Skip to main content

Test how AI agents get on with your site

Agents now search, compare and buy on behalf of people. They read the markup, not the design, and a button with no label or a form that only works with a mouse is a wall. An actor is an agent in a real browser, so a run shows you the step where an agent gives up.

Agentic user journey
  1. 01An assistant is sent a task
  2. 02Your homepage
  3. 03The pricing page
  4. 04The contact form

How it works

Captures on this page are from a workspace testing Parlour, a demo app simulating a restaurant booking flow.

  1. 01

    Give an agent a job

    Write the tasks an assistant would be sent to do on your site: find a price, compare two plans, book a slot, get in touch.

    The checklists in a Stunt Double project, each with the result of its last run.
    Checklists, each with the result of its last run.
  2. 02

    Run it and read the stall

    Each task comes back passed or failed with the screenshot of the moment, and the reasoning shows what the agent failed to find or operate.

    A checklist in Stunt Double with its run history, one row per run.
    A checklist and its run history, one row per run.
  3. 03

    Fix the page, not the prompt

    Label the control, move the price out of the image, make the form work from a keyboard. Then run the same tasks again.

    A failed check raised as an issue in Stunt Double: the deposit refund window is easy to miss, with its evidence attached.
    A failed check, raised as an issue with its evidence.

The requirements you would write

Plain language, one per line. An actor checks each one in a real browser.

  1. 01An agent can find the price of the standard plan from the homepage
  2. 02An agent can complete the contact form
  3. 03An agent can find the returns policy in two steps or fewer
  4. 04The site publishes an llms.txt file
A completed checklist run in Stunt Double: three checks on a booking flow, two passed and one failed, with a weighted score of 72%.
A checklist run: each requirement comes back passed or failed, with the screenshot the actor took.

What comes back

  • 01

    Each check passed or failed, with the screenshot and the reasoning behind it

  • 02

    A recording of the whole session, step by step

  • 03

    Insights for anything a check did not ask about

Where it stops

An actor is one agent. Assistants read pages in different ways, so treat a pass as a strong sign rather than a guarantee for every one of them.

Questions people ask

How is this different from an accessibility audit?

They overlap, because agents depend on the same labels and keyboard access that assistive technology does. An audit checks the page against a standard. An actor attempts the task and shows you whether it got done.

What is the Stunt Double Index?

A public ranking of how AI agents get on with real websites. It is free to browse at index.stuntdouble.io.

Keep reading

Go further

Related terms

Other use cases

Try it on your own site

Give an actor your URL. It opens a real browser, works through the site, and shows you every step it took.

Free to try, no sign-up needed.