Clicks from ChatGPT often land in Yandex Metrica as visits with no source. Metrica checks for UTM tags in the first pageview; if none are found, it reads the referrer. When both are absent, the visit goes to direct. Tenpoint Labs put the hidden AI share of direct traffic at 6.3% at most.
Direct includes clicks from mobile apps with WebView, messengers without tags, ChatGPT and Yandex Alice chat. The landing page itself can strip the referrer via referrer-policy rules.
What AI tools send
ChatGPT adds utm_source=chatgpt.com to some links. Seer Interactive found that even links in the sources block do not guarantee the tag. Tenpoint Labs notes ChatGPT apps may drop the referrer.
Alice AI answers in search arrive with referrer yandex.ru/search, same as organic clicks. Alice chat goes to direct.
Gemini appears in the GA4 AI Assistant channel list, determined by referrer. GigaChat, Perplexity and DeepSeek lack confirmation for web or app.
Metrica segment
AI traffic in Metrica needs a segment with two OR conditions: referrer domain and utm_source value. Metrica uses RE2 syntax for regex and treats tag values as case-sensitive, so the utm condition requires the (?i) flag.
Create a visit segment with two conditions: one for referrer domain, one for utm_source value. Set both to match by regular expression.
Domain yandex.ru stays out: Alice search answers share the same referrer as organic, so the segment would pull all search traffic. GigaChat is also excluded until its referrer is confirmed.
Custom channel group in GA4
GA4 has a built-in AI Assistant channel for ChatGPT, Gemini, DeepSeek, Copilot and Grok. Google assigns visits to this channel by referrer: if the referrer matches the AI list, medium becomes ai-assistant. Visits without a referrer do not qualify.
Google defines Direct as saved links or typed URLs, but AI visits with no referrer end up there.
Estimating the hidden share in direct
Tenpoint Labs checked 489,920 direct sessions across 10 sites and found none with tags from ChatGPT, Perplexity, Gemini, Copilot or Claude. Tenpoint assumed all engaged direct visits to deep pages came from AI and arrived at 6.3% of direct traffic on 6 sites. Ramp cites 30–40%.
The estimate needs a segment: entry to deep pages, excluding sections with paid or search traffic, only engaged visits — no bounce and either two pageviews or 30 seconds on site.
Without the engagement threshold the segment collects noise. On 6 of 8 Tenpoint sites, direct visits to deep pages engaged far less than AI referrals and organic. The authors concluded such visits look more like bots and link previews.
The share: segment visits divided by all direct visits in the same period. A result well above 6.3% warrants a check for bots and previews.
Where the count breaks
Alice AI in search arrives with yandex.ru/search and blends into organic. Alice chat goes to direct.
A spike in direct visits after a publication cited by AI tools is correlation: the spike could come from a private chat repost or email. Tenpoint did not confirm that direct spikes link to AI clicks.
Direct visits to deep pages at Tenpoint came mostly from desktop, same as AI referrals, though ChatGPT apps may drop the referrer.
Tenpoint's one weak signal, weekly alignment between engaged direct visits to the homepage and AI referrals (r = 0.45 on one site), held at around five AI visits per week and did not survive correction.
Standard source reports undercount AI traffic: they catch only visits where the AI tool sent a referrer or tag. Reliable counts combine two parts: confirmed traffic visible via referrer and utm, and an estimate drawn from direct visits by session traits. The estimate should come with clear assumptions: the two-pageview or 30-second threshold, excluded sections, bots and link previews the segment does not filter out.
