The number is harder to ignore each month: 9.5 billion web visits across generative AI platforms in the 12 months ending May 2026, up 70% year-over-year, with 655 million unique visitors. AI agents — systems that browse, evaluate, and act on behalf of users — are moving from prototype to mainstream.
Yet a Rankability analysis of the top 1,000 websites found that only 8.7% publish an llms.txt file, the most basic signal that a site is ready for programmatic AI access. The gap between where AI traffic is heading and where most websites are positioned is widening fast.
This post breaks down the technical stack that actually matters for agent-readable websites: what llms.txt does (and what it doesn’t), what WebMCP changes when it ships, the schema property most practitioners skip, and the four-part framework Google uses internally to evaluate agent accessibility.
What llms.txt Actually Does (And the Adoption Trap to Avoid)
llms.txt is a markdown file placed at your domain root that gives AI systems a curated map of your site’s content and capabilities — analogous to how sitemap.xml signals structure to search crawlers. When an AI agent arrives at your domain, a properly formatted llms.txt tells it which pages are authoritative, what the site covers, and which sections to prioritize.
The adoption curve is instructive. An SE Ranking study of 300,000 domains found 10.13% overall adoption as of 2026, up from roughly 1% in mid-2025 — 5.4x growth in 12 months. The outlier in the data: Shopify, sitting at 78.1% adoption, but only because Shopify silently pushed the file to every store by default in late April 2026 with no opt-in or announcement. WordPress sits at 8.7%, which reflects actual deliberate decisions.
What llms.txt does not do: Google has explicitly confirmed it has no effect on rankings or visibility in Google Search or AI Overviews. Its intended audience is programmatic AI access — agents and crawlers that consume site structure directly, not search engines making indexing decisions. Publishing one expecting Google citation gains is solving the wrong problem.
What it does do: it reduces the inference burden on AI agents trying to navigate large sites. A 15,000-page site without a structured content map forces agents to parse the DOM on every page to understand priority. An llms.txt file with clearly labeled sections — docs, pricing, comparison pages, key product pages — gives agents a direct navigation path and cuts traversal time significantly.
WebMCP: The Standard That Shifts Agent Interaction from Scraping to API
If llms.txt is the passive signal, WebMCP is the active protocol. Announced at Google I/O 2026 and currently entering a Chrome 149 origin trial, WebMCP is a proposed web standard that lets websites expose structured tools directly to browser-based AI agents — rather than requiring agents to simulate mouse clicks and parse visual layouts.
The practical shift: instead of an agent trying to infer how to “add to cart” by locating a styled div that looks like a button, WebMCP lets the website declare: here is a tool called addToCart, here are its parameters, here is the authentication it requires. The agent invokes the tool directly.
Booking.com, Shopify, Instacart, and Intuit have committed to implement WebMCP — a strong signal that transactional sites view agent-accessible capabilities as a near-term commercial priority. For e-commerce and SaaS sites, agent-completion rate (how often an agent can successfully complete a booking, signup, or purchase) is becoming a new performance category alongside conversion rate.
WebMCP is not production-ready for most sites yet — the Chrome 149 origin trial means early adopters can experiment before it becomes a stable spec. But the implementation pattern is becoming visible: map your site’s key user actions, define them as structured tools with explicit parameters, and expose them through the WebMCP interface.
The Schema Property Most Site Owners Skip: potentialAction
Of the structured data types that affect agent readability, potentialAction is consistently underemphasized in implementation guides. Organization, Article, FAQPage, and Product schema get covered. potentialAction does not.
potentialAction tells agents not just what your site is, but what can be done there. A SaaS site implementing it correctly exposes:
{
"@type": "SoftwareApplication",
"potentialAction": {
"@type": "RegisterAction",
"target": "https://yourdomain.com/signup"
}
}This is the schema equivalent of WebMCP’s tool declaration — it tells agents which actions are available and where to invoke them. For agents making “should I complete this task here or on a competitor’s site” decisions, potentialAction is the clearest machine-readable signal of capability.
At minimum, implement potentialAction for your primary conversion actions: signup, trial start, product search, contact, or booking. Agents evaluating multiple options will favor sites that explicitly declare their action endpoints over sites where that capability has to be inferred from DOM structure.
The CARE Framework: Four Tests for Agent Accessibility
Google’s internal guidance on agent-accessible sites organizes around four criteria, documented in Agentic AI Optimization (AAIO) research as the CARE framework:
Crawlability — Can agents reach your content? This means no bot-blocking rules that exclude AI crawlers, server response times under 3 seconds, no JavaScript-required rendering for key content, and robots.txt rules that distinguish between AI agents you want and scrapers you don’t. You can allow specific agent user-agent strings while blocking general crawlers.
Accessibility — Can agents find what matters? Google specifies using semantic HTML elements — button and anchor tags rather than styled divs, label tags linked to inputs, cursor: pointer on clickable elements. The accessibility tree — the structured representation screen readers use — is one of three input modalities agents use to parse pages. Sites built with semantic HTML score measurably better on accessibility tree parsing than div-heavy layouts.
Readability — Can agents parse your content? This overlaps with standard GEO best practices: standalone answer paragraphs, explicit headings that describe section content, FAQ schema on question-and-answer content. Agents operating across multiple tabs simultaneously favor content that provides direct answers without requiring inference from earlier paragraphs.
Executability — Can agents complete tasks? This is where potentialAction, WebMCP, and stable page layouts intersect. Stable layouts (elements that don’t shift after initial load) reduce agent error rates significantly. Forms that accept programmatic input without CAPTCHA work better for agent completion. Checkout flows with clear single-action steps per page reduce the number of DOM traversals required.
The Five-Item Audit You Can Run Today
Based on the CARE framework and current adoption data, the practical starting point covers five items:
1. Publish llms.txt. A minimal version takes under an hour. List your five most important pages, your main content categories, and a one-paragraph description of what your site covers. Place it at yourdomain.com/llms.txt. This is low-effort, high-signal infrastructure that puts you ahead of 91% of sites at equivalent traffic levels.
2. Add potentialAction to your primary schema. If you have Organization schema (you should), extend it with potentialAction entries for your top conversion actions. This is a 30-minute implementation that explicitly signals capability to agents evaluating your site.
3. Audit semantic HTML on key pages. Run your main landing page and product page through Chrome’s Accessibility tab. Identify elements that should be buttons or links but are styled divs. Fix the most critical ones. Agents read the accessibility tree; visual design is secondary to structural signal.
4. Review your robots.txt for agent-specific rules. Check whether your rules block crawlers that identify as AI agents. If you want agent traffic, explicitly allow specific user-agent strings. Blanket bot-blocking rules written before 2025 frequently exclude agents you’d want.
5. Monitor the WebMCP origin trial. If you run a transactional site, start mapping your key user actions now. The spec is evolving through Chrome 149’s origin trial. Sites that participate in early testing will have production-ready implementations before the spec locks and before competitors know it exists.
Two Tracks, Not One
The overlap between sites that have implemented GEO best practices and sites that have implemented agent-accessibility measures is small. Most GEO work has focused on citation optimization for passive AI answers — being mentioned in ChatGPT or Perplexity. Agentic optimization focuses on a different outcome: being the site an AI agent can actually complete tasks on.
These two tracks serve different AI behaviors. Passive AI answers benefit from citation tactics: structured content, FAQ schema, cross-platform brand mentions. Agentic AI — an AI assistant that browses, evaluates options, and completes a purchase — benefits from agent-accessibility infrastructure: llms.txt, potentialAction, semantic HTML, WebMCP.
Both are now worth building in parallel. Citations drive awareness. Agent-ready infrastructure converts the agents that arrive.
Conclusion
The 8.7% of top websites publishing llms.txt aren’t necessarily seeing dramatic traffic gains yet — the standard is still early, and WebMCP is pre-production. What they’re building is position: infrastructure that costs relatively little to implement now and will matter more as agent traffic volumes compound.
The three-layer stack — llms.txt for site navigation, potentialAction schema for action declaration, and WebMCP (when it ships) for full tool exposure — is becoming the technical baseline for sites that want to participate in the agentic web rather than just be scraped by it.
To see how current AI systems are actually navigating and citing your site, the LLMagnet AI Visibility Analyzer shows your citation footprint across ChatGPT, Perplexity, Google AI Overviews, and Claude, and flags the technical gaps limiting your visibility.