ChatGPT traffic showed up as Direct for a week. How to repair your October numbers

KEY TAKEAWAYS

1. The 8 to 15 October drop in ChatGPT referrals was mostly a labelling bug, not lost visitors.
2. Estimate the moved visits only on deep pages where ChatGPT was at least 5% of sessions. The homepage cannot be separated.
3. After the fix, the tag is the only ChatGPT signal on most clicks. It undercounts when it fails and overcounts when Google indexes tagged URLs.
4. Report ChatGPT on session source, check it weekly against citation data, and never judge AI visibility on referral counts alone.

For about a week in October 2026, ChatGPT referrals in most analytics tools fell by a third or more. The visitors did not stop coming. Their clicks arrived unlabelled and were filed as Direct. The tag is back, but the week is still wrong in your reports, and the reason it broke has not gone away. Here is how to repair the gap and make the next break cost less.

What happened between 8 and 15 October?

GPT-6 reached paid ChatGPT plans on 7 October and Free and Go users on 8 October. From about 2pm Pacific on 8 October, many links in answers stopped carrying utm_source=chatgpt.com. That is 9 October for most of Europe and Asia, which is why some reports date the start a day later.

Cedric Huesler of Adobe said OpenAI confirmed to him on 14 October that this was a known bug. A fix began rolling out at about 2am Pacific on 15 October. By 9am that day, he saw tagged ChatGPT clicks back at roughly 95% to 100% of the same hour a week earlier, across thousands of domains. That confirmation is secondhand. OpenAI has published no note about it, and its release notes for 1 to 15 October do not mention link tracking. Glenn Gabe reported that day that he was not yet seeing the fix on his accounts.

The working window, then, is 8 to 15 October, with the first and last days partial.

Why did the clicks land in Direct?

A visit needs one of two signals to be credited to ChatGPT: a referrer header or a tag in the URL. Most ChatGPT clicks had already lost the first. Ahrefs found in 2025 that inline links in paid accounts carried noreferrer, while the same links in free accounts did not. Huesler reported that GPT-6 opened inline links in a way that withheld the referrer. Ethan Smith of Graphite said 95% of mobile ChatGPT visits to the sites Graphite tracks relied on the tag alone.

So when the tag went, nothing was left. In GA4, Direct means source (direct) with medium (none) or (not set). An unlabelled ChatGPT click fits that rule exactly.

Profound offered a second cause. The share of ChatGPT answers with at least one tagged source link fell from about 98% on 8 October to 66% on 14 October, while answers using the new Intelligent UI rose from about 2% to about 35%. Links in side panels and shopping results kept their tags. Inline links often did not. The two explanations fit together. The new formats put more links inline, and inline links were the ones that lost the tag.

How big was the drop, and why do reports disagree?

Every panel saw a fall. None agree on its size.

The spread follows from what each panel holds. Sites with more mobile traffic and more inline citations depended more on the tag. A panel weighted toward desktop visits and shopping links would lose less. No single figure applies to your site. Measure your own.

Did Direct really go up?

This is where two sources disagree. Huesler looked at about 500 pages where ChatGPT drove at least 5% of visits. Comparing 9 to 13 October with the week before, ChatGPT visits fell by about 31,000 a day and Direct rose by about 26,000 a day. Direct rose on two in three pages that lost ChatGPT traffic, against roughly a coin flip the prior week. LLM Pulse found no measurable rise in Direct at all.

Both can be right. Huesler chose pages where ChatGPT was a real share of traffic, so the moved visits stood out. LLM Pulse looked at sites in aggregate, where Direct is far larger than ChatGPT and a shift of that size disappears into normal variation. The evidence supports relabelling, but you can only see it where ChatGPT mattered to begin with.

One more point from Huesler's numbers. If 26,000 of 31,000 lost visits reappeared as Direct, about 84% moved and about 16% did not show up there. Some may be real decline. Some may sit in Unassigned or elsewhere. Do not assume the whole drop was relabelled.

Why can't you repair the homepage?

Peec AI ran a million prompts before and after the change. During the gap, direct URLs appeared in 42% of responses, against about 8% to 10% before. About 54% of those URLs were homepages. These are links shown, not clicks, but they point to where the moved visits went.

Your homepage carries the most Direct traffic on the site: typed URLs, bookmarks, untagged email, returning visitors. A few dozen extra visits a day sit well inside its normal swing. Any estimate there would be a guess presented as a number. Annotate the homepage and leave it.

How do you estimate what moved on deep pages?

The method below is a two-week baseline comparison, the same logic Huesler and LLM Pulse used, applied page by page. Work in GA4 Explorations or a BigQuery export.

1. Set the window. Use 9 to 14 October as full gap days in your reporting time zone. Treat 8 and 15 October as partial and exclude them from the estimate.
2. Pick the pages. List landing pages with sessions whose source contains chatgpt in the four weeks before 8 October. Keep pages where ChatGPT was at least 5% of sessions. Drop the homepage.
3. Build the baseline. For each page, take mean daily ChatGPT and Direct sessions on the same weekdays in the two weeks before the gap.
4. Measure normal noise. Note how much Direct moved day to day on each page during the baseline. A Direct gain smaller than that range is not evidence of anything.
5. Calculate. ChatGPT lost equals baseline minus gap-period ChatGPT. Direct gained equals gap-period Direct minus baseline. The moved estimate is the smaller of the two, counted only where the Direct gain clears the noise range.
6. Check the visitors. Compare new-visitor share and device mix of Direct during the gap against the baseline. Huesler saw the share of new mobile Direct visitors rise from 36% to 44%, toward the 55% typical of ChatGPT visitors. A similar shift on your pages supports the estimate.
7. Store it separately. Add the estimate as its own labelled row, not merged into raw ChatGPT sessions. Anyone reading the report should see which numbers are measured and which are inferred.

A worked example (hypothetical)

Take a software company with 40 guide and comparison pages that ChatGPT cites regularly. In the two baseline weeks, those pages averaged 200 ChatGPT sessions and 500 Direct sessions a day. Direct normally moved by about 40 sessions either side of its mean.

From 9 to 14 October, the same pages averaged 70 ChatGPT sessions and 610 Direct sessions a day. ChatGPT lost 130 a day. Direct gained 110 a day, well outside its normal range of 40. The moved estimate is 110 a day, or 660 sessions over the six full days. Estimated ChatGPT traffic for the gap is 70 plus 110, which is 180 a day, or 90% of the baseline. The remaining 20 a day stay unexplained and are reported as such.

Now the homepage. It averaged 5,000 Direct sessions a day, moving by about 300 either way. A gain of 110 is invisible there, so the estimate stops at the 40 deep pages.

What should you do with each pattern?

Why is ChatGPT attribution still fragile after the fix?

The fix restored the tag, not the referrer. Huesler noted that most ChatGPT clicks still arrive without one. One parameter now carries almost all of your ChatGPT attribution, and the past week showed it can vanish in a single release.

It also fails in the other direction. On 1 October, Aashish Kasma found Google ranking a Disney+ login URL that still carried utm_source=chatgpt.com. Tags copied into public pages get crawled. Because GA4 lets a URL tag override the referrer, a Google click on that result is credited to ChatGPT.

And it may not land where you expect. Google's AI Assistant channel requires the medium ai-assistant, which GA4 sets when the referrer matches a known assistant. A visit with the tag but no referrer has a source and no medium, so it usually ends up in Unassigned. Google's own help page for custom channel groups says to place an AI channel above Referrals, because traffic goes to the first channel it matches.

Three habits follow. Report ChatGPT on session source containing chatgpt, not on the channel. Check monthly for tagged sessions with a search engine referrer. Read referral counts next to citation data, since a fall in clicks with steady citations points to measurement, not lost visibility. Limy saw tagged referrals fall 65% while citations moved about 6%. Our piece on measuring AI visibility when answers change covers how to sample those citations properly.

How does this fit a wider measurement plan?

Referral counts were always a floor. Only some answers are clicked, and some clicks never carry a label. One widely shared estimate claims that 60% of AI traffic hides in Direct. David Rosam traced it to a single vendor's estimate, later replaced by 70.6% of AI visits arriving without a referrer in that vendor's own customer data. Neither figure tells you how much of your Direct traffic is AI.

The sturdier view combines three layers: citations and mentions in answers, referral sessions by source, and server logs. OpenAI documents ChatGPT-User as the agent that fetches pages when a user's request needs them, separate from OAI-SearchBot crawling. Those fetches are not human visits, but logs show which pages ChatGPT reads for its users, a count that does not depend on tags. Our guide to AI crawler access explains how to separate those agents. For Google's surfaces, the Search Console report covered in what an AI impression is worth adds a fourth view. Gemini began tagging its own links on about 1 October, with values Google has not documented, so expect the same kind of break there.

This is the measurement pillar of digital authority: knowing whether you are named, by which engine, and what that is worth, without relying on one fragile signal. If you want that baseline set up for your brand, join the waitlist.

Questions

Is ChatGPT traffic still showing as Direct after the fix?

Mostly not. Adobe's Cedric Huesler reported that tagged ChatGPT clicks were back to roughly 95% to 100% of the prior week by 9am Pacific on 15 October. Some practitioners had not seen the fix yet that day, so check your own tagged sessions against the same weekday a week earlier before you close the annotation.

Can I recover the visits that were counted as Direct?

Not in the raw data. Sessions recorded as Direct stay Direct, and no backfill has been announced. You can estimate the moved visits for deep pages by comparing ChatGPT and Direct against a two-week baseline, then keep that estimate in a separate, labelled row of your report.

Why does the drop look smaller on some reports than others?

Panels differ in device mix, link types and pages. Published declines ran from about 22% to about two thirds. Mobile visits and inline answer links depended most on the tag, so sites with more of both lost more of their labelled ChatGPT traffic.

Why do tagged ChatGPT visits appear in Unassigned in GA4?

The tag sets a source but no medium. Google's AI Assistant channel requires the medium ai-assistant, which GA4 sets when the referrer matches a known assistant. A visit with the tag but no referrer usually lands in Unassigned. Report on session source instead, or build a custom channel placed above Referral.

Sources

1. Search Engine Journal, ChatGPT traffic drop appears to be a tracking problem searchenginejournal.com
2. Search Engine Land, ChatGPT appears to drop UTM tracking, making AI referral traffic harder to measure searchengineland.com
3. Search Engine Roundtable, OpenAI drops UTM parameters from ChatGPT results (GPT-6) seroundtable.com
4. Relevant Audience, ChatGPT UTM tag loss was a bug, fix rolling out relevantaudience.com
5. StartupTalky Community, ChatGPT fixed the UTM issue (Tomek Rudzki, Peec AI) community.startuptalky.com
6. The Prompt Insider, The day ChatGPT traffic disappeared from analytics, and where it went thepromptinsider.com
7. Limy, ChatGPT didn't stop sending traffic. It stopped signing it limy.ai
8. Google Analytics Help, Default channel group support.google.com
9. Google Analytics Help, What's new in Google Analytics support.google.com
10. Google Analytics Help, Custom channel groups support.google.com
11. OpenAI, Overview of OpenAI crawlers developers.openai.com
12. Ahrefs, Generative engines are breaking web analytics (Spanish edition) ahrefs.com
13. Priority Pixels, ChatGPT traffic showing as Direct in GA4 and how to track it prioritypixels.co.uk
14. Medium (Aashish Kasma), Google is indexing ChatGPT's tracking links and it's corrupting your AI traffic data medium.com
15. Search Engine Journal, Google Gemini adds UTM parameters for referral attribution searchenginejournal.com
16. David Rosam, Everyone quotes 60%. Nobody measured it buttondown.com

NEXT STEP

See where your brand stands.

Your Digital Authority Score, who AI recommends for your target questions, and the placements that close each gap. Your first report is free.

Get your free report →