ChatGPT and analytics
ChatGPT visits went missing for a week. Your visibility may not have.
If your ChatGPT line in GA4 dropped around 9 October, you were not alone. For about a week ChatGPT's links reached websites with no tag and no referrer, so analytics tools filed the visits as Direct. The reports point to people still clicking; the label went missing. That week is a good moment to look at what the ChatGPT line ever measured. It counts labelled clicks on links, and in our own answers an AI engine naming a business and linking to its website were two quite different things.
Key takeaways
- 8 to 15 October is a gap in your data, not a verdict. ChatGPT links carried neither a utm_source tag nor a referrer, so the visits landed in Direct. The tag returned on 15 October, and a retest on 16 October found the referrer still missing.
- Count ChatGPT by the tag. With no referrer, utm_source=chatgpt.com is what identifies a ChatGPT visit now.
- Named is not linked. In our data the 6 most-named tools had their own site cited in 12% of the answers that named them, so most mentions could never show up as a referral (1 to 25 September 2026).
- ChatGPT's links move anyway. Asked the same question at the next run, ChatGPT cited 27% of the same sites again, against 55% for Perplexity. Referral counts swing even when tracking works.
What happened, with dates
OpenAI has not published an account of the week; its own pages refused our reads, so what follows comes from three teams that measured it.
- The links lost their tag and referrer. Tony Castillo, in a test he published on 14 October and updated on 16 October, tested 47 links across nine ChatGPT answers on 14 October, from shopping buttons to ordinary citations, and found that every one opened without a referrer or a tracking tag. He dates the gap from 8 to 15 October. He also reports that OpenAI told Cedric Huesler at Adobe the missing tag was a bug, and that a fix began reaching users on 15 October.
- The ChatGPT line fell by about half. LLM Pulse, an AI visibility vendor (Esteve Castells, 13 October, last updated 21 October), found that ChatGPT referral traffic in GA4 dropped about 46% from 9 to 11 October against the same days in the two previous weeks, across the sites it monitors. It gives no site count. More than eight in ten of those sites lost over 30% of their ChatGPT traffic, while their total traffic stayed roughly flat.
- Other properties showed the same shape. SearchIntel (Paul Byrne, 23 October) read the GA4 properties it had access to: ChatGPT sessions from 9 to 13 October ran roughly two-thirds below the daily average of the three weeks before, and total sessions did not fall in step. It does not say how many properties.
- The tag came back, the referrer did not. When he tested again on 16 October, every link in a ChatGPT shopping answer carried utm_source=chatgpt.com, and none passed a referrer.
Total traffic holding level while the ChatGPT line fell is what you would expect if people kept clicking and the visits were simply filed elsewhere. His store data adds a caution: lost clicks to support and how-to pages reappeared as Direct, but on that store lost clicks to product pages did not reappear for US visitors, and only about half did elsewhere. Mostly a labelling problem, then, but not only that. And the missing week cannot be recovered after the fact.
What the ChatGPT line measures, and what it misses
A visit has to get through four steps before it appears in your analytics as ChatGPT. The October gap broke only the last one. The first two were never visible in analytics at all:
| Step | What happens | Who can see it |
|---|---|---|
| Named | The answer mentions your business | Only someone who reads the answer |
| Linked | The answer links to your website | Only someone who reads the answer |
| Clicked | The reader follows that link | Your server, as a visit |
| Labelled | The visit arrives with a tag or a referrer | Your analytics, as ChatGPT |
Named is not linked
We put 20 buyer questions about AI visibility software to five AI engines four times from 1 to 25 September 2026, and kept every answer with the sites it cited. The 6 tools named most often were named in 726 answers between them. Their own website was cited in 89 of those answers, or 12%. The most named, Semrush, was named in 160 answers and had its own site cited in 34.
For a business that means most of the times an assistant recommends you, there is no link to your site for anyone to click. Those mentions can still win you a customer who types your name or address in later, which lands in Direct or branded search, never in the ChatGPT line. That matters: in a Talker Research survey commissioned by We are Talker, of 2,000 Americans who had used an AI tool for a personal purpose in the previous three months (online, 21 to 30 September 2026), 64% said they had bought from a brand they had never heard of until an AI tool mentioned it.
Where ChatGPT's links go
ChatGPT cited sources in 80 of its 80 answers: 613 citations to 330 different sites, and 441 different page addresses. Only 64 of those citations, 10%, went to a software vendor's own site. The rest went to reddit.com (cited in 28 answers), techradar.com (26) and other review, discussion and comparison pages. When ChatGPT recommends you, the click often goes to the page that recommended you, not to you. Which websites ChatGPT cites has the full lists.
The links change from run to run
Asked the same question at the next run, ChatGPT cited 27% of the same sites again (60 pairs of answers), against 55% for Perplexity. The tools it named held better: 53% were named again. So a ChatGPT referral count can rise or fall from one week to the next because the links moved, while the recommendation stayed. Checking every week covers how much each engine moves.
One limit to be plain about: our newest run was on 4 October 2026, before the gap, so we did not measure what ChatGPT said during that week. We cannot tell you your visibility held steady. We can tell you the referral line could not have told you either.
What to do now
- Annotate 8 to 15 October. Mark the week in GA4 and in any report that charts AI traffic, so nobody reads the dip later as ChatGPT dropping you.
- Identify ChatGPT by the tag. With the referrer still missing, filter on utm_source=chatgpt.com. A channel rule that looks only at the chatgpt.com referrer will undercount.
- Look at Direct page by page. Compare Direct visits to the pages ChatGPT usually sends people to for 8 to 15 October against the week before. A rise there is your missing ChatGPT traffic.
- Ask new customers where they heard of you. A plain question on your contact or checkout form catches the mentions that never came with a link, and it survives the next tracking change.
- Measure the answer, not only the click. Put the same buyer questions to ChatGPT on a schedule and note whether you are named and which sites are cited. That is the measurement the October gap did not touch. Free AI visibility checkers are a way to start without paying.
Where Greater Than Services fits
Our own product, so read this as the vendor's description, not part of the measurement above.
Greater Than Services is a workspace per brand that puts your buyer questions to five AI engines every week, records which answers named you and what they cited instead, shows your Google rank beside each answer, and drafts the fixes.
- If you need to know whether ChatGPT still names you, whatever your analytics says, it puts your buyers' questions to ChatGPT every week and records whether each answer named you, with the full text kept.
- If you need to see which sites ChatGPT linked instead of yours, it records the sites each answer cited, so you can see the links that went to a directory, a review site or a rival.
- If you need more than one assistant, the same questions go to Claude, Gemini and Perplexity too on the larger plans.
It reads the answers, not your analytics. It does not count clicks or recover the visits filed as Direct.
Plans start at $49 a month, billed yearly, for one brand on ChatGPT and Google AI Overviews; Growth and Agency cover all five engines. The three-day trial is on one brand and needs a card.
Common questions
Why is my ChatGPT traffic showing as Direct?
Analytics tools know a visit came from ChatGPT in one of two ways: a referrer of chatgpt.com, or a utm_source=chatgpt.com tag on the link. A test by Tony Castillo found that from 8 to 15 October 2026 every ChatGPT link it tested opened with neither, so tools such as GA4 filed those visits as Direct, the same bucket as a typed address or a bookmark.
Is ChatGPT traffic tracking back to normal?
Partly. The same test reported that the utm_source=chatgpt.com tag was back from 15 October, and in its test on 16 October every link in a ChatGPT shopping answer carried the tag while none passed a referrer. So filter ChatGPT visits by the tag, and check your own line from 15 October onwards.
Did ChatGPT stop recommending my business that week?
Nothing in the reports says so. LLM Pulse and SearchIntel both saw ChatGPT sessions fall while total traffic on the same sites held roughly level, which is what relabelled visits look like. Whether ChatGPT names you is a different measurement: you only see it by asking ChatGPT your buyers' questions and reading the answers.
Does ChatGPT referral traffic measure AI visibility?
Not on its own. An answer can name you without linking to you: in our data, the 6 tools named most often had their own website cited in 12% of the answers that named them. Referral traffic counts only the clicks on links, and only when the click arrives labelled.
Sources
All four pages were read on 26 October 2026. OpenAI's help and news pages refused automated reads that day, so no OpenAI statement is quoted directly.
- Tony Castillo, a test of outbound links and shop buttons, published 14 October 2026, updated 16 October 2026
- LLM Pulse, Esteve Castells, ChatGPT Referral Traffic Fell by Almost Half on October 9, 2026, 13 October 2026, updated 21 October 2026
- SearchIntel, Paul Byrne, A Week Without ChatGPT Referrals, 23 October 2026
- Talker Research, Survey finds AI tools drive 64% to buy new brands, 22 October 2026
The measurements are our own, from the answers behind our AI visibility tools ranking, which sets out the questions, the engines and the dates and publishes the data under CC BY. If a figure here does not match its source, tell us and we will correct it.