In mid-October, many shops saw in their web analytics how visits from ChatGPT collapsed overnight. Had ChatGPT suddenly become less visible, or was only the tag missing that identifies these visits? Web analytics alone cannot separate the two. Server logs come closer: next to every click, they also show how often ChatGPT fetched the pages live during a conversation. I analyzed both for a large online retailer, hour by hour, plus my own samples in ChatGPT.
−73 %
tagged ChatGPT clicks
per day, Oct 9–12 vs. Oct 1–7; on Oct 13–15 even −79 %
−20 %
ChatGPT-User fetches
same period, only verified OpenAI addresses; Oct 13–15 −23 %
34 → 11
clicks per 100 fetches
Oct 1–7 vs. Oct 9–12; on Oct 13–15 only 9
≈ 55 %
offset by “direct”
More direct mobile entries on product pages, as many as ≈ 55 % of the missing ChatGPT entries; measurable only in part of the shop (GA4, Oct 9–14)
Server logs (UTC) and GA4 of a large online retailer, Oct 1–15, 2026, anonymized. Method and limitations in the next section.
What happened, according to the sources
As of October 16, OpenAI had not commented publicly. What is known comes from measurements by analytics vendors and from the statement of an Adobe executive that OpenAI confirmed the bug to him:
| When | What |
|---|---|
| Oct 7 | GPT-6 launches in the paid plans, on October 8 for Free and Go (Search Engine Journal). |
| Oct 8, approx. 2 p.m. PT | GPT-6 no longer appends utm_source=chatgpt.com, attributable ChatGPT visits drop to about a third within hours (Cedric Huesler, Adobe, according to Relevant Audience). |
| Oct 8–14 | Answers with at least one tagged source link: about 98 % on the 8th, 75 % on the 9th and 67 % on October 14 until 19:00 UTC. In parallel, the new answer format “Intelligent UI” grows (Profound). |
| Oct 14 | According to Huesler, OpenAI confirms to him that it is a bug (Search Engine Journal, Relevant Audience). |
| Oct 15 | The fix rolls out in stages from about 2 a.m. PT. At 9 a.m. PT, tagged clicks are at 95 to 100 % of the previous week across thousands of domains, according to Huesler (Search Engine Journal, Relevant Audience). |
| Oct 14–16 | From October 14, 19:00 UTC, Profound measures a second drop: for about 27 hours, only 5 to 12 % of answers carry a tagged source link. From October 16, 02:00 UTC, the share rises again; around 06:30 UTC it is at about 99 % (Profound). |
PT = Pacific Time. On October 8: 2 p.m. PT = 21:00 UTC = 23:00 German time. In Germany, the outage therefore began in the night from October 8 to 9.
Method and limitations
- Server logs: all requests of the retailer from September 30 (from 18:00 UTC) to October 15, 2026, 17:40 UTC, analyzed by hour in UTC. A click is a page request without a bot identifier, with status 200 and a
utm_sourceof an AI source. These are requests, not people. - Demand: ChatGPT-User fetches, only from the addresses OpenAI publishes for this fetcher. ChatGPT-User fetches a page when someone asks about it in a conversation. That is not a click, but it shows how often ChatGPT reads the shop’s pages live in conversations. Answers from OpenAI’s own search index without a live fetch are not included.
- Baseline: October 1 to 7, averaged per hour of day. Index 100 means: as much as in the previous week at the same time of day. Adjusting for weekdays was not possible because each weekday occurs only once.
- Web analytics: the retailer’s GA4 only until October 14. From October 15, session counting is not comparable with the previous days for a different reason unrelated to ChatGPT. I leave those days out.
- Own samples: three questions on SEO and websites in the ChatGPT interface, location Germany, on October 3, 15 and 16 (two of the three questions the same on all days), plus on October 3, 15 and 16 via OpenAI’s API. I counted the links in the answers with and without the tag.
- Anonymized: I do not name the retailer, nor absolute numbers. Shares, indices and ratios are enough for the questions here.
- Limitations: a single retailer, a short period, only about three hours of data after the return began. The logs have no referrer; untagged ChatGPT clicks cannot be separated from direct requests there.
The break: one hour on the evening of October 8
During the day on October 8, tagged ChatGPT clicks were well below the previous week (index around 65 to 80), and so were fetches (more on that in the next section). Between 21:00 and 22:00 UTC, the series broke off: per 100 fetches, 40 tagged clicks arrived in the hour from 21:00, 18 from 22:00, 10 from 23:00. In the previous week at the same times it was 45, 38 and 35. That matches Huesler’s figure of about 2 p.m. PT, i.e. 21:00 UTC, to within an hour. ChatGPT-User fetches in the same night, relative to the previous week, stayed at the level of the hours before.
Values as a table
| Period | tagged ChatGPT clicks | ChatGPT-User fetches (demand) |
|---|---|---|
| Oct 6, 4:00–10:00 UTC | 98 | 104 |
| Oct 6, 10:00–16:00 UTC | 94 | 81 |
| Oct 6, 16:00–22:00 UTC | 92 | 112 |
| Oct 6, 22:00–4:00 UTC | 82 | 102 |
| Oct 7, 4:00–10:00 UTC | 106 | 96 |
| Oct 7, 10:00–16:00 UTC | 96 | 68 |
| Oct 7, 16:00–22:00 UTC | 108 | 63 |
| Oct 7, 22:00–4:00 UTC | 70 | 55 |
| Oct 8, 4:00–10:00 UTC | 68 | 84 |
| Oct 8, 10:00–16:00 UTC | 82 | 72 |
| Oct 8, 16:00–22:00 UTC | 65 | 65 |
| Oct 8, 22:00–4:00 UTC | 43 | 63 |
| Oct 9, 4:00–10:00 UTC | 25 | 89 |
| Oct 9, 10:00–16:00 UTC | 28 | 71 |
| Oct 9, 16:00–22:00 UTC | 20 | 68 |
| Oct 9, 22:00–4:00 UTC | 31 | 56 |
| Oct 10, 4:00–10:00 UTC | 39 | 77 |
| Oct 10, 10:00–16:00 UTC | 29 | 78 |
| Oct 10, 16:00–22:00 UTC | 22 | 75 |
| Oct 10, 22:00–4:00 UTC | 26 | 51 |
| Oct 11, 4:00–10:00 UTC | 36 | 78 |
| Oct 11, 10:00–16:00 UTC | 37 | 87 |
| Oct 11, 16:00–22:00 UTC | 24 | 107 |
| Oct 11, 22:00–4:00 UTC | 21 | 70 |
| Oct 12, 4:00–10:00 UTC | 29 | 95 |
| Oct 12, 10:00–16:00 UTC | 24 | 80 |
| Oct 12, 16:00–22:00 UTC | 24 | 83 |
| Oct 12, 22:00–4:00 UTC | 20 | 74 |
| Oct 13, 4:00–10:00 UTC | 21 | 85 |
| Oct 13, 10:00–16:00 UTC | 19 | 68 |
| Oct 13, 16:00–22:00 UTC | 19 | 76 |
| Oct 13, 22:00–4:00 UTC | 31 | 75 |
| Oct 14, 4:00–10:00 UTC | 24 | 89 |
| Oct 14, 10:00–16:00 UTC | 19 | 77 |
| Oct 14, 16:00–22:00 UTC | 22 | 81 |
| Oct 14, 22:00–4:00 UTC | 16 | 63 |
| Oct 15, 4:00–10:00 UTC | 20 | 86 |
| Oct 15, 10:00–16:00 UTC | 24 | 72 |
| Oct 15, 16:00–17:00 UTC | 35 | 57 |
Server logs of a large online retailer, anonymized. Clicks: page requests with utm_source=chatgpt.com (occasionally utm_source=openai). Fetches: ChatGPT-User from the published OpenAI addresses. At night the numbers are small, so the curve fluctuates more there.
A second, smaller step follows on October 12 around 14:00 UTC: the click index falls from about 28 to 21, i.e. by another quarter. This suggests that the change was rolled out gradually. It also fits Profound’s measurement, according to which the share of tagged answers kept falling from October 9 to 14. The sharp drop Profound measures on October 14 from 19:00 UTC does not show for this retailer: the click index stays at about 20.
Demand fell much less than clicks
If simply fewer people had come via ChatGPT, live fetches would have had to fall about as sharply as clicks. They did not. Calculated per 24 hours:
| Period | Clicks | Fetches | per 100 fetches |
|---|---|---|---|
| Oct 1–7 | 100 | 100 | 34.0 |
| Oct 8 | 69 | 72 | 32.7 |
| Oct 9–12 | 27 | 80 | 11.4 |
| Oct 13–15 | 21 | 77 | 9.1 |
Clicks = tagged ChatGPT clicks, fetches = ChatGPT-User fetches, both as an index (Oct 1–7 = 100). Last column: tagged clicks per 100 fetches. Days in UTC; October 8 is the transition day, October 15 counts until 17:00 (last full hour).
During the day on October 8, clicks and fetches fell equally, the ratio stayed at about 33 per 100 fetches. So this decline had nothing to do with the tagging. From the night to October 9, only the ratio collapses. Taking the decline in fetches out, 67 to 73 percent of the tagged clicks that would have been expected at a constant ratio are missing.
The smaller part remains open: fetches had already been falling since midday on October 7, by almost a third until the break, and afterwards they were about a fifth below the previous week. That began before the break and on the same day GPT-6 launched in the paid plans, and clicks stayed at the previous week’s level on October 7. Whether this reflects lower demand or a new model that fetches live less often cannot be reliably separated with this data. In favor of the second explanation: according to an analysis by Peec, ChatGPT also draws on its own search index, internally called “Labrador”, and according to Cedric Huesler (Adobe), OpenAI uses ChatGPT-User much less often in the switch to it. This is not confirmed by OpenAI. It changes little for the result above: fewer fetches with the same number of clicks would make the ratio of clicks per 100 fetches rise, not fall. The share of missing tagging of 67 to 73 percent is therefore more of a lower bound.
All devices, browsers and page types at once
A real loss of visibility would hardly hit every group in the same hour; a technical removal of the tag would. That is exactly what the breakdown shows: no group keeps the tag, all of them drop in the same night.
| Attribute | Oct 9–12 | Oct 13–15 |
|---|---|---|
| Mobile | −77 % | −83 % |
| Desktop | −65 % | −70 % |
| Safari on iOS | −78 % | −85 % |
| Chrome on Android | −77 % | −82 % |
| Chrome on desktop | −65 % | −64 % |
| Product pages | −72 % | −79 % |
| Category pages | −87 % | −88 % |
| Links from product cards | −65 % | −72 % |
| Other links | −77 % | −82 % |
Tagged ChatGPT clicks per 24 hours vs. October 1–7, the retailer’s server logs.
Three observations:
- Mobile more than desktop. The logs do not show why. For web analytics it counts double: visits from the ChatGPT app are almost only recognizable by the tag. According to Search Engine Journal, at Graphite 95 percent of mobile ChatGPT visits depended on it alone.
- Links from product cards fell somewhat less. Their share of the remaining tagged clicks rose from 31 to 41 percent. That fits Profound’s observation that some links in expandable product and info panels kept the tag.
- No replacement. No new ChatGPT user agent, no other parameter, no other spelling. The links simply arrived without the parameter.
For comparison, the other AI sources in the same log: tagged clicks from Perplexity did not fall (index 106 and 144). Copilot and Gemini bring too few clicks to say anything.
Where the clicks went: part of them sits under “direct”
Without a tag and usually without a referrer, a ChatGPT visit ends up under “direct” in web analytics. Whether that happened can only be checked indirectly: do direct entries rise exactly where ChatGPT entries are missing? This works most cleanly for mobile visits that start directly on a product page. Nobody types a long product URL by hand.
- In part of the shop: there, direct mobile entries on product pages had been very stable for three weeks. In the outage week, they rose by an amount equal to about 55 percent of the missing ChatGPT entries (GA4, October 9–14 vs. the previous week, per day).
- In the rest of the shop: not measurable. Direct entries there fluctuate more from week to week and even fell in the outage week. A rise the size of the missing ChatGPT entries gets lost in that.
- The referrer dropped too: only a small share of ChatGPT visits brings a chatgpt.com referrer at all. This remainder also fell by about three quarters from October 9 (median 7.8 to 2.1 per 10,000 sessions, visitors with analytics consent). That fits Huesler’s observation that during this time ChatGPT opened links in a way that passed no referrer (Search Engine Land).
For scale: Cedric Huesler (Adobe) saw, on about 500 sites that previously got at least 5 percent of their visits from ChatGPT, about 31,000 fewer ChatGPT visits and about 26,000 more direct visits per day, arithmetically about 84 percent (Search Engine Land). For this retailer it is less, and only part of it is measurable at all. In the server logs, I looked for a substitute signal, for example a ChatGPT-User fetch shortly before an untagged visit to the same product page. The direction is right, but it is not robust. I therefore cannot put an exact number on how many untagged ChatGPT visits there were.
Addendum, October 28: the same URLs
On Reddit, someone asked whether direct entries rose on exactly the product pages that ChatGPT-User fetched during the outage week. I ran that check: about 12,000 product URLs with at least one ChatGPT-User fetch from OpenAI’s published IP ranges between October 9 and 14, against two control groups of the same size made of product URLs with Google organic entries but no observed ChatGPT-User fetch between September 30 and October 14: one drawn at random, one taken from the URLs with the most Google organic entries between September 18 and October 14, i.e. including the outage week. For each group, direct mobile entries in GA4 without the known bot signature, outage week (October 9–14) against the mean of the three previous weeks (September 18–23, September 25–30, October 2–7), each Friday to Wednesday.
| URL group | direct entries, mobile | Google organic, all devices | URLs above their previous-week maximum |
|---|---|---|---|
| Product URLs fetched by ChatGPT-User | −0.5 % | −3.4 % | 12.7 % (12.1–13.3 %) |
| Control: random product URLs without a fetch | +3.5 % | −1.8 % | 6.0 % (5.6–6.5 %) |
| Control: product URLs with the most Google entries, without a fetch | −1.9 % | −4.7 % | 12.0 % (11.4–12.6 %) |
The retailer’s GA4. Direct mobile sessions without the known bot signature; Google organic across all devices as a measure of general demand. Last column: share of URLs whose direct mobile entries in the outage week exceeded each of the three previous weeks; all URLs in the group count, including those without sessions; 95 % Wilson interval.
Across all language versions there is no robust excess rise on the fetched URLs: −0.5 percent against +3.5 percent in the random control and −1.9 percent in the control with heavy Google traffic; depending on the control, the figure sits below or slightly above. In the part of the shop where the 55 percent above were measured, direct mobile entries on the fetched URLs rose by 17 percent; arithmetically that equals about 65 percent of the mobile ChatGPT entries missing there, a comparison of size, not a demonstrated share of relabelled sessions. In the rest of the shop they fell. The share of fetched URLs above their previous-week maximum, 12.7 percent, is clearly above the random control (6.0 percent, one-sided permutation test p < 0.001); against the control with heavy Google traffic the difference is not statistically clear (12.0 percent, p = 0.06). As a by-product, the same dataset reproduces the figure above: all product pages of that part of the shop against the immediately preceding week again give about 55 percent.
My conclusion: in the conspicuous part of the shop, direct mobile entries rise on exactly the fetched product pages as well. That is a descriptive indication that depends on segment and choice of control; it does not demonstrate that individual ChatGPT sessions ended up under “direct”. Limitations: the control groups are matched by count only, neither by language nor by baseline traffic, and they overlap; the top control uses Google entries up to and including the outage week. “No fetch” only means no hit in the available logs between September 30 and October 14; earlier fetches and unobserved hours remain open, and a fetch is no proof that someone clicked afterwards. The final GA4 raw reports carry no data-loss warning, but the bot-signature queries do; the mobile signature deduction is very small and identical in two query variants, yet further lost signature rows cannot be ruled out. Intervals and p-values apply to the URL model, not to causes.
The return: from the afternoon of October 15
In the log, tagged clicks start to rise on October 15 around 14:30 to 15:00 UTC: between 15:00 and 16:30 UTC, about two to three times as many per half hour as on the previous days at the same time, spread over many different addresses, so not a single automated client. In the hours from 15:00 and from 16:00 UTC, however, they are only at 41 and 35 percent of the previous week, and from 17:00 somewhat lower again. The data series ends at 17:40 UTC. Huesler (Adobe) already reported 95 to 100 percent across thousands of domains for 16:00 UTC; Profound saw the full return only on October 16 around 06:30 UTC. For this retailer, the fix had arrived on the afternoon of October 15, but was far from complete. Whether it reached the full level again, my data do not show yet.
My own samples show the same picture from the other side, from ChatGPT’s point of view:
- ChatGPT interface (utm_source=chatgpt.com)
- Oct 3, 14:00, answer text 100 % (67 of 67)
- Oct 15, 07:00, answer text 0 % (0 of 24)
- Oct 15, 07:00, sources 0 % (0 of 32)
- Oct 15, around 14:35, all links 0 % (0 of 82)
- Oct 16, 06:25, answer text 71 % (25 of 35)
- Oct 16, 06:25, sources 72 % (31 of 43)
- Oct 16, 07:40, all links 70 % (71 of 102)
- Via the OpenAI API (utm_source=openai)
- Oct 15 85 % (28 of 33)
- Oct 16 71 % (25 of 35)
“All links” also counts image URLs and placeholders such as “deine-domain.de”, which never carry a tag, and some pages appear once with and once without the parameter in the same answer. The roughly 70 percent of October 16 are therefore not an exact comparison value to October 3; they only show: the tag is back.
Via the API, OpenAI appends a different tag: utm_source=openai. This tag appeared in my API samples before, during and after the outage (October 3, 15 and 16), utm_source=chatgpt.com never did. In the retailer’s log, with a handful of clicks per period, it plays hardly any role, but it does for shops with visitors from apps and tools built on the API.
Strength of evidence at a glance
| Claim | Strength of evidence |
|---|---|
| Timing (Oct 8, 21:00–22:00 UTC) and shape of the step, size of the decline, all groups at once, fetches only about −20 % | established |
| Own samples: Oct 3 with, Oct 15 without, Oct 16 again mostly with the tag | established |
| The main cause is missing tagging, not lower visibility in ChatGPT | very likely |
| Part of the visits ended up under “direct” (in part of the shop about 55 %) | likely |
| On the product URLs fetched by ChatGPT-User in that part of the shop, direct mobile entries rise by 17 % (addendum, Oct 28); no robust excess rise across all language versions, a ChatGPT-driven rise is not demonstrated | descriptive indication |
| The rollout happened in steps (second step on Oct 12) | likely |
| How large the real decline underneath is, why fetches already fell on Oct 7, whether the recovery reached the full level | open |
What does this mean for you?
- Mark the week, don’t repair it. Add an annotation in GA4 and your reports for October 8 to 16, 2026. According to Profound, the fix only fully arrived for many answers early on October 16. Mind your property’s time zone: in German time, the outage begins in the night from October 8 to 9. Monthly and weekly comparisons with this week need a note; you do not need to rebuild channel definitions and attribution models because of it.
- Direct entries on deep pages as a cross-check. Compare direct entries on product or article pages per device with the previous weeks. If they rise where ChatGPT entries are missing, that is a hint of missing tagging, not proof.
- Count
utm_source=openaitoo. Add it to your AI channel group or AI filter, next tochatgpt.com. That way you also capture visits from apps that use OpenAI’s API. - ChatGPT-User fetches as a demand signal. If you have access to server logs, count ChatGPT-User fetches, only from the published OpenAI addresses. They do not depend on any tag. Clicks per 100 fetches are then a good early-warning value: if it flips overnight, look at the tagging first. Because, according to Cedric Huesler (Adobe), OpenAI uses ChatGPT-User less often in the switch to its own index, the baseline number of fetches may decline over time. Compare short periods and recalibrate the value regularly.
- Never rely on a single tag. Put tagged clicks, referrers, fetches, direct entries on deep pages and a fixed sample of your own questions side by side. If only one of them drops, it is usually a measurement issue. If all of them drop, it is more likely visibility.
How to build the sample with fixed questions is explained in the guide Measuring AI visibility (in German).
Sources
- Matt G. Southern: ChatGPT Traffic Drop Appears To Be A Tracking Problem, Search Engine Journal, October 15, 2026 (Huesler/Adobe, Graphite, LLM Pulse, Peec AI, Profound).
- Antonio Fernandez: ChatGPT UTM Bug Confirmed (Secondhand), Fix Rolling Out, Relevant Audience, October 16, 2026 (timeline in Pacific Time).
- Anu Adegbola: ChatGPT appears to drop UTM tracking, making AI referral traffic harder to measure, Search Engine Land, October 15, 2026.
- Barry Schwartz: OpenAI Drops UTM Parameters From ChatGPT Results (GPT-6), Search Engine Roundtable, October 14, 2026, updated October 15, 2026.
- Josh Blyskal: ChatGPT’s missing UTM tags: what happened from October 8 to October 16, Profound, October 14, 2026, updated through October 16, 2026.
- Peec: ChatGPT has its own search index – the Labrador family, analysis of ChatGPT data May to July 2026, published September 2026. Plus Cedric Huesler (Adobe) in a LinkedIn comment on October 16, 2026: OpenAI uses ChatGPT-User much less often in the switch to Labrador.
- OpenAI: Overview of OpenAI Crawlers (ChatGPT-User, IP address lists).
- All sources retrieved on October 16, 2026. As of then, there is no public statement from OpenAI itself on the bug.
