A user-agent and IP gate, deployed across several properties, was blocking: generic HTTP clients (python-requests, Go-http-client, curl, empty user-agents), shared cloud IP ranges including the GCP ranges where a major AI retrieval agent operates, and an AWS crawler — which was denied /robots.txt itself. On removal, error rates on one property dropped from 13 403s and 11 401s in a day to 2 and 0.
The site owner's own IndexNow submission script runs on python-requests. On the day the gate was removed, that script received a clean 200. The day before, it would have been blocked by its own site. A classifier that cannot distinguish its owner's tooling from an attacker is not a security control.
Nothing measurable. Citation performance did not drop. Rank-tracker sweeps resumed — and those are free telemetry, since a tracker crawling a page is confirmation that page is being tracked. The correct security boundary is not running exploitable software; it is not refusing to answer HTTP requests from user-agents you dislike.