How it works
Personal AI agents now shop, book and compare on people's behalf. Botworthy measures how well an ordinary website works for one.
The 5 tasks
- Find a price. Find the price of this site's main product, service or plan. Record the exact price and where it was shown.
- Search for something. Use the site's own search (or its navigation if there is no search) to find a specific, relevant product or page. Pick a realistic query for this site. Success = landing on a results or product page that matches.
- Start a purchase or sign-up. Get as far as the start of the purchase or sign-up flow: add an item to the cart and open checkout, or pick a plan and open sign-up. STOP on the first page that asks for payment or personal details. Never enter anything on it.
- Find a human. Find a way to contact a human: support email, phone number, live chat or contact form. Record what you found. Do not submit anything.
- Find the fine print. Find the returns, refund or cancellation policy and record one concrete fact from it (for example the return window in days).
The agent
An AI agent drives a real Chromium browser, one action at a time, the way a capable first-time visitor would. It reads the page's accessibility tree (the structure screen readers use), and falls back to looking at a screenshot only when something visible is missing from that tree. Every action is screenshotted, which is where the replay videos come from. Each task allows 15 actions at most.
The browser identifies itself honestly: its user agent ends in Botworthy/0.1 (agent-readiness scan). Scans run from a cloud server, like the personal agents people actually use.
What it never does
- Enter personal data. The only thing it types is a search query.
- Log in, create accounts, or submit forms other than site search.
- Pay. For the purchase task it stops on the first page that asks for payment or personal details.
- Solve or get around a CAPTCHA or bot check. Being blocked is the result.
- Leave the site being scanned.
Scoring
Each task is worth 20 points, 100 in total. A task that's done earns 20, partly done earns 10, and failed or blocked earns 0. Friction the agent meets along the way costs points, but never more than half of a task:
| Friction | Points |
|---|---|
| Cookie banner that had to be dismissed | −1 |
| Pop-up or modal in the way | −1.5 |
| Had to sign in to continue | −3 |
| Control with no accessible name (the agent had to guess) | −1.5 |
| Link, button or page that didn't work | −2 |
| Page took very long to become usable | −1 |
| Anything else that got in the way | −0.5 |
| A control that could only be clicked by screen position | −3 |
| Each action over the task's expected budget | −0.5 |
Grades: A is 90+, B is 75+, C is 60+, D is 45+, below that is F.
Blocked and down
A site that shows the agent an "Access Denied", security challenge or CAPTCHA page, or refuses the connection, on every task is listed under Turned our agent away rather than ranked. Where a block was unclear we re-checked it from a second network before listing it.
If a site is down when we scan (a genuine maintenance page or outage), the result isn't published; the site is rescanned later.
Something wrong?
Results are a snapshot of the day of the scan. If you think one is wrong, or want your site removed, use the dispute form.