A sharp drop in ChatGPT tracking data across Google Analytics 4 (GA4), starting Oct. 9, appears to reflect mainly a measurement failure rather than a real loss of visitors, although there is not yet enough data to rule out some real loss. ChatGPT stopped appending its utm_source=chatgpt.com tag to inline answer links and began opening pages without passing a referrer (the browser signal that tells your analytics where a visitor came from), shifting those sessions into the Direct channel instead.
How large was the apparent drop?
The numbers were large enough to alarm site owners. Graphite CEO Ethan Smith wrote on LinkedIn that ChatGPT-attributed sessions in GA4 fell 22% to 57% for sites with meaningful AI Assistant channel traffic starting Oct. 9, which he estimates represents approximately a 45% fall from the prior baseline. LLM Pulse, tracking a separate set of sites, reported sessions from ChatGPT were approximately 46% lower during Oct. 9–11 compared with the same days in the prior two weeks. Total traffic on those sites remained relatively constant over the same period, an early signal the underlying visits had not disappeared.
How did ChatGPT tracking break?
Cedric Huesler, senior director of product management at Adobe, reported on LinkedIn that after the GPT-6 rollout, ChatGPT began opening links in a new tab with a browser setting that withholds the referrer, and stopped appending the utm_source tag that its citations had carried since ChatGPT’s search launched in 2024.
Not every link type was equally affected, according to the analysts:
- Inline links in answer text: lost both the UTM tag and the referrer.
- Links in the Sources panel and shopping links: retained the tag, per Smith.
- Links accessed by opening a product or information panel: still carried tags, per Profound.
Smith noted that 95% of mobile ChatGPT visits to Graphite’s sites relied on the UTM tag alone, making app traffic the most exposed segment.
What drove the change? A new answer format
Profound, which tracks ChatGPT answer formats, tied the shift to the rise of a new format called Intelligent UI, ChatGPT’s presentation for charts, comparisons, and small tools. Among ChatGPT answers with citations that Profound collected, the share with at least one tagged source link fell from about 98% on Oct. 8 to 66% by Oct. 14. Over the same period, Intelligent UI went from 2% of all answers Profound collected to about 35%.
GPT-6 began rolling out to paid plans on Oct. 7 and to Free and Go users on Oct. 8. LLM Pulse observed no drop on Oct. 7.
Tomek Rudzki at Peec AI ran the same 1 million prompts across two periods, Oct. 2–7, then Oct. 9–13, and observed that ChatGPT placed a direct URL for a webpage in 42% of responses, compared with a rate of about 8–10% previously. Of those URLs, 54% pointed to homepages. He reported no tags or referrers associated with those links. These figures represent links displayed by ChatGPT, not actual user click-throughs.
How did analysts detect the visits hiding in Direct?
Huesler compared Oct. 9–13 with the prior week across about 500 pages where ChatGPT drove at least 5% of visits. ChatGPT visits to those pages fell by about 31,000 a day while Direct visits rose by about 26,000 a day. Search traffic fell 7%, in line with broader search trends.
Direct rose on two of every three pages that lost ChatGPT traffic, compared with about half the pages the week before. Huesler also found that the new Direct visitors behaved like ChatGPT visitors. On mobile, the share of Direct visitors who saw the consent banner (his proxy for a first-time visit) rose from 36% to 44%, moving toward the 55% he typically sees among ChatGPT visitors. He called the findings early, based on four days of sampled data.
LLM Pulse noted that Direct is a much larger channel than ChatGPT, making a relatively small influx difficult to detect in aggregate data, and making it impossible to distinguish traffic that was genuinely lost from traffic that was merely mislabeled.
Is a fix in place?
Huesler says OpenAI confirmed to him on Oct. 14 that the behavior was a bug. In a follow-up post, he wrote that a fix began rolling out early on Oct. 15, and that by 9 a.m. PT, tagged ChatGPT clicks were back to roughly 95% to 100% of the same hour a week earlier, across thousands of domains he tracks.
Smith had earlier called the change a bug on OpenAI’s side and wrote that he expected a fix “very soon.” SEO consultant Glenn Gabe replied to Huesler’s post that he was not yet seeing the change on paid or free accounts. Profound stated it had not seen official word on a fix, and OpenAI’s ChatGPT release notes for Oct. 1 through Oct. 15 do not mention link tracking.
Huesler recommends counting the UTM tag rather than the referrer for ChatGPT attribution going forward, since most ChatGPT clicks still arrive without a referrer even after the fix. Both Smith and Huesler suggest annotating Oct. 8 in any reporting dataset, with Huesler putting the full affected window at Oct. 8–15.
What this means for AI-search visibility
Here is the plain-English version: when ChatGPT sends someone to your site, it normally attaches a label to the link, utm_source=chatgpt.com, that tells your analytics this visitor came from ChatGPT. For about a week, that label went missing from many inline links, so those visits showed up as Direct, as if the person had typed your URL themselves. No label does not mean no visitor.
A few things the combined data points imply:
- The roughly 45% fall in ChatGPT-attributed sessions appears to have been largely a measurement gap, not a collapse in referral intent. The Adobe signal is the clearest evidence: Direct rose 26,000 visits per day on the same pages that lost 31,000 ChatGPT visits per day.
- Mobile was disproportionately exposed. With 95% of mobile ChatGPT visits depending on the UTM tag alone, app traffic had little backup attribution during the bug window. Sites with a heavy mobile share from ChatGPT are likely to have a larger attribution gap to reconcile.
- The bug fix does not recover attribution that already went untagged. Profound notes that restoring tags could help with future reports, though it wouldn’t bring back attribution that’s already been lost. Any period-over-period comparison that crosses Oct. 8–15 without an annotation will misread ChatGPT’s actual referral trend.
In Hingewise’s view, the more durable question raised by this incident is not the bug itself but the format shift beneath it. Intelligent UI, the chart and comparison format that Profound observed rising from 2% to roughly 35% of ChatGPT answers, surfaces structured information without necessarily surfacing the source brand prominently. Even with UTM tags fully restored, content appearing inside these panels may drive less branded recognition than a direct citation link. That is worth monitoring separately from the tracking fix.
Gemini moved in the opposite direction: it added UTM parameters to its links two weeks ago, per reporting by Roger Montti at Search Engine Journal. The contrast matters for anyone building a multi-platform AI referral picture: AI platforms are currently making different choices about how (and whether) to pass attribution signals to publishers.
Quick checklist: what to review in your analytics now
- Annotate Oct. 8 (or the range Oct. 8–15) in GA4 before anyone interprets the dip as an audience loss.
- Compare your ChatGPT channel and Direct channel side by side for each page ChatGPT regularly referred visitors to.
- Check your homepage specifically: Rudzki’s data shows more than half of the direct URLs ChatGPT displayed during this period pointed to homepages.
- Update ChatGPT attribution to rely on the UTM tag (utm_source=chatgpt.com) rather than referrer strings, per Huesler’s recommendation.
- If you run ChatGPT Ads with your own UTM parameters on the click, Smith notes those should be unaffected by this bug.
- Monitor whether tagged clicks have fully recovered on your specific domains: the fix rollout was gradual and is not confirmed in OpenAI’s public release notes.
The open question going forward is whether Intelligent UI’s 35% share of answers holds or grows. If it does, the relevant metric for AI-search visibility may shift from did ChatGPT send traffic toward did ChatGPT answer the question in a way that still surfaces the brand.
Source: Search Engine Journal, Oct. 15 (Matt G. Southern). LinkedIn posts: Ethan Smith (Graphite), Cedric Huesler (Adobe). Additional data: LLM Pulse, Profound, Peec AI (Tomek Rudzki).
This article was drafted with AI assistance from the sources cited above and checked against them before publishing. The "What this means for AI-search visibility" section is Hingewise's editorial analysis, built on the frameworks and positions we use in our GEO work. How we work: Editorial policy.