Key takeaways
- Around October 8-9, 2026, many ChatGPT outbound links stopped carrying
utm_source=chatgpt.com. Most ChatGPT clicks already had no referrer, so GA4 filed those visits under Direct. - Your ChatGPT line probably fell. Your actual visitors probably did not. LLM Pulse saw ChatGPT sessions fall about 46% on monitored sites while total traffic stayed flat (about -2%).
- Treat October 8-16 as a candidate disruption window, narrow it with your own data, and annotate it in every report people actually read.
- Do not reassign Direct sessions to ChatGPT without independent evidence. Preserve recorded counts and label any model as estimated.
- Stop depending on a single tag. Pair source data with citation tracking, landing-page patterns, and a self-reported field.
What actually happened
GPT-6 reached ChatGPT paid plans on October 7 and free plans on October 8, 2026. It came with a new interactive answer format OpenAI calls Intelligent UI: widgets, maps, product panels, small tools. According to LLM Pulse's analysis, the first signs of a traffic drop showed up on October 8 and the full change landed on Friday, October 9.
The mechanism, found by Adobe's Cedric Huesler, is boring and nasty. Links in the new format open in a new window with noopener and noreferrer, and without utm_source=chatgpt.com. No tag and no referrer means GA4 has nothing to identify, so the session becomes Direct.
That undid the UTM tagging OpenAI had rolled out to more ChatGPT links earlier. Search Engine Roundtable covered the finding.
The numbers worth knowing
- LLM Pulse: daily ChatGPT sessions fell about 46% across October 9-11 versus the same days in the prior two weeks. More than 8 in 10 monitored sites lost over 30%. Perplexity and Claude referrals showed no break.
- Huesler looked at about 500 pages where ChatGPT drove at least 5% of visits. Comparing October 9-13 with the week before, ChatGPT lost about 31,000 visits a day while Direct gained about 26,000 a day. Direct grew on two in three of the pages that lost ChatGPT traffic.
- Profound reported teams seeing 30% to 40% declines in reported ChatGPT referrals. Its tracking of collected answers showed the share with at least one tagged source link falling from about 99% to roughly 5-12% at the worst point on October 14-15.
- Webflow reported 56% fewer ChatGPT-attributed visits on its own site.
None of these samples is a census. Sample sizes are mostly undisclosed, and each source measures something slightly different. Read them as direction, not as multipliers for your own data.
Did it recover?
Partially, and the sources disagree. Arcalea says tagging recovered by October 16, and Profound saw about 99% of collected answers with a tagged link again by 06:30 UTC that day. LLM Pulse, writing around the same time, said the tag was still missing on roughly a third or more of answers. Behavior is also uneven: links reached after opening an extra panel inside an answer can still carry the tag, while a link shown directly may not. That is why GA4's ChatGPT line fell by about half rather than to zero.
OpenAI had not publicly commented as of mid-October, per LLM Pulse. Adobe's Huesler says OpenAI told him it was a bug. Treat the fix as partial and verify it on your own site.
Why GA4 files these visits as Direct
A visit with no referrer lands in Direct for several ordinary reasons: pasted links, the mobile app, in-app browsers that strip the referrer, and noreferrer links. GA4 shows only the ChatGPT traffic that successfully passed source information, which makes it a minimum baseline rather than a count.
The October change did not create this weakness. It removed the one thing papering over it. Arcalea notes that most ChatGPT clicks still arrive without a referrer, so the UTM tag had become the only marker identifying them. Pull the tag and the traffic vanishes from the channel report.
A few platform quirks make it messier:
- Adobe Analytics may show a visit with
utm_source=chatgpt.combut no referrer as Direct, while GA4 may classify it as Unassigned or Referral. - ChatGPT's referrer policy (
strict-origin-when-cross-origin) passes origin only, so even when a referral appears you get no path context. - iOS uses WKWebView and Android uses Chrome Custom Tabs or in-app WebViews, both of which can drop the referrer on external link taps.
Step 1: confirm whether you were affected
Don't assume. Plenty of sites will see the pattern, but your own data decides.
Compare the same landing pages across sources
Pick the pages that historically get ChatGPT traffic. Compare a matched baseline (same weekdays) against October 8-16, and look at ChatGPT, Direct, and total sessions side by side. Convert UTC times into your reporting timezone first. A calendar day in your property can straddle two UTC days.
| Pattern on the same page cohort | Reading | What it does not prove |
|---|---|---|
| ChatGPT sessions fall, Direct rises, total sessions and key events stay stable | Attribution loss is plausible | That every extra Direct session came from ChatGPT |
| ChatGPT sessions rise after tags return, total sessions unchanged | Better attribution explains part of the rebound | That demand grew by the same percentage |
| ChatGPT falls, Direct flat, total sessions fall | A real traffic decline is plausible | Anything about cause. Check other channels |
Screpy's worked example is a good gut check. Baseline: ChatGPT 200, Direct 800, other 1,000, total 2,000. Affected period: ChatGPT 120, Direct 875, other 1,005, total 2,000. ChatGPT lost 80, Direct gained 75, and the total did not move.
Be careful with the Direct bump. An untagged newsletter can inflate Direct at the same moment real ChatGPT visits decline, so a rise in Direct is supporting evidence, not proof.
Run the cheaper cross-checks
- Look at new-user traffic on pages that previously got AI referrals.
- Check Direct traffic to the deep pages ChatGPT tends to cite. Homepage Direct is noise. Direct to a niche comparison page is a signal.
- Watch branded search trends over the same window.
- Confirm Perplexity and Claude referrals did not move. If they stayed level and only ChatGPT broke, you are looking at a tracking event.
Step 2: fix your channel rules in GA4
Arcalea's advice here is the most actionable: confirm whether your reports classify ChatGPT by referrer or by session source. If they lean on the referrer alone, you lose everything the moment it goes missing.
- Open Traffic acquisition and filter session source for
chatgpt.com,chat.openai.com, andopenai. - Check which default channel those sessions fall into. Some get misbucketed under Referral or Unassigned.
- Build a custom channel group with a source-based AI rule (a regex covering chatgpt, openai, perplexity, claude, gemini, copilot) and place it ahead of the default Referral rule.
- GA4 now has an AI Assistant channel. Check what it catches and what it misses, and document your own definition of identifiable AI traffic so everyone reports the same number.
- Save the segment so you can compare AI-referred sessions against your site average on conversion rate, engagement, and revenue.
A source-based rule only catches visits that arrive with a source. It cannot rescue sessions that arrive with nothing. That is why the next steps matter.
Step 3: annotate the window where people will see it
Annotate October 8 through 16 in GA4. Then do the part most teams skip. Pravin Kumar makes the point well: an annotation buried in GA4 does nothing if the founder reads a Looker Studio dashboard or a weekly HubSpot email. Put a short note beside the chart in the report itself that states the dates and the reason.
If you want a structure, Screpy publishes a note template. Condensed, it looks like this:
October 2026 ChatGPT attribution note
External context: published monitoring observed missing ChatGPT
source tags during October 8-16, recovery varying by observation.
Our scope: property, timezone, landing-page cohort,
baseline dates (matched weekdays), affected period on our site.
Recorded change (before -> during):
Identifiable ChatGPT sessions
Direct sessions on the same cohort
All sessions on the cohort
Key events or revenue
Concurrent site, campaign or consent changes
Interpretation: attribution disruption may explain part of
the ChatGPT movement. True visitor change is unknown unless
supported by separate evidence.
Reporting treatment: preserve recorded counts. Do not reassign
Direct without independent source evidence. Label modeled
scenarios as estimated.
Next verification: owner, date, current referral check.
The rule that matters most is the one in the middle: preserve recorded counts. If you want to show a scenario where Direct gave back 80 sessions, label it estimated and keep the raw numbers next to it. Replacing history with a guess presented as fact is how reporting loses trust.
Step 4: stop relying on one signal
The real lesson from October is that any AI measurement plan built on one platform-controlled tag will break eventually. A second witness is cheap.
Layer your attribution signals
| Signal | What it catches | Weakness |
|---|---|---|
| Source and UTM in analytics | Tagged and referred visits | Breaks when the tag or referrer disappears |
| Direct traffic to deep cited pages | Untagged AI visits, roughly | Newsletters, bookmarks, and dark social add noise |
| Self-reported field on forms | Buyers who say "I found you via ChatGPT" | Memory and incomplete completion |
| Question on sales calls | Qualified-lead attribution | Manual, inconsistent |
| Citation and mention tracking | Exposure inside the answer, before any click | Not a traffic or revenue number |
On the systems side, Pravin Kumar suggests storing the raw first-touch source, the referrer, and the self-reported answer as separate CRM properties in HubSpot or Salesforce instead of blending them into one field. Then you can see UTM-tagged ChatGPT visits drop while self-reported AI mentions hold level. That divergence is the exact signature of a tracking change rather than a demand change.
Measure exposure, not just clicks
A click is a small slice of what an AI answer delivers. Promptwatch Data found that 47.6% of cited ChatGPT responses cite 10 or more unique domains (Sep 28 to Oct 4, 2026), so any one source gets a thin share of the clicks even when it is in the answer. If you only watch post-click traffic, a tracking glitch and a visibility loss look identical.
That is the argument for tracking citations and visibility directly. Tools like Promptwatch monitor how ChatGPT and other engines mention and cite your brand, so those numbers do not depend on what happens after the click. Promptwatch also has visitor analytics for AI-referred traffic, and it added an optional setting after October 8 that counts visits tagged utm_source=chatgpt.com as ChatGPT traffic. Its attribution guide is a good read on combining signals.
Other options in the AI visibility category include Profound, whose tagged-link analysis produced much of the public timeline, and LLM Pulse, whose monitored-sites data is cited above.
For server-side evidence, AI crawler logs help separate a crawl or access problem from a content problem, though they show bots, not human clicks. Promptwatch's AI crawler logs are one way to see that, and DarkVisitors tracks AI agents and LLM referrals visiting a site.
If you are comparing options more broadly, the GEO software directory lists the field.
Step 5: tag what you control
You cannot tag ChatGPT's links, but you can tag everything else, which stops your own untagged traffic from muddying Direct.
- Add UTMs to every newsletter, social post, and email link. Otherwise your Direct baseline is contaminated and the "Direct rose by X" test gets weaker.
- For ChatGPT ads, static UTMs on every destination URL survive referrer loss. Promptwatch Data found ads in an average of 20.1% of ChatGPT web-search responses between May 20 and August 17, 2026, so paid clicks are a separate attribution problem from organic citation clicks. The DigitalApplied playbook adds that OpenAI's
opprefclick identifier should be captured at landing and attached to Conversions API events. - Treat server-side tracking and CRM-stage attribution as the backstop. Tools in our catalog such as Cometly and Dreamdata sit in this attribution category, though I have not tested how either handled the October gap.
A caution on the new interface
One thing no analytics tweak can settle: Intelligent UI may also answer more questions inside the chat and send fewer clicks in the first place. LLM Pulse flags this. So some of the drop could be real. You can't tell from analytics alone, which is one more reason to track citations and self-reported discovery alongside sessions.
What to do this week
- Pull ChatGPT, Direct, and total sessions for your top AI-landing pages, matched weekdays, October 8-16 versus baseline.
- Annotate the window in GA4 and in the dashboard or email your decision-makers read.
- Check that your channel rules key off source, not referrer.
- Compare tagged ChatGPT sessions from the last two weeks with early October to see if tagging truly recovered on your site.
- Add one non-tag signal: a self-reported field, or citation tracking.
- Keep your AEO strategy as it is. The evidence points to a labeling problem, so cutting investment would be reacting to a tracking bug.
The history will not come back. Even if tagging fully returns, the missing October days stay mislabeled unless someone writes down what happened. A one-paragraph note next to the chart is the cheapest fix you will do all year.