AI agent readiness audit: can AI agents use your website?
Axpecto's AI Usability audit tests whether an AI agent can find its way around your website and complete tasks on it. A headless browser crawls up to 25 pages and scores how easily an agent can navigate, understand and act on them. When you are signed in, a real AI agent then attempts two or three tasks drawn from what your site offers and is judged on the page's own evidence.
The crawl is free and needs no account. It is deterministic: no AI model is involved, so the same site gives the same result.
Last updated
How do I check if an AI agent gets confused by my website?
Run the audit on your homepage. It reports where an agent gets stuck, page by page, with the element that caused it. The seven dimensions it scores:
| Dimension | What it checks |
|---|---|
| Navigation | Whether key pages can be reached through real, named links |
| Interaction | Whether buttons, links and form fields have names an agent can read |
| Semantic clarity | Whether headings, labels and page structure say what each part is |
| Discoverability | Whether the main actions (search, sign up, contact, buy) can be found |
| Interaction efficiency | How many steps a task takes, observed when a real agent runs it |
| Resilience | What happens on errors: broken links, missing pages, failed forms |
| Predictability | Whether the same thing behaves the same way across the site |
What does the real AI agent do?
- It is given two or three tasks a visitor would really do on your site: find a page, find an answer, search, or fill in a form or cart.
- The main task is run twice, because one run can be luck.
- Each task is judged by what the page shows, never by what the agent says it did.
- A guard in the page stops it sending a form, paying or placing an order.
- Task results make up 60% of the score and the seven dimensions the other 40%, so a site an agent can't use never reads as good.
What steps improve a website's usability for AI agents?
- Give every icon-only button a text name (aria-label), so an agent knows what it does.
- Use real links with real URLs for navigation, not clickable divs that run JavaScript.
- Label every form field with a visible label tied to it, not only placeholder text.
- Serve the main content in the HTML, so assistants that don't run JavaScript can still read it.
- Keep one clear primary action per page, named the same way everywhere.
- Return a real 404 page with links back, and fix broken internal links.
- Don't block the AI crawlers and agents you want visiting in robots.txt or bot protection.
Can an AI agent test a specific task, like checkout?
Yes, with AI Task Completion on paid plans. You describe the task, for example "add the cheapest shirt to the cart and reach checkout", optionally say how success will look, and a real AI agent drives a browser to attempt it. The verdict (Success, Blocked or Unverified) comes from evidence recorded during the run, with a step-by-step replay.