ChatGPT Referral UTM Parameters Dropped

Yesterday, your analytics report may have shown ChatGPT referrals falling off a cliff. That doesn’t mean ChatGPT stopped sending people to your site.

Search Engine Roundtable reported that ChatGPT had dropped UTM parameters from links. Today, AI analytics company Profound published its own analysis and found that most links it checked directly in ChatGPT answers were missing utm_source=chatgpt.com.

OpenAI hasn’t explained the change. Until it does, I’d treat this as a measurement problem rather than a traffic problem.

Two explanations, neither confirmed

Barry Schwartz at Search Engine Roundtable said the change seemed to start around the GPT-6 release about a week ago. Adobe’s Cedric Huesler suggested that GPT-6 opens links with window.open using noopener and noreferrer. That can remove both the UTM tag and the referrer, leaving your analytics platform to call the visit direct.

Huesler checked about 500 pages. From October 9 through October 13, those pages lost roughly 31,000 ChatGPT visits per day compared with the prior week. Direct traffic to the same pages rose by about 26,000 visits per day, and those new direct visitors behaved like the missing ChatGPT visitors. Sébastien Billard also reported a sharp decline in human traffic attributed to ChatGPT after October 9.

Profound points to a different cause. It ties the missing tags to ChatGPT’s “Intelligent UI” rollout rather than GPT-6 itself. Profound also found inconsistent behavior inside the product: a link shown directly in an answer could lack the UTM tag while a link reached deeper inside the interface still had one.

So which explanation is right? We don’t know. I couldn’t find an OpenAI statement, help-center update, or confirmation that this is either a bug or a deliberate change. The two reports agree on the useful part, though: links that consumers see in ChatGPT answers are often arriving without the attribution analytics teams expect.

Your direct traffic may now contain ChatGPT traffic

If ChatGPT removes utm_source=chatgpt.com but still passes a referrer, you may retain some source information. If it removes the referrer too, the visit lands in direct. That’s the ugly version because your analytics system can’t recreate information the browser never sent.

This matters beyond a marketing dashboard. If you’re an SE or solutions architect validating an analytics integration, attribution is often part of the acceptance criteria. A buyer may ask whether an AI discovery program increased qualified product-page visits, demo requests, or revenue. You can’t answer that cleanly if the source changes its link behavior halfway through the test.

I’ve worked across solutions engineering and frontend engineering at Feedonomics, FullStory, Shogun, and AutoZone. One lesson carries across those jobs: when a third-party system changes what it sends, you need to separate a broken integration from broken measurement before everyone starts fixing the wrong thing.

In this case, don’t let a source report alone tell you demand disappeared.

How I’d monitor it

First, I’d annotate October 9 in every affected dashboard. Include links to the Search Engine Roundtable and Profound reports, and say that OpenAI hasn’t confirmed the cause. Six weeks from now, nobody will remember why the line moved, and someone will confidently present the decline as a change in customer behavior.

Next, I’d build a temporary comparison using landing page, date, device, geography, engagement, and conversion behavior. Start with pages that historically received meaningful ChatGPT referrals. Compare their attributed ChatGPT traffic and direct traffic before and after October 9.

You’re looking for a pattern like Huesler found: ChatGPT falls, direct rises on the same pages, and the new direct cohort behaves more like the old ChatGPT cohort than your normal direct visitors. That won’t give you perfect attribution, but it’ll tell you whether the missing traffic probably moved buckets.

I’d also split new and returning visitors where your consent setup and analytics platform allow it. A returning customer typing your homepage into a browser isn’t the same as a new visitor landing deep on a product detail page. A sudden rise in new direct sessions entering through pages that ChatGPT previously cited is a useful signal.

Then I’d monitor server or CDN logs alongside client-side analytics. Logs won’t magically restore a stripped referrer, but they can help you verify that requests to the affected landing pages didn’t fall as sharply as the attributed channel did. If total requests and conversions stay steady while ChatGPT attribution drops, you’ve got more evidence that the label changed rather than the audience.

Finally, I’d save the raw daily data now. Profound correctly warns that restoring the tags would improve future reporting but wouldn’t recover the missing history. Don’t wait for OpenAI to fix this and assume your analytics platform can backfill what it never received.

Workarounds without pretending they’re fixes

You can’t force ChatGPT to preserve a UTM parameter or referrer on an outbound link. You can improve your own evidence.

For high-value conversion flows, capture the landing page and first-party session context before the visitor moves deeper into the site. Keep any self-reported attribution question short and optional. “How did you hear about us?” isn’t precise, but “ChatGPT or another AI assistant” can give you a directional check against the analytics trend.

If you run controlled campaigns or publish pages specifically intended for AI discovery, use distinct landing paths where that makes sense. Don’t create hundreds of thin pages just to manufacture attribution. A stable, useful destination gives you another signal when the browser strips the source.

You should also report a range rather than quietly relabeling all new direct traffic as ChatGPT. Build a conservative estimate from historically ChatGPT-heavy landing pages and matching behavior, then show attributed traffic separately from estimated unattributed traffic. Your stakeholders can see the uncertainty instead of getting a made-up exact number.

What I’d do now

I’d leave the existing channel definitions alone, annotate the change, and add a temporary “possible unattributed ChatGPT” view for the affected landing pages. I’d compare that cohort with historical ChatGPT visitors every day until OpenAI comments or the link behavior changes again.

I wouldn’t tell a client that ChatGPT traffic dropped based only on the referral line, and I wouldn’t tell them every extra direct visit came from ChatGPT either. Both claims go beyond the evidence. For now, the honest answer is that ChatGPT appears to have changed what its links send, and your measurement needs to account for that before you make a budget, product, or integration decision.