In February 2026, I wrote that serving markdown to AI agents solves the studying drawback and never the doing drawback. Seven months on, one platform, Shopify, has touched the doing drawback, and the markdown mirrors, the readiness scores, and GEO, the self-discipline for getting cited in AI solutions, haven’t.
The text-only variations now being served to AI don’t merely fail to hold actions; they strip them out. A markdown mirror of an online web page is prose. No matter a customer might have pressed is gone by the point the machine receives it.
A Web site No Human Will Ever Open Wants No Heavy JavaScript And No Visible Layer
Heavy JavaScript is totally pointless for an internet site that’s not some type of net app, and the visual layer is not essential for machines both, particularly if the structural layer is wholesome. So put up an internet site that you recognize for sure no individual will ever open, as an train, and each go. Format, navigation, the design system, the picture therapies, and the hierarchy that tells an eye fixed the place to land all go together with them, and nothing is misplaced.
Structured Information Is The Machine-Solely Floor Web sites Already Have
Structured knowledge is content material written for machines and invisible to folks. JSON-LD alone is on 55.6% of the websites W3Techs measures, as of September 2026. Structured knowledge survives the train as a result of an individual was by no means its viewers.
The markdown mirrors and the readiness scores launched since February are newer members of the identical household. Each reply the identical query, which is what this web page is. The mirrors reply nothing about what might be carried out with it. The readiness scanner checks whether or not WebMCP instruments are registered on the web page, and says nothing about whether or not they work.
Semantic HTML Is The Ground For Exposing Actions And A Declared Instrument Floor Is The Ceiling
A web site has two routes to structurally expose its actions to a machine, semantic HTML and a declared instrument floor, and each work. Again within the train, the web site no individual will ever open; the machine turns as much as do one thing relatively than learn one thing. It needs to cancel a subscription or purchase the factor, and there’s no display and no person to click on.
- The ground is semantic HTML and accessibility features. Native components say what a factor is and what it does. A
is a button, anames a area, ahas a goal and a technique. Any web site with a kind both has this or ought to have it already, for causes that predate brokers by a long time. - The ceiling is a declared instrument floor. WebMCP is the current proposal: the web page registers the features an agent might name, and the agent calls them with out inferring something from the markup.
Ideally, you construct the ceiling. The ceiling continues to be a proposed commonplace, so the ground is what issues, and the ground is damaged for many web sites.
Fixing The Ground Is The Work, And It Is Measured Each 12 months
WebAIM’s 2026 analysis of the highest million residence pages discovered 95.9% failing WCAG 2, up from 94.8% in 2025, reversing six years of small enhancements. Errors averaged 56.1 per web page, up 10.1% in a 12 months. Pages utilizing ARIA, the attributes for including accessibility semantics to markup, averaged 59.1 errors in opposition to 42 for pages with out it. WebAIM is cautious to say these pages had been additionally extra advanced. The markup that reaches for additional semantics continues to be the markup carrying extra errors.
Three of the six most typical failures are the actions going lacking. Type inputs with no label had been on 51% of residence pages, empty hyperlinks on 46.3%, and empty buttons on 30.6%. A button with no identify is a button an agent can not inform other than the one subsequent to it. Each a kind of is a repair on the web page each customer already makes use of. A text-only model leaves all of them the place they’re and provides a file.
On these pages the HTML is just not semantic, so the accessibility tree an agent works through (how AI agents see your website covers the mechanism) is lacking the identical issues a display reader is lacking. An unlabelled enter is gone from each. A study accepted to CHI 2026 ran Anthropic’s Claude Sonnet 4.5 as a computer-use agent on 60 on a regular basis duties and measured its success falling from 78.3% underneath default circumstances to 41.7% keyboard-only and 28.3% with the viewport magnified to 150%. It didn’t take a look at damaged markup, it examined the agent working the best way individuals who depend on assistive know-how work.
An Agent That Can’t Inform Whether or not It Labored Will Do It Twice
On a venture I labored on, an AI agent was set to submit net varieties, and two issues broke it: invalid HTML (clearly), and the absence of any programmatic success or error suggestions. The agent would attempt to submit once more, as a result of it didn’t comprehend it had submitted efficiently.
The shape labored. An individual would have seen the affirmation and stopped. The agent couldn’t, as a result of the affirmation was rendered for eyes, so it repeated the request. Each duplicate order, duplicate ticket and duplicate signup out of that sample is the web site’s lacking suggestions, not the agent’s mistake. The invalid HTML is a ground drawback. The lacking suggestions is one a kind can have on a wonderfully legitimate web page.
After a machine does one thing, it must be instructed what occurred, in a kind it could actually learn. Nothing in a text-only model says something about that, and neither does the readiness rating. That’s the listing the train produces: expose what might be carried out, make it callable, report what occurred.
The Actions Machines Can Name In the present day Had been Put in By A Platform
Shopify switched on WebMCP tools for every storefront built on Liquid, its personal theme language, on August 5, 2026: catalog search, cart, checkout and coverage lookup, reside by default, nothing to put in. In August, it was the one place I knew of with a declared instrument floor reside at scale, and the retailers didn’t construct it. I checked three storefronts the next day and all three loaded the identical adapter script from Shopify’s CDN, as a result of the platform wrote it as soon as and serves it in all places. The tool descriptions and results are instructions to the machine caller, the checkout instrument moved the browser to the checkout web page and instructed it to not navigate once more, and a search outcome instructed it which instrument to achieve for subsequent, and no service provider needed to learn a phrase of them.
The instruments learn the identical catalog, cart, and checkout the human storefront does, so they may not drift, and so they price the service provider nothing. On August 6, the learn path labored; a catalog search returned merchandise and costs, and the shopping for path died on an inside error. I’ve not re-run it since. Shopify told its earnings call on August 5 that AI-driven site visitors and orders had tripled 12 months over 12 months, and that declare describes people arriving from AI solutions and shopping for the best way people all the time have, so whether or not a machine has referred to as these instruments continues to be an open query. Elements of that listing, expose, make callable, report, are being happy on any individual else’s schedule, for individuals who weren’t requested.
GEO Optimizes The Half That Can’t Act
We have to speak about what GEO is, and what it isn’t.
GEO, generative engine optimisation, focuses on getting cited, or really useful, in that LLM reply, and it’s completely appropriate that quotation is what pays as we speak, so there is no such thing as a must argue it. It’s how folks arrive; it’s measurable sufficient to promote, and telling a enterprise to disregard it might be incorrect.
GEO has completely nothing to do with the motion. What it does contact, how describable a web page is, is what SEO already touched.
GEO is bought as search engine optimisation for extra highly effective programs. What makes these programs extra highly effective is that they will act. So the self-discipline named for the brand new programs is the one that doesn’t handle what makes them new.
The defence is that motion is a guess on a future that has not arrived. However each single firm that may cite you as we speak is engaged on a model of itself that additionally does extra (the agentic browser landscape lists them), whereas working with the opposite distributors within the area on the protocols to do it. It’s incorrect to cease there.
The Visible Layer Can Wait Till The Construction And The Actions Work
A web site has three layers: the visible one, the construction, and the content material. The actions reside within the construction. Take away the visible layer and preserve the opposite two, and a machine can use what’s left. Take away the construction as properly and preserve solely the content material, which is what a text-only model does, and the machine can solely examine it. Machine-First Architecture, the best way I’ve been arguing web sites must be constructed, means the construction and the actions need to work on their very own, with or with out the visible layer on high. The visible layer stays for the folks, and it’s the very last thing to fret about.
- Kind what you serve machines into describable and doable. Count on a full first pile and an empty second one. The sorting takes a number of hours.
- Discover out who put in no matter is within the second pile. If something is there, a platform in all probability put it there, and the phrases sit in documentation you might by no means have opened. Learn it.
- Test what your kind says again. Submit it and skim the response the best way a machine would. If the one signal it labored is a inexperienced field, an AI will seemingly battle with it.
- Repair the ground earlier than shopping for a ceiling. Run WAVE, the checker WebAIM’s analysis is constructed on, in opposition to the web page along with your most essential kind, and depend the unlabelled inputs, empty hyperlinks, and empty buttons. These three are among the many six most typical failures in WebAIM’s 95.9%, and they’re the three that make actions disappear. A declared instrument floor doesn’t repair the web page it arms the agent again to.
A model of your web site constructed for machines is the entire level. A text-only version with the actions taken out is a brochure. And it’s served to the consumer most desperate to act.
You can make an argument that this makes the brochure pointless.
Extra Assets:
This submit was initially printed on No Hacks.
Featured Picture: Roman Samborskyi/Shutterstock
#TextOnly #Model #Web site #Strips #Incorrect #Layer
