OpenAI made some updates to its ChatGPT systems, which are impacting the amount of traffic ChatGPT is sending out from its interface. Josh Blyskal from Profounds posted on LinkedIn that "ChatGPT referral traffic is down -52% since July 21st."
Josh said this is pulled from the Profound dataset of over one billion ChatGPT citations and over one million referral visits from ChatGPT. "ChatGPT is shifting towards sites that provide 'answers', and the long tail of citations is shrinking as a result," he added.
I will say, a number of people who manage a lot of large sites are saying they do not see the decline. Check your data and see if you are seeing a decline in ChatGPT referrals since July 21st.
Here is a chart showing the recent decline; now keep in mind, ChatGPT usage continues to grow, but clicks from ChatGPT to websites has been declining:
I mean, this isn't a huge deal since, as Glenn Gabe wrote, 1% of your traffic is from AI Search - so that means that maybe you get one less click this week from ChatGPT?
The larger point, as Glenn put it on X is that "OpenAI can turn the dials and rankings and downstream traffic can change radically." And we have no insight into those changes, less than we have with Google and Bing.
So it is hard to know if the traffic drop from ChatGPT is something an SEO (or GEO) did or did not do that caused it.
Josh goes on to explain that citations are away up for Reddit, Wikipedia and some other sites. This means that your site may be shown less often in ChatGPT?
The referral decline started right as citation patterns shifted dramatically. Reddit citations increased +87% starting July 23rd, reaching more than 10% of all ChatGPT citations. Wikipedia simultaneously hit historic highs, up +62% since its July low to nearly 13% citation share yesterday.The top 3 domains (Wikipedia, Reddit, TechRadar) combined have grown +53%, now controlling 22% of all citations. That's 1 in 5 ChatGPT citations going to just three sites.
At the same time, branded websites are seeing fewer citation opportunities: millions of potential referrals are being absorbed by these dominant platforms. This isn't a GPT-5 change because the consolidation started weeks before the model shipped.
He ended saying:
The concerning part for SEOs in the short-term: when OpenAI adjusts these weights, referral traffic swings violently. A manual adjustment created a -52% traffic collapse in under a month. We're all downstream from OpenAI's experiments.
To clarify the 52% drop, it excluded those top citation sources. Josh explained in the comments:
The -52% decline is measured from a sample of sites we maintain across various verticals. This dataset doesn't include Reddit, Wikipedia, or the other top domains we're tracking for citations.So yes, you're understanding it correctly: while Reddit and Wikipedia are surging in citations, the average brand site in our monitoring network is down -52% in actual referral traffic.
The consolidation might be even more severe than it looks because those citation gains are translating to traffic gains for only these heavy-hitting domains.
Finally, I generally like to try to validate these third party tools reports with other third party tools, but I am limited here with the AI tools so far. I mean, I spoke to some people and they are not seeing such a stark decline in referrals from ChatGPT. For example, Tom Critchlow wrote, "Hmm, not seeing the same thing across the Raptive network. ChatGPT traffic continues to rise (though with still small absolute numbers). Don't see any step changes around July 23rd although we see a ~20% increase since Aug 7th (gpt-5 rollout date)...."
How does Profound know referral data? Josh answered that in the comments saying:
We track referral traffic from a network of sites that use our analytics tools, essentially a subset our customers who opt into aggregated data sharing.Can't name specific sites for privacy reasons, but it's a statistically significant sample across B2B SaaS, ecommerce, publishing, and other verticals.
The 1M+ referral visits we're tracking come from actual server-side data, not estimates or third-party panels. Real traffic from real sites experiencing these drops firsthand.
Should mention this is a dataset of thousands of websites and billions of logs.
Forum discussion at LinkedIn.
