Free Agent Readability Tool
Accessibility Tree Audit -- See Your Site the Way AI Agents Read It
AI agents do not look at your website. They read a machine version of it called the accessibility tree: the browser's stripped down map of roles, names, and states. Google now documents it as one of the three ways agents perceive a page, alongside raw HTML and screenshots. Enter a URL below to score how readable your site is to the agents already browsing on behalf of your customers. Free, no signup required.
How AI Agents Actually See Your Page
Google documents three modalities that agents use to perceive a web page. Understanding all three helps you fix the right layer.
- Screenshots: Vision models render the page as an image. This is slow and token expensive, so agents use it as a fallback when the structured layers are not available or are ambiguous.
- Raw HTML: The DOM structure and element relationships. Agents read it to understand the page, but it is verbose and includes layout and styling noise that is not relevant to interaction.
- The accessibility tree: The semantic summary: roles, names, and states of interactive elements. Originally built for screen readers, it is now the map agents lean on for reliability and speed. It strips away everything that is not semantic.
Check It Yourself in Chrome
You do not need a tool to see the accessibility tree. Chrome DevTools shows it directly.
- Right click any page and choose Inspect.
- In the Elements panel, open the Accessibility tab. If you do not see it, click the double arrow to reveal hidden tabs.
- Check "Enable full-page accessibility tree" and reload DevTools when prompted.
- Click the accessibility icon that appears at the top right of the Elements tree. You are now reading your page the way an agent does: every node is a role and a name. If you see unnamed images, generic links, or buttons with no label, so does the agent.
What the Score Measures
The audit scores 100 points across six categories. Each category starts at full points and deducts per issue found, with a floor of zero per category.
- Names and labels (30 points): Links with no accessible name, buttons with no label, form inputs with no associated label, and images missing the alt attribute entirely. Agents cannot operate or describe elements without names.
- Semantic interactivity (20 points): onclick handlers on div, span, li, or img without both a role and tabindex, elements with an interactive role but no tabindex, and anchors without href. Agents expect native interactive elements.
- Landmarks and structure (20 points): Missing or duplicate main landmark, missing nav, missing or duplicate h1, skipped heading levels, missing html lang attribute, and empty or missing title. Structure tells agents what each region is.
- Link text quality (10 points): Generic anchor text like click here or read more, raw URLs as anchor text, and identical anchor text pointing to different destinations. Agents navigate by link text.
- Rendered content signal (10 points): Body text content under a minimum threshold with a single empty framework root div, and missing meta description. Agents that do not run JavaScript see a blank page.
- ARIA hygiene (10 points): aria-hidden on containers with interactive elements, redundant roles on native elements, positive tabindex values, and aria-labelledby or label[for] referencing ids that do not exist.
Why This Matters Now
AI agents are already browsing on behalf of your customers. The accessibility tree is how they read your site, and most sites are hard to read.
- Three modalities: Google's official guidance names the accessibility tree as one of three primary ways agents view a site, and its closing recommendations are to audit your accessibility tree and to read about WebMCP.
- Nearly half: WebAIM's survey of the top one million home pages finds empty links on nearly half of them and empty buttons on almost a third. Those are dead ends for an agent.
- 78 to 42 percent: A 2026 study from UC Berkeley and the University of Michigan found agent task success fell from 78 percent to 42 percent when the accessibility tree was degraded.
The Read Layer and the Act Layer
The accessibility tree is how agents read and navigate your site. WebMCP is how your site hands agents structured tools to act with. Google recommends both. Most of the web has neither. Fixing the read layer makes every agent visit work today. Adding the act layer makes your site a first class citizen when WebMCP capable browsers roll out. Run both audits and you have the full picture.
API Documentation
The audit tool is available as a public REST API. Send a POST request to /functions/v1/agent-readability-audit with a JSON body containing a "url" field and an optional "source" field. The API is open to everyone -- no authentication required.
Request Format
POST with Content-Type: application/json. Required field: url (the website URL to audit). Optional field: source (one of "api", "website", or "webmcp", defaults to "api").
Response Fields
- url: The normalized URL that was audited
- readability_score: A score from 0 to 100 indicating agent readability
- grade: A letter grade: A, B, C, D, or F
- webmcp_detected: Whether navigator.modelContext references were found in the page
- findings: Points and issues per category, keyed by category name
- recommendations: Ranked next steps ordered by impact
- fix_prompt: A complete prompt to paste into an AI coding assistant
- checked_at: ISO timestamp of the audit
Rate Limiting
Maximum 10 audits per IP address per hour. Results for the same URL are cached for 24 hours and returned instantly. Rate-limited requests receive HTTP 429 with a retry_after_seconds field.
Frequently Asked Questions
Is this a WCAG compliance audit?
No. It checks the subset of accessibility signals that AI agents depend on to read and operate a page: names, roles, structure, and link clarity. Fixing them also helps screen reader users, but this is not a legal compliance review.
What is the accessibility tree?
A simplified model of your page the browser builds automatically from your HTML. It lists every element with a role, a name, and a state. Screen readers have used it for two decades. AI agents now read it too.
Do AI agents run JavaScript?
Agent browsers render pages, but many AI crawlers do not. If your content only exists after JavaScript runs, part of your audience sees a blank page. Server rendered content is the safe baseline.
How is this different from your WebMCP audit?
The WebMCP audit measures whether your site is ready to expose callable tools to agents. This audit measures whether agents can read and navigate the site you already have. Read layer and act layer. Run both.
Can a page look perfect to humans and still fail this audit?
Yes. Divs styled as buttons, icon buttons with no label, images with no alt text, and headings that are just bold text all look fine on screen and read as noise in the tree.
My score is low. What now?
Copy the generated fix prompt into Claude, Cursor, or any AI coding assistant and it will walk through the exact issues found on your page. Or have us do it: we fix the read layer and the act layer together.
Need a Full Implementation?
The audit tells you what agents cannot read. For a full implementation, semantic rebuild of key templates, labeling pass, landmark structure, and verification against the rendered tree, Adapt Marketing can help. Complete source code, documentation, and training. No recurring fees.