ChatGPT referral drop in GA4 traced to missing tracking tag

ChatGPT referral traffic in Google Analytics 4 fell by roughly 45% from 9 October, and the cause appears to be a tracking fault rather than a loss of visitors. Inline links in ChatGPT answers stopped carrying the utm_source=chatgpt.com tag after the GPT-6 rollout, so those visits landed in GA4's Direct channel. OpenAI has reportedly confirmed it as a bug and started rolling out a fix on 15 October.

For any Australian business compiling October reports, the practical point is simple. A ChatGPT dip paired with a Direct bump between 8 and 16 October is most likely a labelling gap. Annotate it before anyone treats it as lost demand.

Inline links lost utm_source tag

GPT-6 reached Plus, Pro, Business and Enterprise users on 7 October and Free and Go users from 8 October. The tag started disappearing on the afternoon of 8 October US time, early on 9 October in eastern Australia, and the drop was visible from 9 October. Inline links had lost the utm_source tag that ChatGPT citations have carried since ChatGPT search launched in 2024.

OpenAI's 7 October release notes describe Intelligent UI as a mix of prose, graphics and interactive pieces, from comparisons laid out side by side to small tools such as calculators and bill splitters, with ChatGPT picking the format for each answer and still replying in plain text when that suits. GPT-6 can also keep using its tools after it starts replying and add results as they arrive. The notes say nothing about how source links appear inside those formats.

Cedric Huesler, a senior director of product management at Adobe, says ChatGPT also started opening those links in a new tab with a browser setting that strips the referrer. With neither a tag nor a referrer, GA4 files the visit as Direct.

Graphite CEO Ethan Smith noted that links in the Sources panel and shopping links kept their tags; only inline links in the answer body dropped them. On Graphite's sites the tag was the only signal for 95% of mobile ChatGPT visits, so traffic from the ChatGPT app was the most exposed.

Profound, which collects ChatGPT answers for analysis, ties the change to a new answer format. Of the answers it gathered that cited sources, about 98% included a tagged link on 8 October. By 14 October that figure was about two thirds, while the format OpenAI calls Intelligent UI, used for charts, comparisons and small tools, had risen to about a third of answers. The tagged share then collapsed to around 5% to 12% for about a day from the evening of 14 October UTC, early on 15 October in eastern Australia, so 15 October is likely the deepest dip in local reports. Links opened through a product or information panel inside those answers still carried tags.

Graphite-tracked sites saw ChatGPT-attributed sessions fall 22% to 57%

Smith's figures cover sites Graphite tracks with meaningful traffic in GA4's AI Assistant channel. On those sites, ChatGPT-attributed sessions fell between 22% and 57% from 9 October, about a 45% drop from baseline by his estimate.

LLM Pulse saw a similar pattern on the sites it monitors: ChatGPT sessions were about 46% lower from 9 to 11 October than on the same days in the prior two weeks, while total traffic stayed roughly flat.

Peec AI's Tomek Rudzki measured the change from the other direction, running the same one million prompts from 2 to 7 October and again from 9 to 13 October. In the second run, ChatGPT placed a direct URL in 42% of responses, up from roughly 8% to 10%, and 54% of those URLs pointed to home pages. None carried tags or referrers. Those are links displayed, not clicks, so they describe exposure rather than visits.

Did missing ChatGPT visits show up as Direct?

Huesler looked at roughly 500 pages that get 5% or more of their visits from ChatGPT, comparing 9 to 13 October against the prior week. Daily ChatGPT visits dropped by about 31,000, while daily Direct visits climbed by about 26,000. Search traffic to the same pages slipped 7%, matching the broader search trend.

Two in three pages that lost ChatGPT traffic gained Direct visits, against about half the week before. The new Direct visitors also behaved like ChatGPT visitors: on mobile, the share shown the consent banner, Huesler's proxy for a first visit, moved from 36% to 44%, towards the 55% he sees among ChatGPT visitors. He calls these early findings from four days of sampled data.

LLM Pulse could not see a rise in Direct at first. Direct is a much larger channel than ChatGPT, so a few thousand sessions can disappear inside it, and aggregated data cannot separate lost traffic from relabelled traffic. Its 16 October update says the evidence now points clearly to lost labels, with the Direct rise visible at page level, while allowing that GPT-6 may answer more inside the chat and send somewhat fewer clicks. Some real decline cannot yet be ruled out.

Tags back, release notes silent

Smith called the change a bug on OpenAI's side and expected a fix soon. Huesler says OpenAI confirmed to him on 14 October that it was a bug and that a fix began rolling out early on 15 October. By 9 a.m. Pacific time that day, the early hours of 16 October in eastern Australia, tagged ChatGPT clicks across the thousands of domains he tracks had recovered to between 95% and 100% of their level at that hour seven days before.

SEO consultant Glenn Gabe replied on 15 October that he was not yet seeing the change on paid or free accounts. By the afternoon of 16 October in eastern Australia, Profound's sampling had about 99% of cited answers tagged again, including inside Intelligent UI. OpenAI's ChatGPT release notes do not mention the bug or the fix, so the only confirmation on record is second-hand. Entries dated 15, 16 and 20 October, covering new voices, ChatGPT Slides, a GPT-6 Pro update and locked chats among other things, still say nothing about link tagging.

Two caveats apply to all the figures above. Graphite, LLM Pulse, Profound and Peec AI all sell SEO or AI search products, Huesler works for Adobe, and LLM Pulse and Profound have not published sample sizes.

Google is heading the other way. Gemini links began carrying UTM parameters in late September. Google's John Mueller said he could see the tags too, although Google has not documented the change. Meanwhile, Google is moving inline links higher in AI Mode and AI Overviews, and Search Console counts clicks from those AI results inside its overall web search figures without separating them.

Missing tags leave a false dip and unearned Direct lift in October reports

GA4's AI Assistant channel, introduced in May, is documented as matching on the referrer. Google's help page defines the channel as sessions whose medium is exactly ai-assistant, and says that when the referrer matches its list of AI assistants, GA4 sets that medium and a campaign of (ai-assistant) itself. The page gives ChatGPT, Gemini, DeepSeek, Copilot and Grok as examples of those sources. Clicks from Google's own AI Overviews and AI Mode are excluded and counted under Organic Search. Direct, by contrast, is defined as a source of (direct) with a medium of (none) or (not set). When the referrer is missing, the utm_source tag is what lets GA4 record ChatGPT as the source. Without either, the session reports as Direct. Huesler says most ChatGPT clicks still arrive without a referrer even after the fix, so the tag does most of the work. For a week it was gone, and the sessions it would have claimed sat in Direct.

If ChatGPT sends a small share of visits, the effect on a monthly report is noise. If it sends 5% or more to key pages, the October line will show a dip that was never a drop in interest, and the Direct line a lift that was never earned. Both are reporting artefacts, and both will be mistaken for something real unless flagged now.

Restoring the tags helps from here on but, as Profound notes, does not repair the attribution already lost. That is why the window needs an annotation rather than a quiet correction.

The same discipline applied during the recent Search Console data delay, when the performance report fell more than 64 hours behind after Google's spam update and then caught up. Wait for the data to settle before judging the site, the update or the agency.

Reporting errors run the other way too. Meta Ads was found counting welcome message views as onsite conversions in several Australian accounts, inflating a number instead of hiding one. The fix in both cases is knowing what each line in the report actually counts.

Annotate 9 to 16 October

• Add an annotation in GA4 explaining that ChatGPT inline links lost their tracking tag. For a GA4 property set to Australian time, annotate 9 to 16 October; 8 to 16 October covers every time zone.
• Pull ChatGPT and Direct sessions side by side for every landing page ChatGPT referred traffic to, and for the home page, where Rudzki found more than half of the new links pointing.
• Check whether tagged ChatGPT sessions have returned to their early-October level. If they have not, the fix may not have reached the accounts sending the site its traffic.
• In the Traffic acquisition report, switch the dimension to Session source / medium and look for chatgpt.com. Google's URL builder guidance notes that any UTM field a link does not set shows as (not set). The default channel group cannot be edited, but a custom channel group can be built around that source if the business wants one line for all AI referrals.
• If the business runs ChatGPT Ads with its own UTM parameters, leave them alone. Smith says those clicks should be unaffected.
• Stop treating the utm_source tag as the only measure of AI referrals. Pair it with landing-page trends and Direct, and note the window in any month-end or quarterly report before it goes to a client or a board.

For the domains Huesler tracks, the tag came back within a week. The annotation is what stops that week from being misread in six months' time.