AI agent readiness
AI agents find buttons, links and form fields through the same accessibility tree that screen readers use. We measure how well they can use your site and show you where they get stuck.
The same product page, on the screen and as an agent reads it
A product page with a search box, a product photo, the product name, an Add to cart button and a heart-shaped button with no label.
- Search box: Search
- Image: Product photo
- Heading: Product name
- Button: Add to cart
- Button: No name. An agent cannot use this button.
Button 5 has no name. An agent that reads the accessibility tree cannot tell what it does, so it cannot press it.
Part A of our Agent Readiness Index checks that every link, button and form field has a name and the right role. It carries 35 of the 100 points.
Why agents need the same fixes
A button with a proper name works for a blind customer and for an agent shopping on someone's behalf. Two sources show why.
ChatGPT Atlas reads ARIA roles and labels, the ones screen readers use, to understand a page. OpenAI asks site owners to follow WAI-ARIA.
AI crawlers such as GPTBot and ClaudeBot do not run JavaScript, so text that appears only after scripts run never reaches them.
Source: Vercel and MERJ, December 2024
How the score works
Five parts, weighted out of 100. The index is our own method, and we never present it as a standard or a certification. We always show the score with each part's sub-score and the evidence behind it.
| Part | Weight | What we measure | Passes when |
|---|---|---|---|
| A. Names and roles | 35 | Links, buttons and form fields have a name and the right role. Landmarks are present, with one h1 and no skipped heading levels. | Every interactive element on the key templates has a name, and a main landmark exists |
| B. Content without JavaScript | 20 | The text a visitor sees, compared with the text a request with JavaScript switched off gets back. | At least 80% of the main content arrives without JavaScript |
| C. Crawler and agent access | 15 | robots.txt rules for named AI user agents, error or challenge pages served to them, the XML sitemap, canonical tags and hreflang. | Search crawlers and agents acting for a user are let in, and the sitemap is valid |
| D. Structured data | 15 | Organization data, plus Product and Offer for shops or SoftwareApplication for SaaS, plus BreadcrumbList, checked against the visible text. | The data is valid and does not contradict the page, for example with a different price |
| E. Task success | 15 | An agent tries three tasks we agree with you, three attempts each. | At least 7 of the 9 attempts succeed |
The task test
You choose three tasks
For example, finding a product's price or adding it to the cart.
An agent tries each one three times
It sees only the page's accessibility snapshot and can act only through an element's role and name, so a cart button with no name is a button it cannot find.
The report shows where it failed
The exact step of each failed attempt, and the WCAG criterion behind it.
The agent stops before any payment or step that cannot be undone, uses test accounts only, and treats every word on the page as untrusted, because a page can carry hidden instructions aimed at AI models.
What you get
- Your score out of 100, with the sub-score for each of the five parts.
- The evidence behind every sub-score.
- A list of fixes ranked by effort, each tied to the WCAG criterion it falls under.
Other services
Accessibility audits
We test your key templates and user flows against WCAG 2.2 AA, and report every issue with evidence and a fix.
Fix support
Code examples, reviews of your pull requests and a weekly call while your developers work through the report.
Re-tests
When the fixes are live, we test every reported issue again and update the report.
VPAT and ACR reports
An Accessibility Conformance Report on the VPAT template, written from the evidence of an audit of your product.
EU accessibility statements
The accessibility information the European Accessibility Act asks you to publish, written from a real audit.