How to track AI traffic in GA4 without overclaiming

Three of four AI outcomes never look like AI.

What does AI traffic look like in GA4 by default?

Until May 2026, it looked like nothing. An assistant visit landed in Referral next to blog mentions, or in Direct next to bookmarks, and there was no way to see the group.

That changed. Google's Analytics release notes carry a 13 May 2026 entry titled "New AI Assistant traffic measurement", and the Default channel group help page now defines a channel called AI Assistant: "the channel by which users arrive at your site from sources like ChatGPT, Gemini, Deepseek, Copilot, or Grok." GA4 sets medium to ai-assistant and campaign to (ai-assistant) when the referrer matches Google's internal list.

So open Reports → Acquisition → Traffic acquisition, keep the dimension on Session default channel group, and look for the AI Assistant row. If you have never checked, check before you build anything. If you are still finding your way around the interface, our Google Analytics 4 guide covers the report structure, and GA4 setup covers the property basics.

Three catches come with the native channel.

• It is not retroactive. GA4 processes traffic forward only, so last year's assistant visits stay where they landed. Your AI Assistant row starts the day it switched on for your property.
• Google has not published the list. The help page links a spreadsheet of sources and categories. We downloaded it in October 2026 — it contains search, social, shopping and video categories, and no AI assistant category. The matching list is not public.
• It excludes Google's own AI. The same page says AI Assistant "excludes Google's AI Overviews and AI Mode", because Organic Search is defined as including them.

Which AI assistants actually pass a referrer?

Most assistants pass a referrer on desktop web, which is why this works at all. Native mobile apps and some in-app browsers strip it, and a stripped referrer cannot be recovered. That split is the whole reason your measured AI number is a floor rather than a total.

This list is non-exhaustive and will change. New assistants launch, domains move, and apps change their referrer behaviour without announcing it. Treat it as a starting point you re-check each quarter, not a settled standard.

How do you build a custom channel group for AI traffic?

You still want one, for two reasons: it applies to historical data, and it catches assistants Google's list has not reached yet. Google's own documentation confirms the first point — "custom channel groups can be applied to your reports retroactively." That is how you get a before-and-after view the native channel cannot give you.

1. Go to Admin → Data display → Channel groups. You need Editor access at property level.
2. Click Create new channel group. Name it something a client will understand, not "test 1".
3. Add a channel called AI assistants.
4. Set the condition to Session source → matches regex, and paste the hostname pattern from the next section.
5. Add an OR condition: Session medium exactly matches ai-assistant. This makes your group inherit Google's matching too.
6. Drag the AI assistants channel above Organic Search and Referral. Rules evaluate top down. Below them, your AI traffic gets eaten by Referral and you will think the regex is broken.
7. Save, then wait 24–48 hours before judging it. Set it as your primary channel group only once you trust it.

Standard properties allow two custom channel groups with up to 50 channels each, per Google's custom channel groups documentation, accessed October 2026. GA4 360 allows five. Spend one of your two here and leave the other free.

Is there a lighter option than a channel group?

Yes. If you do not want to spend a channel group slot, or you only need to answer the question once, use a comparison or an exploration segment instead. Build it in Explore with the same Session source regex. It takes two minutes, applies to historical data, and changes nothing in anyone else's reports.

That last row deserves a sentence, because it is widely oversold. You can tag links inside content an assistant might quote — a PDF, a listing, a partner page — and those tags survive even when the referrer does not. What you cannot do is tag the citation link ChatGPT or Perplexity generates to your page. Our UTM parameter conventions post covers the naming discipline; just do not expect UTMs to solve AI measurement.

The rule set we use on client properties

First-hand: this is the definition we apply on client GA4 properties, written out so you can copy it. We are not publishing client traffic figures with it — the rules are the useful part.

Custom channel group:   "DH — AI + search"
Channel name:           AI assistants
Position in group:      ABOVE Organic Search and ABOVE Referral

Conditions (match ANY one):
  Session medium    exactly matches   ai-assistant
  Session campaign  exactly matches   (ai-assistant)
  Session source    matches regex     <pattern below>

Regex pattern — paste as ONE line, no spaces, left unanchored:
  chatgpt\.com|chat\.openai\.com|openai\.com|perplexity\.ai
  |gemini\.google\.com|claude\.ai|copilot\.microsoft\.com
  |grok\.com|deepseek\.com|you\.com|poe\.com|mistral\.ai|meta\.ai

Do NOT add: google.com, bing.com, duckduckgo.com
  (these pull AI Overviews and ordinary search into the bucket)

Review cadence: every quarter, against Session source in
  Traffic acquisition — add whatever hostnames actually appeared

Two things we got wrong before settling on this. First, we put the channel below Referral and spent an afternoon debugging a regex that was fine. Second, we included bing.com early on, which quietly mixed Copilot with ordinary Bing search and inflated the number. Both are easy to repeat.

The unanchored pattern is deliberate. It matches sub-hosts and regional variants you have not seen yet, at the cost of the occasional false positive — so verify it against your own Session source values after a week rather than trusting ours.

What can you still not see?

This is the part most guides skip, and it is the part that keeps you honest with a client. Three categories of AI-influenced visit are invisible as AI, permanently or for now. Any AI traffic number you report is a floor, not a total.

• App traffic with no referrer. When a native iOS or Android assistant app hands off a link, the referrer is often gone. Nice Looking Data (10 July 2026) describes the stripping as architectural and reports that ChatGPT's Atlas browser has been seen suppressing referrers on outbound opens. Those sessions sit in Direct and cannot be separated. Clickport's browser test (23 April 2026, updated 18 September 2026) found the one signal that often survives is the URL tag: "In practice, the tag can identify a source when the referrer can't."
• AI Overviews and AI Mode clicks. These are real clicks. Google's Impressions, position and clicks page states that "Clicking a link to an external page in AI Mode counts as a click", and says the same for AI Overviews. They arrive in GA4 as source google, medium organic, channel Organic Search — counted, but not separable from ordinary blue-link clicks. Search Console's generative AI report shows impressions only, with no clicks, CTR, position or query data, so it cannot close the gap either. Our Search Console guide covers what that report does hold.
• Answers that never produce a click. An assistant can summarise your page, name your brand and resolve the question in-chat. Nothing reaches your analytics. That is the zero-click problem, and no GA4 configuration touches it.

What should you report to a client?

Report the direction and the mechanism, not a share of total traffic. AI referral volumes are small for most Indian sites, and the honest framing is that you are watching a trend line start, not reallocating budget on it.

For scale, Semrush's traffic channel mix study — billions of visits across 50,000+ websites and 17 industries, January to December 2025 — found AI traffic grew 66% over the year while still sitting at under 0.15% of total visits. That is global, not India-specific, and it is a year old. We could not find a credible India-only figure for AI referral share, so we do not quote one.

A defensible monthly slide says four things:

• AI assistant sessions this month, the month-on-month change, and one line on which assistants appeared
• What those sessions did — engagement rate and conversions, against your site average
• Which landing pages they hit, because that tells you which content gets cited
• A standing caveat: this is a floor; app traffic and Google's AI surfaces are not included

Pair it with visibility data rather than treating sessions as the whole story. Our LLM visibility guide answers "do we appear in AI answers at all", and citation share answers "how often, compared with competitors". This page answers the narrower question: did anyone click through, and how would you know. You need both halves, and brand mention tracking fills the middle.

What a clean AI baseline gets you next quarter

The point of building this now is not the number. It is that in six months you will have a trend line nobody else on the account has, starting from a definition you can explain.

Push the channel group into Looker Studio using the Session default channel group dimension, and it becomes one row on an existing report rather than a new deck. GA4's Source Group dimension, added 11 June 2026, consolidates platform sources and is worth a look there too. From there you can see which pages earn assistant citations, write more of those, and tie the whole thing back to attribution that already exists.

And when the next assistant launches, you add one hostname to a regex instead of rebuilding your reporting. If you want the content side of this — getting cited in the first place — start with how to get cited by ChatGPT and the wider AI search guides, or the marketing analytics hub for the measurement stack around it.