ChatGPT traffic dropped on 9 October 2026. Your visitors probably didn’t

The short answer

• What happened: GPT-6 reached paid ChatGPT users on 7 October 2026 and free users on 8 October. From 9 October, clicks from ChatGPT answers arrive without a referrer and without the utm_source=chatgpt.com tag.
• What you see: ChatGPT referral sessions down sharply, Direct sessions up on the same landing pages, total traffic roughly flat.
• What it means: most of the drop is reclassification. A smaller part may be real, because GPT-6 answers more inside the chat.
• What to do: annotate 9 October in your analytics, compare ChatGPT and Direct on the same pages, and stop using the ChatGPT referral count as your measure of AI visibility.

What changed in ChatGPT on 9 October 2026

Analytics tools learn where a visit came from in two ways: the browser sends a referrer (the page the visitor clicked from), or the link itself carries a tag such as utm_source=chatgpt.com. ChatGPT has added that tag to the links in its answers since 2025, which is why “chatgpt.com” became a traffic source many teams report on.

According to Dalin de Graff’s analysis, GPT-6 opens links with the browser setting “noreferrer” and without the UTM tag. Neither signal reaches your site, so analytics has nothing to attach the session to and files it under Direct. In a live test, every link ChatGPT opened arrived this way.

The change is specific to ChatGPT. LLM Pulse saw no break in Perplexity or Claude referrals on 9 October, and Gemini still tags its links.

Is the traffic gone, or just mislabelled?

Two public data sets point the same way:

• LLM Pulse measured daily ChatGPT sessions in GA4 across the sites it monitors: down about 46% between 9 and 11 October, while total traffic on the same sites moved about −2%. More than eight in ten sites lost over 30% of their ChatGPT sessions.
• Cedric Huesler’s comparison of about 500 Adobe pages where ChatGPT drove at least 5% of visits: from 9 to 13 October, ChatGPT-attributed visits fell by 31,000 a day while Direct visits rose by 26,000 a day on the same pages.

So roughly five out of six “lost” ChatGPT visits on those pages reappeared as Direct. The remaining gap may be real: LLM Pulse notes that GPT-6 answers more questions inside the chat, with widgets, maps and small tools, which can mean fewer clicks. Your own split will differ, which is why the check below is worth ten minutes.

How to check your own numbers in ten minutes

1. Annotate the date. Add a note on 8 to 9 October 2026 in GA4 or whatever you report from, so nobody reads the drop as a lost channel next quarter.
2. Compare the same pages. Take the landing pages that got the most ChatGPT visits in September. Compare 9 to 15 October with 1 to 7 October for two sources: ChatGPT and Direct. If ChatGPT fell and Direct rose by a similar amount, the visitors are still coming.
3. Check total traffic and conversions on those pages. Flat totals with a moved label mean nothing was lost.
4. Keep counting the engines that still tag. Perplexity, Claude and Gemini referrals still arrive with a source, so a channel group for them still works.
5. Ask people. Add “ChatGPT or another AI assistant” to the “How did you hear about us?” field on your sign-up or contact form. Self-reported answers are the one source a link change cannot remove.

What still works when the referral count breaks

A referral count was always a partial view of AI visibility. It only counts people who clicked, and most people who read an AI answer don’t click at all. Two measures do not depend on ChatGPT passing anything to your analytics:

• ChatGPT’s own visits to your site. OpenAI documents three crawlers. GPTBot collects pages for training, OAI-SearchBot builds ChatGPT’s search index, and ChatGPT-User opens a page when someone asks ChatGPT a question that needs it. Your server logs or CDN identify them by user agent, not by referrer, so GPT-6 did not hide them. ChatGPT-User is the closest signal to “ChatGPT used my page to answer someone”.
• The answers themselves. Ask ChatGPT the questions your buyers ask and record whether it names you, recommends you, or only cites your page as a source. Ask each question several times, because AI answers vary from run to run. This measures visibility before any click happens, so no change to links or analytics can move it.

More on choosing those questions in finding your customers’ AI prompts, and on what tracking covers in what AI visibility tracking is.

How Far & Wide handles it

Far & Wide was built on the second measure, so the GPT-6 change does not touch its core numbers:

• We measure the answers, not the clicks. Far & Wide asks the questions your buyers ask in the real web apps of ChatGPT, Gemini, Perplexity, Microsoft Copilot and Google AI Mode, logged out, from your buyers’ country and in their language, several times each, and records who is named, who is recommended and which pages are cited. Your visibility score comes from those answers, so a link or analytics change cannot move it.
• We read ChatGPT’s visits from your logs. When you connect Cloudflare or a Vercel log drain, the AI traffic page counts GPTBot, OAI-SearchBot and ChatGPT-User visits by user agent. Asking ChatGPT about your brand also makes ChatGPT-User open your site, so Far & Wide estimates how many of those visits its own measurement caused and shows that as a range rather than counting it as demand.
• We say what is affected. The one number on that page that reads the referrer, people who arrived by clicking a link in ChatGPT, is hit by the same change as your GA4: from 9 October it undercounts ChatGPT. Read it as a floor, and use ChatGPT-User visits and the measured answers for trends.

The broader point: if your report on AI visibility is a referral number, one product update can halve it overnight without a single buyer changing behaviour. Measuring the answers is what stays comparable from month to month. For how to tie that to results, see measuring AEO results and ROI; tools that do this tracking are compared on AEO tracking software.