If your ChatGPT visits fell sharply in October and your Direct traffic rose, your site may not have lost anything. Several reports say that ChatGPT changed how its links are opened, and that analytics tools started counting those visits as Direct. This post explains what is reported, what the numbers look like, how to check your own data in minutes, and what not to conclude. We could not open the original analysis or the articles, so this is based on search summaries and a chart from Adobe’s analysis that a reader shared with us. Treat figures as reported and sampled.
What changed
For months, links in ChatGPT carried a tracking tag, utm_source=chatgpt.com. Search Engine Land reports that OpenAI added it in June. Analytics tools read that tag and credited ChatGPT as the source. Only a small share of clicks, about 12 percent in Adobe’s chart, also carried a normal referrer.
Around 8 to 10 October, with the rollout of GPT-6, reports say links stopped carrying the tag. Adobe’s analysis, as reported, says the links were opened with a script call that includes the “noopener” and “noreferrer” settings, which remove the referrer. With no tag and no referrer, the visit looks like someone typed the address, so analytics files it as Direct.
What the numbers say (reported)
| Measure | Reported result | Source |
|---|---|---|
| About 500 pages where ChatGPT drove at least 5 percent of visits | ChatGPT down about 31,000 visits a day, Direct up about 26,000 a day, search down about 7 percent, comparing 9 to 13 October with the week before | Adobe data, via Search Engine Land |
| Pages that lost ChatGPT visits | 261 of 387 (two in three) gained Direct visits, against 49 percent the week before | Adobe chart |
| Mobile Direct visitors on ChatGPT-heavy pages | New-to-site share rose from 36 to 44 percent | Adobe chart |
| GA4 ChatGPT referral sessions, 9 to 11 October | Down about 46 percent while total traffic moved about 2 percent | LLM Pulse |
Adobe describes its data as sampled and directional, and it excludes bots. Two figures here come from a social post carrying Adobe’s chart, so check them against Adobe’s own release.
Was it a bug, and is it fixed?
Search Engine Land’s headline calls the drop a bug. Reports say OpenAI confirmed that, and that the tag returned around 15 October, with one report saying 16 October. We found no public OpenAI statement, so the confirmation rests on the analyst’s account. Other write-ups tie the change to GPT-6 and note that OpenAI has not explained the cause itself.
Important: reports say your history is not corrected. Visits recorded in that window stay filed as Direct.
Our view
This is our opinion, not OpenAI’s or Adobe’s. Two lessons. First, a drop in AI referrals is not always a drop in AI visibility. Second, a measurement that depends on one tag is fragile, so keep more than one signal.
A five-minute check
- Mark the dates. Add an annotation in your analytics for 8 to 15 October.
- Put Direct and ChatGPT side by side. Pick landing pages that normally get ChatGPT visits. If ChatGPT fell and Direct rose by a similar amount on the same pages, the pattern fits this issue.
- Check search. If search traffic to those pages held, a content problem is less likely.
- Check the tag now. Look for utm_source=chatgpt.com in recent sessions to confirm it is back.
- Do not rewrite pages because of a drop in that window.
What not to conclude
Do not assume all the missing visits were ChatGPT. Direct includes many sources, and a sample of pages is not your whole site. Also, most ChatGPT clicks still lack a referrer, so the tag remains the main signal and will remain fragile.
What to do
- Filter AI traffic by the UTM tag as well as by referrer. Our analytics setup guide shows how.
- Track Direct on ChatGPT-heavy pages as a second line on your report.
- Watch branded search and run our 20-prompt AI visibility test, which does not depend on tags.
- Read what Adobe’s earlier data said about AI-referred shoppers and keep this episode in mind when you compare periods.
- Use our traffic drop routine before changing any page.
