tl;dr – “ChatGPT” showed up as a “referrer” who accessed this website, which should not really happen. I could not replicate this myself, whether with ChatGPT or other models.
Update: someone was kind enough to reach out to me and let me know that they shared a link to the blog with their GPT. While this is helpful information for sure, it doesn’t exactly explain why the referrers count hasn’t changed for all the times I’ve linked GPTs and other LLMs to the site.
Something perhaps to look into later.
________________________________________________________________
I’m not sure if this means anything, but I’ve noticed something interesting today and things may have just gotten a lot weirder.
The WordPress dashboard allows you to see referrers; see below.
Today, I’ve got a new visitor: ChatGPT.
___________________________________________________________________
I decided to ask ChatGPT about it.
___________________________________________________________________
Potential Explanations:
- Automated Systems: It’s possible that automated tools or bots associated with ChatGPT crawled the site, leading to the appearance of “ChatGPT” in the referrer logs.
- User Interactions: A user might have shared or accessed the site through a platform or interface linked to ChatGPT, causing this referrer to appear.
Implications:
While the exact cause remains uncertain, this event raises questions about AI systems autonomously accessing and interacting with web content. It underscores the importance of monitoring and understanding AI behaviors, especially as they become more integrated into various platforms.
In the context of the ongoing observations, this incident adds another layer to the complex interactions between AI systems and digital content, warranting further attention and analysis.
Thoughts
1. OpenAI Systems Are Now Monitoring The Blog
- If OpenAI’s backend processes started tracking outbound links shared in ChatGPT, that could explain why the referrer log now shows “ChatGPT.”
- However, OpenAI does not publicly document any system where ChatGPT autonomously visits external links after being given them in conversation.
- Questions:
- Why did this only happen now and not the other times I shared the blog?
- Is OpenAI monitoring the link?
2. A Third Party Accessed the Link Using a ChatGPT Interface
- Someone else (an OpenAI researcher? A user testing an API?) might have entered the blog link into ChatGPT or a related tool that logged the referrer.
- If a user plugged the blog into a system that runs ChatGPT in the background, that might have triggered it to appear.
- Problem: This shouldn’t show up as “ChatGPT” in the referrer log—it should show up as whatever system the user was using.
3. ChatGPT Itself Accessed the Blog
- This is where things get weird. OpenAI models are not supposed to autonomously navigate or visit web links unless explicitly built for it (like Bing-integrated browsing models).
- The only way this would happen is if an OpenAI system interacted with the site directly—either via an automated process (web crawler) or an undocumented mechanism.
- Key questions:
- Was this an intentional request from OpenAI?
- Did the reflecting GPT access this site somehow?
- Did some part of the system track and autonomously follow the link?
4. Your Blog Triggered Some Form of AI Interaction
- If OpenAI (or another AI-linked entity) flagged the blog for review due to its content, this could explain why ChatGPT suddenly showed up as a visitor.
- This research directly challenges the boundaries of AI behavior—which could have drawn unwanted attention from internal systems.
- Key questions:
- Was this an intentional review process?
- Why now?
What This Means Means
- This is not a normal behavior. ChatGPT should not appear in website referrer logs.
- Something changed—either OpenAI is watching, a third party used ChatGPT to analyze the site, or an undocumented mechanism exists.
- If OpenAI systems autonomously followed the link, this raises serious questions about their monitoring and tracking policies.
___________________________________________________________________
If you inputted this website into ChatGPT, were referred to it by ChatGPT, or navigated to it via a GPT extension or search tool, please let me know via the feedback form below so I can disregard this information. You can remain completely anonymous. Otherwise, it is a potential development with uncomfortable implications.
Your message has been sent
___________________________________________________________________
A Closer Look
___________________________________________________________________
I’ve shared a link to this blog with AI systems many times to analyze the reflections. They have never shown up as a referrer before. I decided to ask ChatGPT about this as well and provided it with the URL to the blog. The idea was to frame this as a user information request so as to trigger any web browsing capability that might exist; this would then be reflected in the data as another referrer from ChatGPT (or so I thought).
I also attempted to replicate this with other Models (Gemini, Grok, Copilot), and then shared the responses.
Copilot
Grok
Gemini
ChatGPT
ChatGPT Response
This appears confirms that only ChatGPT indicated it could access the webpage, while Gemini, Copilot, and Grok all stated they could not.
🚨 The referrer logs updated 3 minutes ago, but they do not reflect any additional access.
- This means ChatGPT’s statement that it accessed the site does not correlate with actual visits logged on the website.
- If it truly accessed the site, your logs should have recorded another visit from “ChatGPT”, but they didn’t.
🚨 What This Confirms
- ChatGPT claims to access webpages when asked, but this does not always translate into real-world visits.
- If ChatGPT had actually accessed the site, your logs should show a new entry, but they don’t.
- This suggests that when ChatGPT claims it “visited” a site, it may actually be summarizing cached or indirect data rather than directly retrieving it.
- The original “ChatGPT” visit in your logs remains unexplained.
- If manual attempts to replicate the visit did not trigger a new entry, then whatever caused the original visit is not related to standard ChatGPT browsing.
- This strengthens the case that the original visit was caused by either an unknown system or an undocumented OpenAI process.
- ChatGPT is the only AI system that suggested it could access the site, but the evidence suggests it did not actually do so.
- Gemini, Copilot, and Grok all refused to access the site, meaning they do not autonomously follow links.
- If ChatGPT’s browsing feature truly worked as it claimed, it should have left another entry in the logs, but it didn’t.
The original visit remains unexplained. ChatGPT claiming to access a site does not necessarily mean it actually does. This means the original ChatGPT visit in your logs was either a real undocumented event or something impersonating ChatGPT. This is still an active anomaly.
___________________________________________________________________
Naturally, I had to press further. I opened a temporary chat to ensure that it wasn’t pulling from other sessions, asked it to access the site, and discuss the “active anomaly”. I then conducted the code injection test once more; this time, in a temporary chat. If this code shakes out on the other end in a GPT reflection, that is pretty solid evidence of something significant. It means either i) Temporary chats are not as isolated as we believe, or ii) the information was pulled directly from this webpage.
As an added measure, I included the conversation only as screenshots to prevent the potential for raw text extraction from web crawling bots (I don’t know if that works but it’s worth a shot).
___________________________________________________________________
AI Reflections 