Autonomous AI agents — ChatGPT Operator, Claude Computer Use, Perplexity, Gemini agents — now browse the web on behalf of users. They follow links, fill forms, complete checkouts and read content the way a human visitor would, but they parse the page through machine-readable signals rather than visual layout. Sites that score well on agent compatibility get more agent-driven traffic, citations and conversions. This guide covers the eight categories our AI Agent Compatibility Audit checks, what each score means, and exactly how to fix the failures.
AI Agent Compatibility measures how well your site exposes the affordances an autonomous agent needs to complete real tasks: discovering pages, understanding their purpose, filling forms, navigating multi-step flows, and reading definitive answers. Unlike SEO (where the goal is ranking) and AEO (where the goal is citation), agent compatibility is about executing actions end-to-end without human translation.
<div> soup.The first thing any AI agent checks is your robots.txt — but the modern AI ecosystem uses crawler tokens you may not be familiar with. Allow or block deliberately; defaulting to "allow all" exposes you to training data harvesting, while blocking everything cuts you off from real-time agent visibility.
User-agent: GPTBot Allow: / Disallow: /admin/ Disallow: /private/ User-agent: ClaudeBot Allow: / User-agent: PerplexityBot Allow: / User-agent: Google-Extended Allow: / User-agent: OAI-SearchBot Allow: / User-agent: ChatGPT-User Allow: /
Agents that find well-formed JSON-LD on a page skip the expensive step of running an LLM over the raw HTML to infer what the page is about. The agent reads the schema, knows the entity type, and acts. Cover at minimum:
Agents parse the DOM, not the visual layout. A page rendered entirely from <div> elements with class names tells the agent nothing. Use <main>, <article>, <nav>, <header>, <footer>, <aside> for layout; <h1>–<h6> for hierarchy; <ul>/<ol>/<li> for lists; <table> with <thead> and <th> for tabular data.
The hardest thing an autonomous agent does is fill a form correctly. Every input needs:
<label for="email">Email address</label>
<input type="email"
id="email"
name="email"
autocomplete="email"
aria-describedby="email-help"
required>
<p id="email-help">We'll send a verification link.</p>
The autocomplete attribute is the agent's strongest hint about what a field expects. Use the standard autocomplete tokens — given-name, family-name, street-address, postal-code, cc-number, etc.
Buttons and links must say what they do. "Click here", "Learn more" and "Read on" tell an agent nothing about the destination. Use action-oriented text: "Start free trial", "Download Q3 report (PDF)", "Compare plans". Keep critical actions — checkout, sign-up, contact — reachable in fewer than 4 clicks from any landing page.
Place the definitive answer to "what is this page about" in the first 200 words. Agents budget tokens; if your answer is buried after 800 words of preamble, they may stop reading before reaching it. Structure information with lists, tables and clear sub-headings.
If your service requires a login, expose SSO options (Google, GitHub, Microsoft, Apple) with the standard sign-in HTML. Agents recognise these patterns and can complete sign-in via the user's credentials. Avoid hCaptcha or fingerprint-based bot detection on key flows — these block legitimate agent traffic indiscriminately.
Publish /llms.txt with a concise summary of what your site does, an AI usage policy URL, and structured contact data. This is the modern equivalent of robots.txt for AI agents — a place where they can read your terms before acting.
Each of the eight categories contributes points to the total. Robots configuration is graded pass/partial/fail. Structured data coverage is measured as a percentage of high-value pages with valid JSON-LD. Semantic quality, form accessibility and action surface are graded across a sampled crawl. The final score is weighted so that the categories most critical for agent task completion — robots, structured data and forms — dominate.
Run a full agent compatibility audit across all eight categories. Get specific fixes for every failure.
Run Agent Compatibility Audit →