In one sampled week of website logs, over 3,000 requests identifying themselves as ChatGPT-User passed our verification check. Hundreds failed. For other requests with the same label, the logs couldn’t give me a verdict at all. Counting the label alone would have put all three groups in the same bucket.
I built an experimental collector to preserve that distinction. The workplace figures came from existing logs. I tested the collector separately on sites I control and compared its results with an independent log baseline.
Recording verification results
The collector records three verdicts: verified, failed and unavailable. A verified result means the request passed the configured vendor check. A failed result means a completed check produced a negative verdict. Unavailable means the collector could not establish a verdict with the evidence or capabilities it had.
An unavailable result includes a reason. The client address might be missing, the vendor evidence might have expired, or an address might fall outside a list whose coverage is not known to be exhaustive. These conditions need different follow-up, even though none supports a positive or negative identity verdict.
Some vendor lists support only positive matches. If the vendor has not established that the list covers every legitimate request, a nonmatch remains inconclusive. Where a negative verdict is supported, it still does not establish malicious intent. A positive infrastructure match also leaves the request's purpose and the identity of the person initiating it unknown.
The workplace logging system called the last group unchecked. The experimental collector records it as unavailable and keeps the reason.
Trusting the client address
An IP-range check depends on receiving the correct client address. Behind a proxy, the application's immediate peer may be a load balancer. A forwarded header can carry the original address, but a sender may also supply that header.
The deployment needs an explicit rule for which proxy hop it trusts and how that hop provides client information. Without that boundary, the collector can end up checking an address supplied by the requester. An accurate vendor list cannot correct an untrusted input.
The evidence used for the check also has a retrieval time, an expiry and a version. Once it expires, the collector records verification as unavailable until fresh evidence is installed. It retains the evidence version with each observation because updates can change the proportion of verified requests. A chart spanning an update needs that context before it can support a claim about changing traffic.
Limiting what leaves the application
The host application classifies and verifies requests locally and excludes human traffic before sending events. Outgoing events omit IP addresses, full user-agent strings, query strings, cookies, raw request headers and bodies. Reviewed, bounded derivatives can be included under the event schema. Paths pass through customer normalisation and bounded sanitisation before transmission.
Eligible live fetches receive individual observations. Failed live claims and crawlers retain aggregate counts, calculated per minute inside the process. Queues, timeouts and retries are bounded so that a reporting outage cannot cause unbounded resource use in the host application. These delivery bounds can mean losing analytics data under pressure; collection must not prevent the application from serving requests.
Checking collection coverage
A collector sees only requests that reach its integration point. An upstream cache may answer before the origin executes. A route may bypass the instrumentation, or the deployed collector may differ from the revision being inspected. An empty chart needs checking against these possibilities.
Controlled requests to the relevant routes and cache states help establish coverage. An independent log baseline can reveal gaps. The findings should describe where collection runs so that readers can interpret the counts without assuming complete visibility.
Response status needs its own integration. Middleware observing a request does not automatically establish the final response. A missing-page finding requires a response-aware collector that observed the 404. That status can justify investigating the path, although it cannot establish that an agent invented the URL or reveal what it intended to find.
What I’d look for in the chart
If the number of verified requests falls, I want to see whether more checks failed, more became unavailable or the collector saw fewer requests in the first place. Each result sends me to a different part of the system. I can investigate a change in vendor evidence or check whether an upstream cache is hiding traffic before I draw a conclusion about how often assistants read the site.