ChatGPT referral traffic fell off a cliff. It was a tagging bug

If your ChatGPT referral traffic dropped by roughly half around October 8 while Direct traffic crept up, your content probably did not fall out of favor overnight. After GPT-6 rolled out, many ChatGPT links stopped carrying the utm_source=chatgpt.com tag and passed no referrer, so analytics tools filed those visits as Direct. A fix began rolling out on October 15, but the week showed how thin the thread is that AI attribution hangs on.

What happened to ChatGPT referral traffic

The timeline, pieced together from analyst reports, runs like this. GPT-6 reached paid ChatGPT plans on October 7 and Free and Go users on October 8, according to Search Engine Journal. Within days, sites saw their ChatGPT line sag. Cedric Huesler, a senior director of product management at Adobe, traced it to links in the new answers opening in a new tab with a browser setting that withholds the referrer, and without the usual utm_source tag.

Search Engine Roundtable reported the change on October 14 and said it was unclear whether it was deliberate. That same day, Huesler said OpenAI confirmed to him that it was a bug. He reported that a fix started rolling out early on October 15, and that by 9 a.m. Pacific, tagged ChatGPT clicks were back to roughly 95% to 100% of the same hour a week earlier across the domains Adobe tracks.

Two caveats. First, that confirmation is secondhand: OpenAI has not said anything publicly, and Relevant Audience found no mention of link tracking in ChatGPT's release notes for October 1 to 15. Second, the fix may be uneven. SEO consultant Glenn Gabe said he was not yet seeing tags return on his paid or free accounts when the fix was first reported.

How big the measurement gap was

Several data sets point the same way, though each comes from a company that sells analytics or AI search tools, and most samples are small or undisclosed:

• LLM Pulse found ChatGPT sessions on its monitored sites were about 46% lower on October 9 to 11 than on the same days of the prior two weeks, while total traffic stayed roughly flat. Perplexity and Claude referrals showed no break, per its analysis.
• Graphite CEO Ethan Smith saw ChatGPT-attributed GA4 sessions fall 22% to 57% across sites with meaningful AI assistant traffic, about 45% overall.
• Adobe's Huesler looked at about 500 pages where ChatGPT drove at least 5% of visits. Comparing October 9 to 13 with the prior week, ChatGPT visits fell by about 31,000 a day while Direct rose by about 26,000 a day. Direct grew on two of every three pages that lost ChatGPT traffic. He called it early, sampled data.
• Profound found the share of ChatGPT answers with at least one tagged source link fell from about 98% on October 8 to 66% by October 14, as answers using ChatGPT's new interactive format rose from 2% to about 35%.

Notice the leftover in Adobe's numbers: about 5,000 visits a day that did not reappear as Direct. Nobody has explained it yet. One possibility LLM Pulse raises is that richer in-chat answers resolve more questions without a click, which analytics alone cannot measure.

A traffic source you can only see through one URL parameter is a source you don't really measure.

Fix your reports this week

Even with tags returning, sessions recorded as Direct between October 8 and 15 will stay there. Nobody has mentioned a backfill. Here is how to keep the week from distorting your reporting:

1. Annotate October 8 to 15 in GA4, Looker Studio and any client dashboards. Label it as a tracking incident, and don't rebuild attribution models or channel rules around it.
2. Reconcile at the page level. For the pages ChatGPT normally sends visitors to, add ChatGPT and Direct sessions together and compare against the prior weeks. Include the homepage: Peec AI found that 54% of the direct URLs ChatGPT showed during the incident pointed to home pages.
3. Check your AI channel definition. If you use a custom channel group for AI assistants, make sure it matches on both the source tag and the referrer domain, and that it sits above the default Referral rule so AI visits aren't swallowed.
4. Look at Gemini too. LLM Pulse saw visits tagged "gemini" nearly disappear between September 28 and October 2 while visits from gemini.google.com held steady, which points to a change in how Gemini labels its links. Make sure both labels roll into the same channel.
5. Report on outcomes. Judge AI search performance on total visits, leads and revenue on your key pages, not on the ChatGPT referral line alone.
6. Leave paid links alone. Smith said UTM parameters you set yourself on ChatGPT Ads links should be unaffected.

OnlyaClick tip: In a GA4 exploration, put landing page in rows and session source / medium in columns, then filter to the 20 pages that earned the most chatgpt.com sessions in September. Compare October 1 to 7, October 8 to 15, and the days since. If Direct swelled on exactly those pages during the middle week, you've found your missing ChatGPT visits, and you have a chart to show the client.

ChatGPT referral traffic is the floor, not the ceiling

The bigger lesson is that clicks from AI assistants were always an undercount. Huesler noted that most ChatGPT clicks still arrive without a referrer even after the fix, so the tag is doing almost all of the work. On mobile, Graphite's Smith said 95% of ChatGPT visits to its sites relied on the tag alone. Any change to how ChatGPT opens links can blank your numbers again.

So build a measurement stack that does not depend on one parameter:

• Visibility tracking. Track whether you are mentioned or cited in AI answers for the prompts that matter to your business. Citations don't depend on tags.
• Server logs. OpenAI documents a ChatGPT-User agent for pages fetched when a user asks a question. Requests from it are a sign ChatGPT is reading your pages, independent of clicks.
• Self-reported attribution. Add "an AI assistant like ChatGPT" as an option on your "How did you hear about us?" field, and ask sales teams to log it.
• Crawl access. Confirm your robots.txt, CDN and firewall aren't blocking the AI search crawlers you want to be found by. A technical SEO audit should now include this check.

There is a content angle as well. Peec AI found ChatGPT displayed a direct URL in 42% of responses during October 9 to 13, up from about 8% to 10% before, though that measures links shown, not clicks. As answers get more interactive, the pages that still earn the click will be the ones offering something the answer can't: a tool, a quote, a booking, original data or a real point of view.

The bottom line

The October dip in ChatGPT referral traffic was mostly a labeling problem, and tagging appears to be coming back. Annotate the week, reconcile it page by page, and stop treating one UTM tag as the source of truth. Pair referral data with citation tracking, logs and self-reported attribution, and keep improving the pages AI assistants send people to. If you want a second set of eyes on how you show up in AI answers, our AI search optimization team can help.