Why did traffic from ChatGPT disappear from GA4 and how can you measure the impact of chats on sales?

Since October 9, Google Analytics 4 reports have shown up to half as many visits from ChatGPT. The data points to an error in link tagging. However, companies that measure marketing solely by clicks may draw the wrong budget conclusions from such reports.

What disappeared from the reports

Since October 9, website owners have seen a sharp drop in visits attributed to ChatGPT in Google Analytics 4 (GA4, Google’s free tool for analyzing website traffic). Ethan Smith, CEO of Graphite, a company that provides SEO services (search engine optimization), said on LinkedIn that on the sites it monitors, the number of such sessions fell by 22-57 percent. According to his estimate, that is about 45 percent less than before on average. LLM Pulse, a provider of a tool for monitoring AI traffic, measured a drop of about 46 percent between October 9 and 11. Total traffic on those same sites remained roughly stable during that time.

According to Smith and other analysts, the cause is a change in the links. Links in the content of ChatGPT responses stopped carrying the utm_source=chatgpt.com parameter. UTM is an addition to a URL that analytics tools use to identify the source of a visit. The links also did not pass a referrer, meaning the browser information about the page the user came from. Such a visit ends up in GA4 under the Direct channel, as if someone had typed the address manually.

Customers probably did not disappear, they just changed labels

The most interesting data was presented by Cedric Huesler, Director of Product Management at Adobe. He compared the period from October 9 to 13 with the previous week across about 500 sites where ChatGPT generated at least 5 percent of visits. Visits from ChatGPT there fell by about 31,000 per day, while direct visits increased by about 26,000 per day. Search traffic fell by 7 percent, similarly to other sites.

Huesler points out that the new “direct” visitors behave like visitors from ChatGPT. On mobile devices, the share of Direct visitors who were shown a cookie consent banner (for him, a signal of a first visit) rose from 36 to 44 percent. Among users coming from ChatGPT, this indicator is 55 percent. The author himself calls these results early, because they are based on four days of sample data.

The change coincided with the rollout of the GPT-6 model. It appeared in paid plans on October 7, and for Free and Go users a day later. LLM Pulse did not record a drop on October 7. According to Smith, links in the response text lost the tag, but links in the sources panel and shopping links kept it. On sites handled by Graphite, 95 percent of mobile visits from ChatGPT could be identified only by the tag. So the counting of traffic from the app was hit hardest.

Profound links the issue to a new response format. From October 8 to 14, the share of responses with citations that included at least one tagged link fell from about 98 to 66 percent. At the same time, the Intelligent UI format (charts, comparisons, and small tools in the response) grew from 2 to about 35 percent of collected responses. Tomek Rudzki from Peec AI ran the same million queries through ChatGPT twice. After the change, a direct website address appeared in 42 percent of responses, compared with 8-10 percent before. More than half of those addresses (54 percent) led to homepages. These are data about displayed links, not actual clicks.

Has the bug already been fixed? It depends on who you ask

Huesler wrote that on October 14, OpenAI confirmed to him that it was a bug. The fix started rolling out to users on the morning of October 15. By 9:00 a.m. Pacific Time, tagged clicks from ChatGPT had returned across the thousands of domains he tracks to about 95-100 percent of the level from a week earlier. SEO consultant Glenn Gabe replied that he still does not see this on his accounts, both paid and free. Profound is not aware of any official information about the fix, and OpenAI’s ChatGPT change notes for October 1-15 do not mention link tracking. We did not find a public statement from OpenAI on this matter.

These figures should be treated with caution. Graphite, LLM Pulse, Profound, and Peec AI sell services or tools related to SEO and AI search, and LLM Pulse and Profound do not disclose their sample sizes. LLM Pulse admits that it did not notice an increase in Direct traffic because it is much larger than ChatGPT traffic and the change is lost in the background. According to LLM Pulse, aggregated data also does not make it possible to distinguish lost traffic from traffic that only changed labels. None of the available sources settles this conclusively.

What to do with the reports right now

Before someone in the company presents the drop as a loss of customers, it is worth cleaning up the data. Huesler and Smith both advise starting with annotations in the reports.

Further steps follow from analysts’ observations:

• compare ChatGPT and Direct traffic side by side for pages that ChatGPT recommended,
• check the homepage separately, because more than half of the new direct links point there,
• count visits by the UTM tag, not by the referrer, because most clicks from ChatGPT still arrive without it,
• use your own UTM parameters in ChatGPT ads, which according to Smith were not affected by the bug.

A click is only part of the story

This mishap exposes a broader problem. The AI Assistant channel in GA4 has been operating since May and, without a referrer, it cannot recognize a visit from a chat, so the tag was the only thread connecting the conversation with the visit. Providers are moving in different directions. Gemini has been adding UTM parameters to links for the past two weeks. Google announced this week that in AI Overviews and AI Mode (AI-generated responses in Google Search), it is placing links in the response text more often and higher up, but it did not provide the scale of the change. Search Console (Google’s free tool for website owners) assigns the same position to all links within one AI Overview, so simply moving a link higher will not change its reported position.

The author of the “Ask A PPC” series in Search Engine Journal, an employee of Microsoft Advertising who discloses this relationship in the article, sees it similarly. Some purchase decisions are made in conversation with a chat and do not end with a click at all. The author suggests treating ad impressions as a directional signal rather than proof of sales. The foundation should remain accurately measured conversions compared with your own source of truth: CRM (customer relationship management system), store data, and sales data. He also advises separating the causes of data gaps: lack of cookie consent, tagging errors, on-site problems, and differences in attribution models, meaning the rules for assigning sales to channels.

Our conclusion: attribution based on a tag and the last click describes reality less and less well today, and a single supplier-side error can change the picture of a channel from one day to the next. It is not worth basing budget decisions on a weekly chart from one tool. It is better to compare several independent signals: conversions, sales in CRM, brand searches, and direct traffic on pages recommended by chats.