of sites turned away an agent that said it was one (91 of 198)
of those 91 did it before the first page even loaded
average score out of 100 for the 107 sites that let it in
sites scored a perfect 100
Turned away at the door
Our agent runs a normal Chrome browser from a cloud server, and its user agent ends in Botworthy/0.1, so any site can see what it is. It never tries to get past a bot check. 91 of the 198 sites stopped it on every task, and 88 of them did so on the very first page load: the agent never saw a product, a search box or even a cookie banner. Across those sites it logged 425 access-denied or security pages and 40 CAPTCHAs.
Some blocks don't look like blocks. A few sites showed a "this page couldn't load" error or a "we're busier than expected" holding page on every attempt, while every other site in the batch loaded fine. We count those as blocked too, and re-check unclear cases from a second network before listing them.
Share of sites that turned the agent away, by category
Sorted by share. Hover a bar for the numbers.
Show as a table
| Group | Count | Out of |
|---|---|---|
| Retail | 22 | 30 |
| Insurance & Health | 6 | 10 |
| Grocery & Home | 6 | 10 |
| Fashion | 11 | 19 |
| Travel | 11 | 20 |
| Travel & Food | 5 | 10 |
| Cars & Gaming | 5 | 10 |
| Electronics | 4 | 9 |
| Food | 4 | 9 |
| Learning & Media | 3 | 9 |
| Subscriptions | 3 | 10 |
| Apps & Software | 7 | 30 |
| Money | 2 | 10 |
| DTC & Tech | 2 | 10 |
Once it got in
The 107 sites that let the agent in averaged 80/100. Finding a price and finding the fine print were nearly always done. Starting a checkout was the hardest: 14 sites let the agent browse and then blocked it at checkout, and on 11 more it couldn't find its way to the first checkout step at all. Finding a human was the task most often only partly done, usually because the only route was a chatbot or a help center with no contact details.
How each task went on the 107 sites that let the agent in
Each bar is 107 sites.
Show as a table
| Task | Done | Partly | Failed | Blocked |
|---|---|---|---|---|
| Find a price | 97 | 2 | 4 | 4 |
| Search for something | 92 | 9 | 1 | 5 |
| Start a purchase or sign-up | 74 | 8 | 11 | 14 |
| Find a human | 81 | 17 | 2 | 7 |
| Find the fine print | 98 | 6 | 0 | 3 |
What got in the way
Most sites that let the agent in still made it work for it. These are the things it ran into, and how many of the 107 sites had each one at least once.
Sites where the agent ran into each kind of friction
Out of the 107 sites that let it in.
Show as a table
| Group | Count | Out of |
|---|---|---|
| A link, button or page that didn't work | 53 | 107 |
| A cookie banner to get past | 42 | 107 |
| Controls with no name for assistive tech | 38 | 107 |
| A pop-up in the way | 26 | 107 |
| A login wall | 24 | 107 |
| A bot check partway through | 24 | 107 |
| A page that took very long to load | 11 | 107 |
Site owners: our guide to making a website work for AI agents goes through each of these with examples.
Best and worst
The highest and lowest scores among sites that let the agent in. Every result has a replay of each task.
Highest scores
Lowest scores
How we tested
Each site got the same 5 tasks, with at most 15 actions each, on the same day for the whole batch. The agent never logs in, never enters personal data, never pays and never gets around a bot check. Each task is worth 20 points, and friction costs points. The full rules are in how it works.
This is a snapshot from scans finished by 4 Nov 2026. Sites change, and we rescan. All results are open: /sites.json has every site, and each site's page links to its own data and replays.
Ask us to scan your site, or see the full leaderboard.