says back. Submit it and read the response the way a machine would. If the only sign it worked is a green box, an AI will likely struggle with it.
4. 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. Those three are among the six most common failures in WebAIM's 95.9%, and they are the three that make actions disappear. A declared tool surface does not fix the page it hands the agent back to.

A version of your website built for machines is the whole point. A text-only version with the actions taken out is a brochure. And it's served to the user most eager to act.

You could make an argument this makes the brochure pointless.