How to find ChatGPT traffic in Metrica and GA4: up to 6,3% may be hiding in direct visits

You open Metrica and see plenty of direct visits, but no clear indication of where they came from. Some may be people who clicked a link in a ChatGPT answer. The standard traffic sources report, however, will not show that. Here’s how to build your own segment and avoid misleading yourself with the numbers.

Why traffic from AI services gets lost in direct visits

Metrica identifies the source using tags and the referrer. It checks for a tag only on the first pageview of a visit. If there is no tag, it checks the referrer: the address of the page the person came from, which their browser passes in the Referer header. If neither a tag nor a referrer is present, the visit goes into direct traffic. This is known as dark traffic: the visit exists, but its source is missing.

Clicks from mobile apps with WebView, untagged links in messengers, ChatGPT and Alice chat can all end up in direct traffic. The source page can also strip the referrer through referrer-policy.

Confirmed evidence is limited. Only some ChatGPT links carry utm_source=chatgpt.com, and even a link in the sources block does not guarantee it, according to Seer Interactive. An Alice AI answer in search sends yandex.ru/search as the referrer, just like an ordinary click on a search result. There is no confirmation for GigaChat, Perplexity or DeepSeek, either on the web or in their apps.

The exact portion: a segment based on referrer and tag

In Metrica, create a visit segment with two conditions joined by OR. The first matches the referrer domain against a regular expression. The second matches the utm_source value. Regexes use RE2 syntax, and Metrica treats tag values as case-sensitive, so the tag condition needs the (?i) flag.

Referrer domain: ^(.*\.)?(chatgpt\.com|chat\.openai\.com|perplexity\.ai|gemini\.google\.com|copilot\.microsoft\.com|claude\.ai|deepseek\.com|grok\.com)$

utm_source: (?i)^(chatgpt(\.com)?|perplexity(\.ai)?|gemini|copilot|deepseek|claude|grok)$

Of these tag values, only chatgpt.com is confirmed. The other variants are plausible and should be checked against your Metrica counter’s tag report. Do not add yandex.ru: the segment would capture all organic traffic from it. GigaChat is not included yet either.

GA4 has a built-in AI Assistant channel for ChatGPT, Gemini, DeepSeek, Copilot and Grok. AI Overviews and AI Mode are not included. GA4 assigns visits to the channel based on the referrer, so visits without one remain in Direct. For services outside the list, create a custom channel group using the same regex for session source and place it above Referrals.

Estimating the hidden share: treat it strictly as an estimate

The Tenpoint Labs authors examined 489 920 direct sessions across 10 sites and found none tagged as ChatGPT, Perplexity, Gemini, Copilot or Claude. That leaves visits without a tag or referrer. The researchers made the most generous assumption: every engaged direct visit to a deep page comes from AI. Even then, the result was at most 6,3% of direct visits across the 6 sites with enough AI visits to calculate it. Ramp, which the authors cite, puts the figure at 30–40%.

The author of this article proposes a segment based on the same assumption: a direct visit that starts somewhere other than the homepage or advertising and search sections (blog, promo, landing; replace this list with your own), does not bounce, and includes two pageviews or 30 seconds on the site. These thresholds are the author’s choices and should be challenged. The regex applies to the full URL, including the protocol and domain, so it starts with https?://.

An engagement threshold is necessary; otherwise, the segment will fill with junk. On 6 of Tenpoint’s 8 sites, direct visits to deep pages had lower engagement than AI referrals and organic traffic. The authors concluded that these visits looked more like bots and link previews. The share is straightforward to calculate: divide visits in the segment by all direct visits over the period. If the result is noticeably above 6,3%, look for bots first.

Where the count misleads

• Alice AI: answers in search blend into organic traffic, while Alice chat ends up in direct traffic and appears only in the estimate.
• Spikes in direct visits after publishing something AI services cite: this is correlation; a repost in a private chat or a newsletter could have caused it. The study did not confirm a connection.
• Device: Tenpoint’s direct visits came mostly from desktop, as did AI referrals, although ChatGPT apps can lose the referrer.
• Small numbers: the only weak signal (r = 0,45 on one site) rested on roughly five AI visits per week and did not survive correction.

What to show in the report

You cannot present a single number. The report needs two rows: the exact portion identified through referrer and UTM, and an estimate—the number of direct visits that match the segment’s criteria. Together, they give a range: from the exact portion at the lower end to the exact portion plus the estimate at the upper end. Record the assumptions beside the figure: thresholds, excluded sections, and bots and link previews the segment does not filter out. That way, you will neither overestimate the impact of AI services nor miss it entirely.