In February 2026, industry analysts observed a shift in how websites prepared for the rise of artificial intelligence agents: the adoption of markdown-based mirrors. While this trend was initially hailed as a breakthrough in helping AI "read" the web, seven months of implementation have revealed a significant, structural flaw. By prioritizing content readability, many platforms have inadvertently discarded the "doing" problem—the ability for an AI to execute tasks on behalf of a user. The current trend of serving text-only versions to AI agents does not merely simplify the web; it surgically removes the interactive components that allow machines to navigate, purchase, or perform services, rendering these sites effectively inert for agentic workflows.
The Evolution of the Machine-Readable Web
The move toward markdown mirrors, readiness scores, and Generative Engine Optimization (GEO) was intended to bridge the gap between human-centric design and machine consumption. However, the current execution is inherently backward. A web page served to an AI as a markdown file is stripped of its JavaScript-heavy visual layer, which—while unnecessary for a machine—is often where the programmatic hooks for interaction reside. When a website is reduced to plain prose for an LLM, the buttons, forms, and navigational triggers that a user might interact with vanish. Consequently, an AI agent tasked with, for example, renewing a subscription or completing a purchase, finds itself staring at a text-only brochure rather than a functional interface.
This oversight is compounded by the fact that the structural layer of the web—the underlying HTML—is already capable of communicating with machines. Structured data, such as JSON-LD, is currently present on approximately 55.6% of websites measured by W3Techs as of September 2026. This data serves as a machine-only surface that persists regardless of the visual layout. By prioritizing "markdown mirrors" over existing semantic standards, developers are effectively building a second, inferior layer of information that offers no mechanism for action.
The Accessibility Crisis and the Failure of Markup
The struggle to make the web "agent-ready" is inextricably linked to the ongoing failure to make it accessible to humans. According to WebAIM’s 2026 evaluation of the top one million home pages, 95.9% of sites fail to meet Web Content Accessibility Guidelines (WCAG) 2, a figure that marks a regression from 94.8% in 2025. This downward trend is particularly concerning given that the "accessibility tree"—the programmatic representation of a page used by screen readers—is the exact same architecture that AI agents use to interpret a site’s functionality.
Common failures, such as form inputs lacking labels (found on 51% of home pages), empty links (46.3%), and empty buttons (30.6%), are not just accessibility hurdles; they are fatal errors for AI agents. A button without a descriptive name is invisible to an AI’s navigation logic. A 2026 study accepted to the CHI conference demonstrated this, showing that when Anthropic’s Claude Sonnet 4.5 was tasked with 60 everyday computer-use operations, its success rate plummeted from 78.3% under standard conditions to 28.3% when the viewport was magnified, simulating the navigation challenges faced by those relying on assistive technology. When the markup is broken, the agent is effectively blinded.
The Feedback Loop: Why AI Performs Double Tasks
One of the most persistent issues in agentic automation is the lack of "success feedback." In many deployments, AI agents attempt to submit forms repeatedly because the page provides no programmatic confirmation that the initial attempt was successful. For a human, a "Thank You" message or a redirect is visually obvious. For an AI, if the success feedback is rendered only for human eyes and lacks semantic structure, the machine perceives the request as still pending.
This results in "ghost" errors—duplicate orders, redundant signups, and double-billing. These errors are not failures of the AI’s intelligence, but failures of the website’s "read-back" mechanism. As the industry moves toward more agent-driven commerce, the responsibility falls on web developers to ensure that after a machine performs an action, the site communicates the result in a machine-readable format.
The Shopify Model: A Shift Toward Declared Tool Surfaces
The industry saw a potential turning point on August 5, 2026, when Shopify implemented WebMCP (Model Context Protocol) tools for every storefront built on its Liquid framework. This initiative, which automates the exposure of catalog searches, cart functions, and checkout processes, represents the first time a major platform has provided a "declared tool surface" at scale.
By serving a standardized adapter script from its content delivery network (CDN), Shopify enabled its merchants to provide a functional API to AI agents without requiring individual store owners to write a single line of code. This model solves the "doing" problem by providing a formal pathway for machines to interact with store functions. Shopify reported that AI-driven traffic and orders tripled year-over-year in their August 2026 earnings call, suggesting that the integration of functional tool surfaces is a viable path forward. However, this also highlights a broader concern: the infrastructure for the "Agentic Web" is currently being built by centralized platforms on their own terms, often without the explicit input or understanding of the smaller merchants relying on them.
GEO vs. Action: The Misalignment of Objectives
Generative Engine Optimization (GEO) has become the dominant discipline for firms seeking visibility in AI-generated answers. While GEO is essential for driving human traffic—as citation remains the primary currency of the current digital economy—it is fundamentally decoupled from the action-oriented future of the web. GEO focuses on how a page is described; it does not address how a page is used.
Critics argue that by focusing exclusively on GEO, businesses are optimizing for the "half" of the web that cannot act. This strategy creates a paradox where a website is designed to be highly discoverable by an AI, but once the AI arrives, it finds no tools to execute the user’s intent. As AI agents evolve from research bots into "do-ers"—capable of navigating complex browser environments—this strategy is likely to become a strategic liability.
Conclusion: A Call for Machine-First Architecture
The current reliance on text-only versions is an unsustainable stop-gap measure. To build a truly functional web, developers must adopt a "Machine-First Architecture." This approach treats the structural layer and the action surface as the primary requirements, leaving the visual layer as an optional, albeit necessary, add-on for human users.
The "floor" of this architecture is semantic HTML. When buttons are labeled, inputs are correctly associated with their functions, and forms provide machine-readable feedback, the web becomes accessible to everyone—both humans relying on assistive technology and the next generation of AI agents. Until the industry shifts its focus from merely mirroring content to exposing functionality, the promise of the agentic web will remain unfulfilled, and the "doing" problem will continue to haunt the digital infrastructure of the future. The challenge for 2027 and beyond is not just to make the web readable, but to make it programmable by default.
