Shuyang Zhang measures what web tools actually receive when a server reports success. Across 10,307 URLs collected over three days, many HTTP 200 responses contain application frames, bot checks or substitutes for the requested page.
- Automatic labels classify 20.9% of successful plain-request page responses as unusable. Review lowers the corresponding test-split estimate from 25.5% to 20.0%.
- The review is mainly model-based: an assistant reviews 300 examples and the author checks 60. The author’s independent blind pass agrees on usability for 47 of 57 decidable examples; later discussion retains all reviewed usability labels.
- Combined text-length, redirect and nonexistent-page checks catch 86% of unusable reviewed responses, with a measured 5.6% false-alarm rate. Excluding nonexistent comparison pages lowers detection to 79%.
- Treating every 200 as usable inflates apparent citation drift, where a source seems to stop supporting its claim, from 5.8% to 21.9% in the study’s cited-source analysis.
The practical change is to record whether the requested content was obtained before checking its support for a claim. Browser rendering also needs content checks: it sometimes fails where the plain request works. The paper promises data and code rather than linking an available release.
Caveat: Most labels came from models, collection used one network location, and Appendix C discloses changes made after results were visible. These are author-reported measurements, not universal web-failure rates.