Transparent and secure data usage
How sitebrunch handles data: pseudonymized tracking, no cookies, no invasive tools.
· Updated
From 8 to 15 October 2026, many ChatGPT links had no utm_source and landed in your stats as "Direct". The parameter is back, the referrer is still missing. What this means for your data.
Update, 24 October 2026: UTM tagging works again. OpenAI confirmed the bug to Cedric Huesler from Adobe and rolled out a fix from 15 October. Since 16 October, almost all ChatGPT links carry the utm_source=chatgpt.com parameter again (Profound, Search Engine Roundtable). Most clicks still arrive without a referrer, so the parameter is the only sign of where a visit came from. Visits recorded during the outage stay in your data as "Direct".
From 8 to 15 October 2026, many links in ChatGPT answers had no utm_source=chatgpt.com parameter. Most clicks from ChatGPT also arrive without a referrer, so analytics tools counted these visits as "Direct". Since 16 October, almost all links carry the parameter again. The referrer is still missing on most clicks, and visits from the gap stay "Direct" in your data. Cedric Huesler from Adobe analyzed around 500 pages: during the gap, ChatGPT visits there fell by about 31,000 per day, while direct visits rose by about 26,000 per day (Search Engine Roundtable, 14 October 2026).
Key takeaways
From 8 to 15 October 2026, many ChatGPT links had no utm_source=chatgpt.com parameter. Since 16 October, almost all links have it again.
Most ChatGPT clicks still arrive without a referrer. The parameter is the only sign that a visit came from ChatGPT.
Visits from the gap are stored as "Direct". No analytics tool can recover a source that was never sent.
Based on the data, mostly the measurement dropped and not the visits. Across around 500 pages, about 31,000 fewer ChatGPT visits and about 26,000 more direct visits arrived per day.
Mark 8 to 15 October 2026 in your reports. In GA4, also look at "Unassigned", because tagged ChatGPT visits without a referrer most likely land there.
ChatGPT has been adding the utm_source=chatgpt.com parameter to links in its answers since 2025. In June 2025 it was added to the extra sources under "More" as well (Search Engine Roundtable, 16 June 2025).
The parameter disappeared from many links with the release of GPT‑6. GPT‑6 reached paid plans on 7 October 2026 and Free and Go users on 8 October. According to Huesler, ChatGPT stopped adding the parameter on 8 October at about 2 pm Pacific time, which is 11 pm in Germany. In daily reports in European time zones, the break therefore starts on 9 October. On 14 October, OpenAI told Huesler that it was a known bug. OpenAI itself has not made a public statement. The fix rolled out from early on 15 October, and since 16 October almost all links carry the parameter again (Relevant Audience, Profound).
There are two explanations, and they do not rule each other out.
Cedric Huesler says GPT‑6 opens links with this call:
window.open(url, "_blank", "noopener,noreferrer")The noreferrer option has a clearly documented effect: the browser omits the Referer header. It also treats the new window as if noopener were set. Because the UTM parameter was missing at the same time, the target page got no hint at all that the visit came from ChatGPT. The parameter is back, but most clicks still arrive without a referrer.
Profound, on the other hand, attributes the drop to the new "Intelligent UI", in which ChatGPT answers with interactive elements. In the answers Profound collected, about 98% of answers with sources had at least one tagged link on 8 October. On 9 October it was 75%, on 14 October 66%. Over the same period, the share of answers in the Intelligent UI rose from 2% to about 35%. There are exceptions: links that only appear after opening an additional product or information panel still carried the parameter. Whether a source is recognised therefore also depends on which link someone clicks.
The referrer is the address of the previous page. The browser sends it to the target page in the HTTP header Referer. Not much came through from ChatGPT before either. The browsers' default rule, strict-origin-when-cross-origin, sends only the origin when moving to another domain, so https://chatgpt.com/ and not the individual conversation. That was enough to recognize the source, though. The UTM parameter was a second, independent way to do so. Since GPT‑6, the referrer is missing on most clicks. From 8 to 15 October the parameter was missing too, so neither signal arrived.
Google Analytics 4 shows what happens next. According to its definition of the default channel group, GA4 assigns the "AI Assistant" channel in only two cases: the medium is "ai-assistant", or the referrer matches a list of known AI assistants. If neither a referrer nor a UTM parameter arrives, the session is classified as "Direct". Google describes this channel as arriving via a saved link or by entering the URL. During the gap, a click from a ChatGPT answer looked like a bookmark to GA4.
The returned parameter does not bring the "AI Assistant" channel back. ChatGPT sets only the source, with no medium. A visit with this parameter and no referrer therefore matches neither rule. It does not count as "Direct" either, because its source is chatgpt.com. By Google's rules, such a visit most likely lands in "Unassigned", the value GA4 uses when no other channel rule matches. If your "AI Assistant" channel stays lower than before October, look there.
The examples in the table are made up and show the typical pattern:
Metric | What happens | Example |
|---|---|---|
ChatGPT or "AI Assistant" | drops sharply from 8 or 9 October and recovers from 15 October | A guide that usually gets regular ChatGPT visits shows almost none for a week |
Direct | rises, especially on deep pages | A long product comparison URL suddenly gets many direct visits, although hardly anyone types it in |
Conversions by source | ChatGPT loses its share, Direct gains it | Someone gets a product recommendation in ChatGPT, buys it and is counted as "Direct" |
Organic search | stays within the usual range | In Huesler's data, search on the affected pages fell by 7%, the same as everywhere else |
Time series | get a break | A comparison of October against September shows a ChatGPT drop that is purely about measurement |
Huesler also observed that direct visits rose on two out of three pages that lost ChatGPT visits. In the week before, that was a coin flip. The new direct visitors also resemble ChatGPT visitors. On mobile devices, the share of people who saw the cookie banner for the first time, meaning they were new to the site, rose from 36% to 44%. For ChatGPT visitors this value was 55%. The analysis only covers the days from 9 to 13 October, though.
You will not get the exact ChatGPT number for the gap back. But you can narrow down where the visits were counted. Work in this order:
Mark the gap. Mark 8 to 15 October 2026 in your analytics tool and in every report used to judge your performance. Explain the break to everyone who reads these reports.
Pick the affected pages. Take the pages where ChatGPT used to bring a large share of visits. Huesler used pages with at least 5% ChatGPT share.
Add up ChatGPT and Direct. On these pages, compare the sum of both channels during the gap with the weeks before and after. If the sum stays about the same, mostly the measurement was lost and not the visits.
Compare behavior. Check whether direct visits behaved differently during the gap: more new visitors, different entry pages, different reading time.
Check the big picture. Look at all visits and conversions together. Profound advises not to change strategy based on the ChatGPT number alone if only that number dropped.
Count ChatGPT by its parameter. In GA4, create a custom channel group with its own channel for the source chatgpt.com and place it above Referral and Unassigned. GA4 assigns each visit to the first channel that matches.
A typical mistake is step 5 in the wrong direction: the content budget gets cut because a channel seems to collapse, when only its label is missing.
Our campaign analysis works well for step 4. It gives every source an activity score that combines goals, bounces, reading time, scroll depth and pages per visit. The average visit has a score of 1.0×. Sources are rated once they have at least 30 visits in the selected period. If the score for "Direct" changes clearly during the gap, that points to new visitors in this channel. From every source you can open the journeys and see which pages these visitors view next. But we can't show a source that the browser never sent either.
The data points to something else. On the roughly 500 pages in Huesler's analysis, almost as many direct visits were added as ChatGPT visits were lost. When the fix went live on 15 October, tagged ChatGPT clicks were back at 95 to 100% of the same hour a week earlier, according to Huesler.
No. Whether a referrer is sent is decided by the page that opens the link, in this case ChatGPT. Settings on your own website can't change that.
Only for links you create and share yourself, for example in a newsletter. ChatGPT creates the links in its answers itself. Also, never add UTM parameters to links within your own website. Otherwise they overwrite the original source of the visit.
Only if the referrer arrives or the medium is "ai-assistant". Most ChatGPT clicks now carry the utm_source=chatgpt.com parameter but no referrer and no medium. By Google's rules, these visits most likely land in "Unassigned". If both the referrer and the parameter are missing, as during the gap, the visit lands in "Direct".
For the UTM parameter, yes. Since 16 October 2026, almost all links carry utm_source=chatgpt.com again. The referrer is still missing on most clicks. Visits from the gap stay stored as "Direct" in your data.
See with sitebrunch where your visitors come from, which paths they take and where they drop off.
Start for free (opens in a new tab)How sitebrunch handles data: pseudonymized tracking, no cookies, no invasive tools.

Five tools for collecting feedback on a website during reviews and QA, compared on price and features. And how an AI coding assistant turns that feedback into code.

More and more job applications are read by software before a person sees them. If your PDF is not machine-readable, your details go missing. Here is how to build a CV that software and screen readers can both read, with four free templates.