Why do browser agents get blocked by websites?
In our benchmark, 10 of 21 sites blocked a cloud-style browser agent. The agent ran Browser Use with Claude Sonnet 5.5 in a fresh headless Chromium with an empty profile. Sites such as bahn.de, Zalando and Walmart stopped it with bot protection or checks. The same tasks, run as Wield recipes in an everyday browser, passed on 7 of those 10 sites. These are results from our runs, as of 7 November 2026.
What a site sees from a cloud browser agent
A cloud browser agent starts a new browser for each task. It has no cookies, no history and no earlier visits. The benchmark's browser-agent side used exactly that setup, with a fresh headless Chromium and an empty profile, because that is how a cloud agent runs.
Sites decide which visitors get their data. In our runs, Akamai blocked curl and every automated browser on bahn.de's timetable pages. AWS WAF challenged the automated browser on ImmoScout24 and served an HTML check instead of the JSON results. We do not know the exact rules each vendor uses, and we do not claim to.
Which sites blocked our benchmark agent
The table lists the 10 sites where the browser agent was flagged as blocked. It also shows how the Wield recipe for the same site did in our runs.
| Site | Browser agent blocked | Browser agent success | Wield recipe success |
|---|---|---|---|
| bahn.de | Yes | 0% | 100% |
| Lieferando | Yes | 0% | 100% |
| Zalando | Yes | 0% | 0% |
| ImmoScout24 | Yes | 0% | 0% |
| Idealo | Yes | 0% | 100% |
| Eventim | Yes | 0% | 100% |
| mobile.de | Yes | 0% | 56% (5 of 9) |
| Walmart | Yes | 0% | 100% |
| Zillow | Yes | 0% | 100% |
| OpenTable | Yes | 0% | 100% |
Wield's own misses are part of this data. The Zalando recipe failed in all 9 benchmark runs. The ImmoScout24 recipe failed in its 2 runs, and the AWS WAF check on that site was not retried. Treat the blocked list as a snapshot. It is preliminary and the benchmark is still running.
Why the same task works in your own browser
Wield runs each recipe in a tab of the site, inside your own Chromium browser. Chrome, Edge, Brave, Arc, Vivaldi, Opera and Chromium are supported. The request carries the session you already have. No model runs during a recipe, so the browser sends only the few requests the site's own page makes, at a human pace.
A recipe passing does not mean the site allows it. Each site has its own terms. Wield's public recipes are reviewed against each site's terms, and every site page states the recipe's known restrictions in plain words.
What Wield does when a check appears
When a site shows a check, such as a CAPTCHA or an "are you human" page, Wield stops with site_challenge. It does not solve the check. The extension retries once by itself. If the check is still there, you open the site in your own tab, use it once, and solve the check if one appears. Then you ask your agent again.
FAQ
Why does a site block one agent but not my browser?
We cannot say why each site decides. We can report what we saw. The blocked runs used a fresh headless browser with an empty profile. The Wield runs used a normal browser that already had a session with the site.
Does Wield get past blocks?
No. Wield does not solve checks, and it does not try to get past bot protection. It stops at the first check and asks you to handle it in your own tab. The benchmark did not retry ImmoScout24 after its AWS WAF check, for the same reason.
Will a browser agent be blocked on every site?
No. In our runs, 10 of 21 sites blocked the cloud-style agent. Most of the others did not flag a block. The results vary by site and by date, so the numbers are not a general rule.
Where do the numbers come from?
The full method, the per-task rows and the raw logs are on the benchmark page. The logs hold destinations, timing, token counts and check results. They do not hold page contents.