If your ChatGPT referral numbers fell off a cliff around 9 October 2026 while Direct traffic went up, your site is not alone. After GPT-6 rolled out to ChatGPT users around 7 and 8 October, many links in ChatGPT's new answer format opened without a referrer and without the utm_source=chatgpt.com tag. With neither signal present, GA4 and most other analytics tools had nothing to go on and counted those visits as Direct.
The data so far suggests this was a labelling problem, not a traffic problem: visits kept coming, they just lost their name. On 15 October, the day this article was published, reports said the tags were coming back. The episode is still worth understanding, because it shows how fragile AI referral tracking is: it depends on signals that the AI platform controls, not you.
What happened to ChatGPT referral traffic
Three independent observations line up.
LLM Pulse published an analysis on 13 October of the sites it monitors. ChatGPT referral traffic in GA4 dropped about 46% between 9 and 11 October, compared with the same days in the two previous weeks. Total traffic on the same sites moved only about −2%, and Perplexity and Claude referrals showed no break on 9 October. The drop appeared across countries and site sizes, which LLM Pulse reads as ruling out holidays and seasonality. LLM Pulse does not say how many sites are in its sample.
Cedric Huesler of Adobe looked at the mechanism. According to Search Engine Roundtable's write-up of his LinkedIn post, GPT-6 opened links with window.open(url, "_blank", "noopener,noreferrer") and dropped the utm_source tag. The noreferrer flag tells the browser not to send a Referer header, so the destination site cannot see where the visitor came from.
His numbers, covering about 500 pages where ChatGPT drove at least 5% of visits and comparing 9–13 October with the week before:
| Signal | Change |
|---|---|
| ChatGPT visits | about −31,000 per day |
| Direct visits on the same pages | about +26,000 per day |
| Search traffic | −7%, in line with everywhere else |
| Pages that lost ChatGPT traffic and gained Direct | about two in three |
Most of the "lost" ChatGPT visits reappear as Direct on the very same pages. That is what a broken label looks like.
On 15 October, Huesler updated his post to say it was a bug and that ChatGPT was tagging its links with utm_source=chatgpt.com again, and Search Engine Roundtable reported the same. At the time of writing, OpenAI has not made a public statement about the change or the fix.
One caveat is worth keeping. LLM Pulse notes that analytics alone cannot rule out that some answers now resolve inside the chat and earn fewer clicks. The evidence favours the labelling explanation, but it is not proof that every missing visit was relabelled.
Why ChatGPT traffic shows as direct: how AI referral tracking works
An analytics tool can only attribute a visit to ChatGPT if the visit carries some trace of ChatGPT. In practice there are two:
- The HTTP referrer. When someone clicks a link, the browser normally tells the destination which page the click came from, for example
https://chatgpt.com/. A link opened withnoreferrer, or a referrer policy that strips it, removes this signal. - The UTM tag. ChatGPT appends
utm_source=chatgpt.comto many links. That tag survives even when the referrer is missing, which is why it has been the safety net for ChatGPT referral traffic. (If UTM tags are new to you, see our guide to UTM parameters for campaign tracking.)
When both are gone, the visit looks exactly like someone typing your URL or opening a bookmark. Every tool files that under Direct or "none". There is no setting in GA4, or anywhere else, that can recover a signal the browser never sent.
That is the uncomfortable lesson: AI referral attribution depends on choices the AI platform makes. OpenAI decides whether its links send a referrer and whether they carry a tag. The same applies to Gemini, Perplexity, Claude and Copilot. LLM Pulse notes, for example, that Gemini referrals fell earlier, between 28 September and 2 October, after a change in how Gemini tags its links, while visits from gemini.google.com held steady.
Why it matters: ChatGPT is most of your AI referral traffic
The stakes are higher for ChatGPT than for any other assistant. BrightEdge said on 24 September that ChatGPT accounted for 95.1% of AI referral traffic in August 2026, with Gemini at 2.4%, Perplexity at 1.1% and Claude at 0.8%. BrightEdge has not published its methodology, so treat the exact figures with care, but the direction matches what most site owners see: when ChatGPT attribution breaks, your AI channel breaks with it.
If you report AI traffic to a client or a board, a 46% overnight drop is the kind of number that triggers a strategy change. In this case the right response was to do nothing except annotate the chart.
How to detect dark AI traffic
"Dark AI traffic" is the share of visits that came from an AI assistant but arrived without any label. You cannot measure it exactly, but you can find its footprints.
1. Watch Direct on deep pages
People rarely type /blog/how-to-configure-x into an address bar. Direct visits normally land on your homepage or on short, memorable URLs. A rise in Direct entries on deep pages such as blog posts, documentation and product detail pages is a strong hint that something upstream lost its label.
2. Compare your top ChatGPT landing pages
Take the pages that received the most ChatGPT visits in the weeks before 9 October. If ChatGPT traffic on those pages fell while Direct traffic on the same pages rose by a similar amount, you are looking at relabelled visits, exactly as in Huesler's data.
3. Check total traffic and conversions first
A channel that halves while total visits and conversions stay flat has not disappeared. Before you change anything based on the ChatGPT line, check the totals.
4. Annotate the gap
Visits that arrived as Direct between 9 October and the fix will not be reattributed later. Add an annotation for the period in your dashboards, and avoid before-and-after comparisons that span it.
5. Track visibility, not only clicks
Whether ChatGPT cites you at all does not depend on referral tags. Tracking how often AI assistants mention your brand gives you a second line of evidence when click data goes dark. Our LLM visibility tool does exactly that.
Referrer + UTM + landing page: why you need all three
Each signal fails in its own way:
| Signal | What it tells you | How it fails |
|---|---|---|
| Referrer | The exact AI domain, such as chatgpt.com or perplexity.ai | Stripped by noreferrer, in-app browsers and strict referrer policies |
| UTM tag | The source, even without a referrer | Only present if the platform adds it, and the platform can drop it |
| Landing-page heuristic | That a no-referrer visit to a deep page is unlikely to be typed in | Cannot name the assistant and also catches bookmarks, email clients and chat apps |
No single signal is reliable on its own. Used together, with the heuristic clearly marked as an estimate, they give you the most honest picture. We covered the wider landscape in AI and LLM referral tracking: the next frontier in analytics.
What this means for your analytics
ClearAnalytics recognises AI visits in three layers. A visit whose referrer is a known AI domain (chatgpt.com, claude.ai, perplexity.ai, gemini.google.com and others) is counted with high confidence, as are visits with experimental AI referral headers. UTM tags are stored and shown in your campaign reports. An entry visit with no referrer that lands on a deep page is flagged separately as a low-confidence, possible AI visit, without naming an assistant, so you can see when dark traffic is rising without mixing estimates into your confirmed numbers. For a step-by-step setup, read how to track AI and ChatGPT traffic to your website.