Your ChatGPT sessions in GA4 fell by roughly half around Oct. 8, and someone is about to ask why. Before you tell them AI visibility collapsed, check where those visits went. Several analysts say a change in how ChatGPT links open stripped the tracking tag, which can push visits into Direct instead of removing them.
This guide gives you a repeatable audit to size that effect on your own property, plus report wording that claims no more than the data supports. The audit cannot recover lost attribution, and it says nothing about rankings.
What the early reports say
Treat everything below as early, partly vendor-sourced and unsettled. The Search Engine Journal report collects the findings. It notes that Graphite, LLM Pulse, Profound and Peec AI sell SEO or AI search services, and that Cedric Huesler works at Adobe. It also says the LLM Pulse and Profound posts give no sample sizes.
Graphite CEO Ethan Smith reported that ChatGPT-attributed GA4 sessions fell 22% to 57% on sites with meaningful AI Assistant traffic, starting Oct. 9. He estimates the fall at roughly 45% from the previous baseline. He says inline links in answer text lost the utm_source=chatgpt.com tag and the referrer, while links in the Sources panel and shopping links kept the tag. LLM Pulse saw about 46% fewer ChatGPT sessions on Oct. 9–11 than on the same days over the prior two weeks. Total traffic on those sites stayed relatively constant.
Huesler compared Oct. 9–13 with the week before on about 500 pages where ChatGPT drove at least 5% of visits. He reported ChatGPT visits down about 31,000 a day and Direct up about 26,000 a day, with search down 7%. He calls his results early and based on four days of sampled data.
Profound found that the share of cited ChatGPT answers with at least one tagged source link fell from about 98% on Oct. 8 to 66% by Oct. 14. Over the same days, Intelligent UI (ChatGPT's format for charts, comparisons and small tools) rose from 2% to about 35% of the answers Profound collected.
Peec AI's Tomek Rudzki ran the same 1 million prompts twice, Oct. 2–7 and Oct. 9–13. ChatGPT put a direct page URL into 42% of responses, up from about 8–10%, and 54% of those URLs pointed to home pages. Those are links displayed, not clicks.
The gap that matters: LLM Pulse did not observe Direct rise. It noted that Direct is so much larger than ChatGPT that a small shift would be hard to detect. Its aggregated data also could not separate lost visits from visits no longer labeled as ChatGPT. Your audit tries to answer that question for one site.
Step 1: Fix your comparison windows
Pick two equal-length windows so weekday patterns cancel out. Oct. 1–7 as the baseline and Oct. 8–14 as the affected window work well, since each covers seven full days. Review Oct. 15 on its own, because a fix reportedly began rolling out that day (more on that below).
Write the windows down before you pull any numbers. Changing them after you see the data is how audits turn into storytelling.
Step 2: Pull ChatGPT sessions by landing page
In GA4, open Explore and build a free-form report. Use Landing page + query string as the rows, and Sessions, New users and Engaged sessions as the values. Filter Session source to values containing chatgpt.com. If your property has the AI Assistant channel, you can filter on that instead, but the source filter is more precise.
Export the top landing pages, ideally everything that earned at least a handful of ChatGPT sessions in the baseline. Include your home page, since more than half of the bare URLs Rudzki saw pointed there.
Step 3: Pull Direct for the same pages and windows
Repeat the report with Session default channel group = Direct. Keep the same landing pages, the same two windows and the same metrics. Add Device category as a second breakdown for the mobile check in Step 5. This pairs well with ai video ad workflow audit for agencies: 3-step checklist.
Then build a control: Direct sessions for pages that never received ChatGPT traffic. If those pages also swung up or down, your Direct movement has another cause, such as a campaign, an email send or a seasonal shift.
Step 4: Compare ChatGPT and Direct per page
Put the numbers side by side and calculate the change for each channel on each page. You want three readings:
- How many pages lost ChatGPT sessions.
- How many of those pages gained Direct sessions.
- How much of the ChatGPT loss the Direct gain covers, after you subtract the control's normal drift.
Worked example (hypothetical numbers)
The figures below are invented for illustration. They come from no real site and from none of the cited analysts. Sessions are seven-day totals.
| Landing page | ChatGPT Oct. 1–7 | ChatGPT Oct. 8–14 | ChatGPT change | Direct Oct. 1–7 | Direct Oct. 8–14 | Direct change | |---|---|---|---|---|---|---| | / | 420 | 240 | -180 | 3,100 | 3,260 | +160 | | /pricing/ | 150 | 85 | -65 | 900 | 940 | +40 | | /guides/crm-setup/ | 310 | 170 | -140 | 640 | 740 | +100 | | /blog/seo-audit-checklist/ | 260 | 150 | -110 | 410 | 450 | +40 | | /tools/roi-calculator/ | 90 | 80 | -10 | 520 | 500 | -20 | | Total | 1,230 | 725 | -505 | 5,570 | 5,890 | +320 |
ChatGPT sessions fell by 505 (about 41%). Direct rose by 320 on the same pages, and it rose on four of the five.
Now suppose the control pages' Direct sessions drifted up 1% over the same period (also hypothetical). Applying that drift to these pages' Direct baseline accounts for about 56 of the 320 sessions. That leaves roughly 264 excess Direct sessions, which covers about 52% of the ChatGPT drop.
Step 5: Check the Direct shift against behavior signals
Excess Direct is only a hint. Next, check whether the new Direct visitors look like ChatGPT visitors.
- First-visit share: Compare New users as a percentage of Direct sessions on these pages before and after. Huesler used the consent banner as his first-visit signal. He saw the share among mobile Direct visitors move from 36% to 44%, toward the 55% he sees for ChatGPT visitors. GA4's New users metric is the closest widely available stand-in, though a blunter one.
- Mobile share: Compare Device category for Direct before and after. Smith says 95% of mobile ChatGPT visits to Graphite's sites relied on the tag alone, which makes app traffic the most exposed.
- Landing pages: Check whether the Direct gain concentrates on pages ChatGPT used to cite, and on the home page, rather than spreading evenly across the site.
- Total traffic: Check whether all-channel sessions held steady. LLM Pulse reported flat totals on the sites it monitors. A flat total with a ChatGPT drop and a Direct rise is the pattern relabeling would produce, while a falling total points elsewhere.
In the hypothetical example, the new-user share of Direct on these pages moves from 38% to 44%, mobile share from 45% to 52%, and total sessions from 41,200 to 40,900 (all invented). All three would be consistent with relabeling.
A rule of thumb for reading the result
This rule is an inference, not a validated standard. Call the drop "consistent with misattribution" if all of the following hold: We cover related ground in how to build an ai visibility prompt set by tag (template) in depth.
- Excess Direct covers about half or more of the ChatGPT drop.
- Direct rose on most pages that lost ChatGPT sessions.
- New-user and mobile shares moved up.
- Total traffic held.
If Direct stayed flat or the total fell, call it "unresolved, possible real loss." Either way, whatever Direct does not explain stays unresolved.
Small sites should be cautious. A few dozen ChatGPT sessions per page make every percentage noisy, and a large Direct baseline can hide a shift, the same problem LLM Pulse flagged.
Step 6: Check whether tagged sessions are returning
Huesler reported that OpenAI confirmed a bug to him on Oct. 14, and that a fix began rolling out early Oct. 15. By 9 a.m. PT, he said, tagged ChatGPT clicks were back to roughly 95% to 100% of the same hour a week earlier, across thousands of domains he tracks.
Reports conflict. SEO consultant Glenn Gabe replied that he was not yet seeing the change on paid or free accounts. Profound says it has seen no official word on a fix, and OpenAI's release notes for Oct. 1 through Oct. 15 do not mention link tracking. Verify on your own data.
Build a free-form report with Date and Hour as rows, filtered to chatgpt.com sessions. Compare each hour from Oct. 15 onward with the same weekday and hour one week earlier. Allow time for GA4 to finish processing recent data. Also break returning sessions out by landing page, since inline links reportedly lost the tag while Sources panel links kept it.
Huesler adds that most ChatGPT clicks still arrive without a referrer, so he recommends counting the UTM tag instead. In practice, filter on source or campaign values tied to utm_source=chatgpt.com rather than relying on referrer detection. Smith says UTM parameters you set yourself on ChatGPT Ads clicks should be unaffected.
Restored tags help future reporting. Profound notes they will not bring back attribution already lost, and this audit will not either.
Also read: see how to fact-check ai-generated metadata before publishing
Step 7: Annotate and report
Add a note for Oct. 8–15 wherever your team keeps annotations, whether that is a dashboard text box, a shared change log or your reporting tool. Huesler puts the affected window at Oct. 8–15. Keep the wording factual.
Annotation template: "Oct. 8–15: ChatGPT-attributed sessions fell. Analysts report ChatGPT links lost their UTM tag and referrer, which can shift visits into Direct. A fix was reported starting Oct. 15; status on our property: [confirmed/not yet seen]. Interpret this window with caution."
Report template (one short paragraph per heading):
- Known: ChatGPT sessions fell [X]% (Oct. 1–7 vs. Oct. 8–14). Direct rose on [N] of [M] pages that lost ChatGPT sessions. Total sessions changed [Y]%.
- Estimated: After adjusting for control-page drift, about [Z]% of the ChatGPT drop is matched by excess Direct sessions. New-user and mobile shares of Direct [rose/did not rise]. This is an estimate, not a measurement of individual visits.
- Unresolved: The remaining [100-Z]% could be lost visits, noise or untagged visits hidden in Direct. GA4 cannot separate these. The cited sources are early and partly vendor-reported.
Avoid phrases like "recovered traffic" or "ChatGPT visibility fell by X." Neither claim follows from this audit.
What to expect next
The open question is how fast tags return for every account, since OpenAI's release notes do not mention the fix. Rerun the Step 6 check daily for a week, then repeat the full comparison with a clean post-fix window.
If tagged ChatGPT sessions climb back toward the baseline while Direct eases, that supports your "consistent with misattribution" label. If they do not, the unresolved share of your report is the part to keep investigating.
