The ChatGPT referral traffic decline that many sites saw in October 2026 was mostly a loss of attribution, not a loss of visitors. From about October 8, links in ChatGPT answers stopped carrying the utm_source=chatgpt.com tag and did not pass a referrer, so analytics tools filed those visits under Direct. The tag came back from October 15, but the referrer is still missing on most clicks, and the visits from the gap stay recorded as Direct.
If your ChatGPT line fell off a cliff in the second week of October, this article explains what changed, how large the shift was in public data, how to check your own numbers, and how to track AI referral traffic so the next change of this kind does not mislead you. It also describes what we did with the AI search share figures we publish, because the same change hit them.
Did ChatGPT referral traffic really drop in October 2026?
For most sites, no. The visits kept arriving; they stopped saying where they came from.
The clearest evidence comes from Cedric Huesler at Adobe, who looked at about 500 pages where ChatGPT had driven 5% or more of visits. In a LinkedIn post on October 13, he compared October 9 to 13 with the week before. On those pages, ChatGPT visits fell by about 31,000 a day while Direct visits rose by about 26,000 a day. Search traffic on the same pages fell about 7%, in line with the rest of the web. Direct rose on about two in three of the pages that lost ChatGPT traffic, against roughly a coin flip the week before.
He was careful about the limits: four days of sampled data, and only pages where ChatGPT was a large share of traffic could show the pattern. But the shape is hard to read any other way. One channel lost almost exactly what another gained, on the same pages, in the same days.
What changed in how ChatGPT links pass UTM and referrer?
Two separate signals tell an analytics tool that a visit came from ChatGPT, and both went missing at once.
The UTM tag. ChatGPT normally appends utm_source=chatgpt.com to the links it shows. Analytics tools read that parameter from the landing URL and assign the visit to ChatGPT, whatever the browser reports.
The referrer. When a browser follows a link from one site to another, it usually sends the origin of the page it came from in the Referer header. Under the browsers’ default policy, strict-origin-when-cross-origin, a click from a conversation sends only https://chatgpt.com/, not the conversation itself. That is enough for analytics to label the visit as a ChatGPT referral.
According to Huesler’s follow-up post, the tag stopped being added around 2 p.m. Pacific time on Thursday, October 8, and attributable ChatGPT visits fell to about a third of normal within hours. Sitebrunch reported that the links were being opened with window.open(url, "_blank", "noopener,noreferrer"). The noreferrer option tells the browser to omit the Referer header entirely. With no tag and no referrer, the visit has no source, and analytics tools file sourceless visits as Direct.
Huesler wrote that he asked OpenAI about it on the morning of October 14 and that OpenAI told him the same day it was a known bug, with a fix due that week. The fix began rolling out in stages around 2 a.m. Pacific on October 15. By 9 a.m. that day, tagged ChatGPT clicks in his telemetry were back to roughly 95% to 100% of the same hour a week earlier. Sitebrunch’s update of October 27 says almost all ChatGPT links have carried the tag again since October 16. We could not find a public statement from OpenAI itself; the confirmation reached the public through Huesler.
Two things did not come back. The referrer is still missing on most ChatGPT clicks, so the tag is now the only reliable sign of the source. And visits recorded during the gap stay in your data as Direct. Analytics tools do not reprocess past sessions when a tag reappears.
How big was the shift?
Large enough to change the picture of the whole AI referral market. Statcounter publishes AI chatbot referral shares from the sites that run its tracker. We downloaded its weekly worldwide figures for all devices on October 27, 2026. Weeks run Monday to Sunday.
| Week | ChatGPT | Gemini | Perplexity |
|---|---|---|---|
| Sep 28 - Oct 4 | 78.77% | 12.51% | 6.21% |
| Oct 5 - Oct 11 | 67.96% | 18.73% | 9.67% |
| Oct 12 - Oct 18 | 19.70% | 47.31% | 25.10% |
| Oct 19 - Oct 25 | 18.13% | 49.85% | 25.92% |
ChatGPT’s share fell by about 60 points in three weeks. The partial drop in the week of October 5 matches a break that started on Thursday, October 8.
The last row is the interesting one. Tagging came back on October 15 and 16, yet ChatGPT’s share in the week of October 19 was lower still. A count that relies on the referrer rather than the UTM tag would behave exactly this way, because the referrer is the part that has not returned. We have not confirmed which signal Statcounter uses, so treat that as a likely explanation, not a finding. Either way, a referrer-based report is the slowest kind to recover.
Why did Gemini and Perplexity look like they grew?
Because shares are fractions. When most of ChatGPT’s referrals disappear from the numerator, they also disappear from the denominator, and every other engine’s slice grows even if its own traffic did not change at all.
Gemini went from 12.51% to 49.85% of Statcounter’s AI referrals in four weeks, and Perplexity from 6.21% to 25.92%. Nothing in the reporting above suggests those engines quadrupled their traffic in October. A site that sees Gemini rise as a share of AI referrals should look at Gemini’s raw sessions before reading it as growth. If the raw count is flat, the share moved because ChatGPT left the denominator.
This is the general lesson of the episode. A single engine changing how it opens links can redraw a market-share chart without a single user changing behavior.
How do you check whether your own AI referral traffic was affected?
You can test it in your own analytics. The idea is the one Huesler used: if ChatGPT visits were relabeled, one channel should lose about what another gains, on the same pages.
- Pick the pages ChatGPT used to send people to. In your analytics tool, list the landing pages with the most sessions from source
chatgpt.comin a normal week before October 8. Deep pages such as documentation, help articles, and product details are the best test, because few people type those URLs by hand. - Compare two windows on those pages. Put October 1 to 7 next to October 9 to 14, and look at sessions by source. If ChatGPT sessions fell and Direct sessions rose by a similar amount on the same pages, the drop was attribution.
- Check the recovery. Look at ChatGPT sessions from October 16 onward. If tagged sessions came back while Direct fell back, your tool is reading the UTM tag and you are now seeing most ChatGPT visits again. If ChatGPT stayed low, check whether your reports or filters rely on the referrer rather than the landing URL’s parameters.
- Annotate the gap. Mark October 8 to 15 in your analytics tool. Huesler suggests doing this in GA4 or Adobe Analytics, so anyone reading a year-over-year chart in 2027 knows why October dipped.
- Report ChatGPT and Direct together for that window. Moving visits from Direct back to ChatGPT in a report is an estimate, not a measurement. If you do it, label it as one.
If your pages never received much ChatGPT traffic, the comparison will be noise. Direct traffic moves for many reasons, and the pattern only shows where ChatGPT was a meaningful share to begin with.
How should you track AI referral traffic from now on?
Assume this will happen again. AI assistants change how they render answers and open links, and this change was not announced; site owners found it in their own charts. We made the broader case in GEO looks like early SEO, except the ground keeps moving. A few habits make the next change visible within days instead of a month.
Rely on the tag first and the referrer second. For now, utm_source=chatgpt.com on the landing URL is the more reliable signal. Make sure your channel rules read it, and that nothing on your site strips query parameters before the analytics tag fires.
Watch Direct on deep pages. A sudden rise in Direct visits to pages nobody types by hand is the earliest sign that some source stopped identifying itself. A simple weekly check of Direct sessions on your top AI landing pages would have caught this break within days.
Read shares and raw counts side by side. A share chart can move because of one engine’s bug. A raw count of sessions per engine shows which engine actually changed.
Measure the answers, not only the clicks. Referral traffic tells you how many people clicked through. It cannot tell you whether an assistant mentioned or cited you when they did not click, and as October showed, it can lose its labels overnight. Generative engine optimization is about appearing in the answer itself, so the most direct measure is to ask the engines the questions your buyers ask and read what comes back. That is what the extension and the console do, and it is why a change in how ChatGPT opens links leaves your AI visibility readings unaffected: they never depended on the click. What the score means and what it leaves out is on our measurement page.
None of these replaces referral analytics. Clicks are where revenue starts. But a team that watches both can tell “ChatGPT stopped recommending us” apart from “ChatGPT stopped telling us it sent the visitor”, which were indistinguishable in a referral chart for a week in October.
What did we do with our AI search share numbers?
We publish an AI search share that combines four public sources into one estimate of how much each AI assistant is used. Statcounter is one of them. On October 27, the first run of our new weekly refresh replaced Statcounter’s September figure with the average of its four most recent weeks, the weeks in the table above.
Had we published that result, the combined share for ChatGPT would have fallen from 65.9% to 47.0%, and Perplexity would have risen from 5.6% to 15.1%. Nothing about how people use these assistants changed that much in a month. A tracking bug did.
So we held it. The weekly refresh is built to stop when one engine moves sharply in a single week: the update waits for a person, and the cause is looked up in primary reporting before anyone accepts it. In this case the cause was documented, and we kept the page on Statcounter’s September figures. If the weekly figures return to their earlier range, the refresh resumes on its own. If they do not, a person decides what the page should show.
We caught this on the first weekly run, held the numbers the same day, and wrote this article so the reasoning is on the record. That is how we want to work. AI assistants will keep changing how they search, cite, and link out, usually without notice, and each change will show up first as a strange number. We would rather find the strange number ourselves, trace it to its cause, and tell you, than let it pass into a chart you rely on.
The same habit runs through the rest of the product. Our AI search share is refreshed every week instead of being copied by hand. The mention detector is versioned, so when we fix how a name is recognized, old and new numbers are never quietly mixed. The recommended steps are checked against the engines’ own documentation and new research every month, and a step the evidence no longer supports is revised or removed. None of this is finished, and it is not meant to be. A measurement of something that moves this often is only as good as the team that keeps correcting it.
The habit worth building on your side is the same one: treat a sharp move as a question about the measurement before treating it as news about the market.
FAQ
Did ChatGPT stop sending traffic to websites in October 2026?
Not for most sites. From about October 8, ChatGPT links lost their UTM tag and referrer, so analytics counted the visits as Direct. Adobe’s Cedric Huesler found that on pages ChatGPT used to drive, ChatGPT visits fell by about 31,000 a day while Direct rose by about 26,000 a day.
Is ChatGPT referral tracking fixed now?
Partly. The utm_source=chatgpt.com tag returned from October 15, and almost all links carried it again from October 16. The referrer is still missing on most clicks, so reports that rely on the referrer may still undercount ChatGPT.
Can I recover the ChatGPT visits that were counted as Direct?
Not in the raw data. Analytics tools do not reprocess past sessions, so visits from October 8 to 15 stay as Direct. You can estimate them by comparing ChatGPT and Direct sessions on the same landing pages before and during the gap, and label that figure as an estimate.
Why did Gemini and Perplexity’s share of AI referrals jump?
Mostly because ChatGPT’s visits dropped out of the total. When the largest engine’s referrals stop being counted, every other engine’s share rises even if its own traffic is flat. Check raw sessions per engine before reading a share increase as growth.
