A sudden drop in the ChatGPT row does not tell you, by itself, that fewer people reached your website. Some may have arrived without a usable source label. It also does not tell you that everything is fine.
Researchers reported missing ChatGPT link tags during October 8–16, 2026, with different samples showing their return at different times. If your report changed around those dates, document the possible attribution gap before judging the channel. A tagging problem and a real decline can happen together.
If you only receive occasional identifiable ChatGPT visits, start with the reporting note. You can record what is unknown without a baseline or a change to GA4.
Checked October 16, 2026. This article reviews public reporting and proposes a way to read your own reports. We have not reproduced the incident or measured a MAXUOD or client traffic recovery.
The dates do not tell the whole story
The useful question is not simply “was it fixed on Thursday?” The sources measured different things.
Profound's updated analysis followed the share of collected answers with cited links that contained at least one tagged source link. It reports a decline beginning October 8 and coverage returning to approximately 99% at 06:30 UTC on October 16. That is answer-level tag coverage, not the percentage of visitors recovered. Some links reached through additional panels retained tags during the gap.
Adobe's Cedric Huesler reported that OpenAI had privately told him the missing tags were a bug. His follow-up described tagged-click recovery on October 15 in his telemetry. That is his account of a private conversation, not a public OpenAI incident statement. Other practitioners in the same discussion had not yet seen the change.
LovedByAI's October 16 update separately reports a return of tagged arrivals from October 15 in its site logs. Its main comparison cohort contained 184 sites that already averaged at least one identifiable ChatGPT arrival a day before the break. That selection is not a minimum traffic requirement for your business, nor a representative sample of every small website.
These observations support a reporting caveat. They do not establish one universal outage window, unchanged visitor demand, or a date when every account and link type behaved the same way. Treat October 8–16 as the reported context, then record what your own evidence can support.
A missing label is different from a missing visit
A UTM parameter is a label in a link. The relevant example here is utm_source=chatgpt.com. A referrer is a separate signal that can identify the page a visitor came from. If a link loses its UTM but still passes a useful referrer, the origin may remain identifiable.
When usable source information is missing, a visit may appear under Direct. Google's explanation of Direct traffic also lists other causes, including direct navigation, offline documents, missing campaign information and tracking interference. Direct is not a hidden ChatGPT total.
A custom channel group can organise source information you actually collected. It cannot retrieve a source that never arrived. Nor can a recovered tag today repair an unidentified visit from last week. Keep the original data; do not move an estimated slice of Direct into an AI bucket.
For the broader distinction between seeing a brand in an answer, receiving a visit and receiving a useful enquiry, use our AI visibility baseline method. This incident concerns identifiable referrals, not proof that your content stopped being cited.
Leave a note your future report can explain
Start with a dated note outside GA4 if that is all your access allows. You do not need to change a property to record uncertainty.
Read the property's existing reporting timezone first. Google documents that the selected timezone determines the dates assigned to visits. Do not change that setting to make the incident line up. For example, Profound's 06:30 UTC observation on October 16 is 03:30 on that date for a property using Halifax time. It is still a source observation, not your site's recovery timestamp.
If you have permission, GA4 supports annotations on reports with line graphs. Right-click a chart point, choose “Add annotation”, and enter a title, description and date or range. Google requires Analyst access or higher to create one. A Viewer can keep the same information in the reporting document instead.
Suggested title: Possible ChatGPT attribution gap
Suggested description: Oct 8–16: reported link-tag loss; site impact unconfirmed. Compare consistent landing-page data before interpreting the trend.
Report and metric: [report name, Sessions and Session source / medium, landing-page definition and filters]
Property timezone: [existing reporting timezone]
Known concurrent changes: [website, consent, tracking or campaign changes; unknown if not checked]
Missing evidence: [baseline, source information or other gaps]
Next action and owner: [specific check, responsible person and review date if useful]
Those fields fit Google's 60-character title and 150-character description limits. Use the broader reported dates provisionally; narrow the note only when your own evidence justifies it. The remaining fields belong in your reporting document, not in GA4's short description.
If you have no usable baseline, or only occasional identifiable arrivals, stop at that record: insufficient evidence to tell whether actual visits changed. That is not reassurance. It means this report cannot answer the question yet.
Compare the same pages without guessing their source
Where you have enough comparable data to inspect, use Sessions with Session source / medium. Do not switch halfway to First user source / medium or compare a source row with an entire channel group. Google's Traffic acquisition guidance distinguishes the session-level question from first-user acquisition.
Keep the same landing-page definition and filters throughout. Look at pages that previously received identifiable ChatGPT sessions, comparing complete matched weekdays with comparable earlier weeks where available. Keep countries, devices, consent conditions and release changes visible. Use absolute counts as well as percentages; a change of a few sessions can look dramatic.
Direct is optional supporting context. A rise on the same pages is consistent with missing labels, but email, bookmarks and other traffic can produce it too. No visible rise does not rule out a gap. LLM Pulse's original analysis explicitly could not separate lost visits from lost labels in its aggregates, and did not find a clear Direct increase. Its result is a useful warning against treating one offset as proof.
Check known changes before reaching for a platform explanation: a new consent banner, a redirect that drops parameters, a tag release, an email campaign or a broken landing page may need its own investigation. If Google clicks and GA4 sessions also disagree, our GSC and GA4 comparison guide covers that separate counting and collection problem.
Branded queries in Search Console can add context about Google search demand. They cannot identify unlabelled ChatGPT visits. Google's data notes also explain that daily Search Console reporting uses Pacific Time outside the 24-hour view and omits some queries for privacy. Stable branded searches are not proof that ChatGPT clicks stayed steady.
Flat totals can hide a real problem
Illustrative numbers only, not client data: imagine this report for two otherwise comparable periods.
| Observed metric | Before | After |
|---|---|---|
| ChatGPT-labelled sessions | 20 | 5 |
| All-source sessions | 1,000 | 1,000 |
| Qualified enquiries, all sources | 10 | 10 |
One possible explanation is that people kept arriving from ChatGPT but 15 sessions lost their labels. Another is that 15 ChatGPT sessions genuinely disappeared while other sources supplied 15 more. The totals above cannot choose between those explanations. Flat site traffic and enquiries do not exclude losing a small channel.
Now change just the last row: enquiries fall from 10 to 6. Missing source labels could still be present alongside a separate commercial problem. A broken contact form, a different visitor mix or ordinary variation may need checking. The decline does not disprove attribution loss; attribution loss does not explain it away.
What to check after tags return
A tagged link shows that source information is present in that instance. It does not prove your tag collected it correctly, that every ChatGPT surface behaves alike, or that your channel's previous volume returned. Keep link behaviour, analytics collection and actual business activity as separate checks. Respect consent; do not disable privacy controls to make the numbers look complete.
If further comparison is worthwhile, choose a review date appropriate to the complete, comparable data you can collect. Record “consistent with the earlier baseline”, “below baseline, cause unresolved”, or “inconclusive”. None identifies the cause. Do not keep excusing later declines as this October incident without fresh evidence, and do not wait for a reporting review when orders or qualified enquiries are deteriorating.
At MAXUOD, the useful handoff would be the unchanged source report, a dated incident note, the definitions used, known concurrent changes and the next check with an owner. Our weekly service-progress reporting should show both completed work and the questions the evidence still cannot answer.
Related reading and sources
Read next on MAXUOD
External references
Need a clearer measurement handoff?
Start with your public website and a short description of the reporting concern. We can review what to check next, not invent missing attribution. Do not send passwords, personal identifiers or raw visitor records.
