Most teams assume their AI-ready page is a text-only mirror: strip the JavaScript, serve markdown, and let the model read. That solves the reading problem. It does not solve the doing problem. Flatten a page to prose and every button, form field and purchase path disappears before the machine arrives. You are not making your site machine-friendly; you are making it a brochure for the user most likely to act.
The doing layer is the new battleground
Since February, one platform has actually shipped an action layer. On August 5, Shopify switched WebMCP tools on for every Liquid storefront by default—catalog search, cart, checkout and policy lookup—with nothing to install. The tools read the same catalog and cart the human storefront uses, so they cannot drift, and merchants pay nothing. Shopify told its August 5 earnings call that AI-driven traffic and orders tripled year over year, but that describes humans arriving from AI answers and buying as usual—not a machine completing checkout. On August 6, the read path worked; the buying path failed with an internal error, so the promise is still maturing.
By contrast, markdown mirrors and readiness scores still answer only ‘what is this page?’ A readiness scanner can verify that WebMCP tools are registered without telling you whether they work. Structured data is useful—JSON-LD sits on 55.6% of measured sites—but it describes content, not actions.
Machines need two surfaces
A website can expose actions to a machine in two ways:
- The floor: semantic HTML and accessibility features. A real button is a button, a label names a field, a form has a target and a method.
- The ceiling: a declared tool surface such as WebMCP, where a page registers the functions an agent may call.
The floor is broken. WebAIM’s 2026 audit of the top million home pages found 95.9% failing WCAG 2, up from 94.8% in 2025, with errors averaging 56.1 per page. Three of the six most common failures are actions going missing: form inputs without labels on 51% of 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 beside it.
Tell the machine what happened
Even a perfectly valid page can fail an agent. If a form only shows a green confirmation rendered for human eyes, the agent may not know the submission succeeded—and it will try again. That pattern produces duplicate orders, tickets and signups, and it is a website feedback problem, not an agent mistake.
The stakes are becoming measurable. A CHI 2026 study ran Anthropic’s Claude Sonnet 4.5 as a computer-use agent on 60 everyday tasks. Success fell from 78.3% under default conditions to 41.7% keyboard-only and 28.3% when the viewport was magnified to 150%. That tested the agent working the way assistive-technology users work, not broken markup.
GEO optimizes citation, not action
Generative engine optimization is sold as SEO for LLMs, and citation absolutely pays today. But GEO only makes a page more describable—something classic SEO already did. It does not address the reason these systems are new: they can act. Every company that can cite you today is already working on a version of itself that does more.
What to fix this week
- Sort what you serve machines into describable and doable. Expect a full first pile and an empty second one.
- Run WAVE on your most important form and count unlabelled inputs, empty links and empty buttons.
- Submit that form and read the response the way a machine would. If the only signal is a green box, an AI will likely struggle.
- Fix the semantic HTML floor before buying a declared tool surface.
Machine-first architecture means structure and actions must work with or without the visual layer. Keep the visual layer for humans, but make the action layer readable by the agent that is ready to buy.
Source: Search Engine Journal
