How to Track AI Citations and AI Traffic

For about two years, the honest answer to how to track AI traffic in GA4 was "you cannot, not really". That changed this year, and most guides have not caught up.

Your analytics already has an AI channel

Google added a default channel called AI Assistant, which rolled out gradually and was widely available by mid 2026.

Here is how to find it. Open GA4, go to Reports, then Acquisition, then Traffic acquisition. The first column is already set to Session default channel group. Scroll the list. If you have had any AI referral traffic at all, AI Assistant will be sitting there next to Organic Search and Direct.

Google's own definition is one sentence: "AI Assistant is the channel by which users arrive at your site from sources like ChatGPT, Gemini, Deepseek, Copilot, or Grok. It excludes Google's AI Overviews and AI Mode."

Read that last part twice.

The mechanism is simple. When someone clicks a link inside one of these tools and lands on your site, GA4 checks the referrer against a list it keeps. If it matches, GA4 sets the medium to ai-assistant and the campaign to (ai-assistant). The channel is then just one rule: medium exactly matches ai-assistant.

Build your own AI channel anyway

Google gives five example sources and stops. It does not publish the list, so you do not know whether Perplexity is on it, or Copilot, or the smaller tools your buyers use.

You also cannot see your own past. The medium is stamped when the visit happens, not when you run the report, so sessions from last year do not carry it.

A custom channel group built on Source fixes both problems, because Source was always recorded.

The path is Admin, then Data display, then Channel groups, then Create new channel group. You need Editor access. Add a channel called AI and give it two conditions joined with OR:

• Source matches regex: chatgpt|openai|perplexity|claude\.ai|gemini\.google|copilot|deepseek|grok|meta\.ai|you\.com|notebooklm
• Default channel group exactly matches: AI Assistant

The second condition inherits whatever Google adds to its list. The first catches what Google missed, and it reaches back through your old data.

One limit to know first. A standard property allows two custom channel groups and a 360 property allows five, with fifty channels in each.

Check the raw strings before you trust any of it

Change the dimension in Traffic acquisition from Session default channel group to Session source / medium, set the range to ninety days, and read what is there.

The hostnames that turn up are chatgpt.com, chat.openai.com, perplexity.ai, claude.ai, gemini.google.com, copilot.microsoft.com, deepseek.com, grok.com, meta.ai and you.com. Yours will not include all of them, and may include one you have never heard of.

ChatGPT referral traffic also arrives tagged. ChatGPT has appended a utm_source parameter to cited links for some time, and in June 2025 the SEO consultant Glenn Gabe spotted it extended to the "More" links too. The exact value has been written up differently in different places, so read it off your own report rather than copying it from anyone, including me.

The three things this channel does not catch

Google's own AI surfaces. By Google's definition, clicks from AI Overviews and AI Mode land in Organic Search as google / organic, exactly like a blue link. You cannot separate them in GA4 at all. Search Console's generative AI report is the only place they show up, and it covers Google only.

Visits with no referrer. When someone taps a link inside a phone app, the referrer header often is not sent, and the visit lands in Direct. Nobody has a trustworthy figure for this gap. One agency guide puts it at 35 to 70 percent of AI sessions, but it cites other people's research rather than its own accounts, so treat that as a shrug. Better to run your own check: sort Direct by landing page and look for deep URLs nobody would type by hand. One agency found five ChatGPT-cited pages drove around 40 percent of direct visits inside a single content group.

Copy and paste. Someone reads the answer, copies your link, opens a new tab. Nothing is passed, and that visit is Direct forever.

A citation with no click leaves no trace at all

If ChatGPT names your company in an answer and the person reads it and closes the tab, your analytics records nothing. Not a session. Not a direct visit. Not an impression. There is no event to catch, because nothing ever touched your server.

So how to track AI citations is a different job from tracking clicks, and it needs different tools.

The cheapest is your server logs. OpenAI publishes its crawler names: OAI-SearchBot, which it says is "used to surface websites in search results in ChatGPT's search features", and ChatGPT-User, used for certain user actions. A hit from either means a page of yours was fetched while an answer was built. It does not prove you were cited, or that anyone read it. But it is real, it is free, and it is your data rather than a vendor's.

The scale of the gap shows up in Cloudflare's crawl-to-refer ratio, which divides crawler requests by referral clicks. In August 2025 Cloudflare published ratios of nearly 50,000 to 1 for Anthropic and 887 to 1 for OpenAI. By April 2026 a practitioner reading the same dashboard reported roughly 13,528 to 1 and 1,252 to 1, against about 5 to 1 for Googlebot. Narrowing, and nowhere near closed.

Past that you are into asking the engines your buyers' questions on a schedule and counting how often you appear. That is a separate discipline, and it is what our AI visibility audit does.

How big should your number be? The studies disagree

Two large studies, both paid for by companies selling into this market, both looking at 2025. They do not agree.

Conductor's 2026 AEO/GEO Benchmarks Report covered 13,770 domains and over 3.3 billion sessions from May to September 2025. It puts AI referral traffic at just over 1 percent of all website visits, with ChatGPT responsible for 87.4 percent.

Semrush looked at over 50,000 websites across 17 industries for all of 2025 and puts AI traffic below 0.15 percent of total visits, growing 66 percent across the year while organic grew 2.38 percent.

That is close to a sevenfold difference on the same question, and nobody has reconciled it. The panels and definitions differ, but that is a guess at the cause, not a finding. Do not build a forecast on either number.

Conversion is just as unsettled. StudioHawk analysed 21 months of GA4 data from 100 Australian ecommerce brands and found 340,000 ChatGPT sessions converting at 0.9 percent, while the figures circulating in the industry run from four to fourteen times better than organic. They may be measuring different things, a purchase against a form fill, but that is my guess, not a published explanation. Your own number beats all of them.

What to do this week

1. Look for AI Assistant in Traffic acquisition.
2. Read your raw Session source / medium strings for the last ninety days.
3. Build the custom channel group, to look backwards and cover the gaps.
4. Sort Direct by landing page and check for deep URLs nobody types.
5. Check your server logs for OAI-SearchBot and ChatGPT-User.

Then leave it alone for a month. These numbers are small and jumpy, and checking them weekly teaches you nothing except how to worry. Measurement is half the job. The other half is what you change on the page.

Sources

• Google Analytics Help, "[GA4] Default channel group", including the AI Assistant definition and the ai-assistant medium rule. support.google.com/analytics/answer/9756891
• Google Analytics Help, "[GA4] Custom channel groups": creation path, permitted dimensions, retroactive application, and per-property limits.
• OpenAI developer documentation, "Bots": OAI-SearchBot, ChatGPT-User, GPTBot and OAI-AdsBot descriptions.
• Conductor, 2026 AEO/GEO Benchmarks Report, published 13 November 2025. 13,770 domains, 3.3 billion sessions, May to September 2025 data. Vendor-funded. Reported by Search Engine Land.
• Semrush, "We analyzed billions of web visits: How AI is reshaping traffic channels", 27 April 2026. 50,000+ websites, 17 industries, January to December 2025. Vendor-funded.
• Lawrence Hitches, StudioHawk, analysis of 21 months of GA4 data across 100 Australian ecommerce brands, July 2024 to March 2026, published 5 April 2026. Agency data.
• Cloudflare blog, "A deeper look at AI crawlers: breaking down traffic by purpose and industry", 28 August 2025. Crawl-to-refer ratios by platform.
• Jarred Smith, "What my own analytics say about AI search traffic in 2026", 26 May 2026. Source of the April 2026 crawl-to-refer readings. The author sells a visibility tool.
• MadX Digital, "GA4's AI Assistant Channel: What It Shows and Hides", 29 August 2026. Source of the referrer hostname list and the 35 to 70 percent Direct estimate. Agency.
• Wheelhouse DMG, "AI traffic is already in your analytics", 23 March 2026. Source of the client example on direct traffic to cited pages. Agency.
• PPC Land, "ChatGPT adds UTM parameters to more links for better analytics tracking", 16 June 2025, reporting Glenn Gabe's observation.