Measuring AI traffic in GA4: what you see and miss

You see your own site appear as a source in ChatGPT. Nice, but what does it bring in? Until recently Google Analytics 4 could not tell you without building your own rules. Traffic from AI tools landed in the “Referral” bucket, next to blogs and news sites. If you want to know why you show up there at all, read how ChatGPT picks its sources.

That has changed. Since 13 May 2026 GA4 has a dedicated channel for AI assistants. And since June 2026 Search Console has a separate report for generative AI. Two new measurement points within three weeks.

There is a catch: they do not measure the same thing. The GA4 channel explicitly excludes Google’s AI Overviews and AI Mode. The Search Console report covers exactly those two, and shows no clicks. Add the numbers together and you get a figure that means nothing.

This article explains what each one measures, where the gaps are, and how I check what is happening in my own data before drawing a conclusion from it.

What GA4 has done since May 2026

GA4 has a new channel in the default channel group: AI Assistant. Google describes it as follows: “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.”

Under the hood, three things happen when the referrer (the site the visitor came from) is recognised as an AI assistant:

The medium is set to ai-assistant.
The campaign is set to (ai-assistant).
The session lands in the AI Assistant channel.

You do not have to configure anything. Whether Google also applies this assignment retroactively to older sessions is not stated in the documentation. Further down you will find how to check that in your own property.

If you want your own breakdown, for example to track ChatGPT and Perplexity separately, a custom channel group will do it. According to Google those do apply to historical data: “Custom channel groups can be applied to your reports retroactively.” In a standard property, so not 360, you can create two extra groups, with a maximum of 50 channels per group.

Google’s guide includes an example for AI assistants with a ready-made regex. Do not copy it unchanged. Its first part is ^.*ai, which matches every source containing the letters ai. A newsletter link with mail in the source name falls into it too. Spell out the sources you want to track instead.

Clicks from AI Overviews and AI Mode sit somewhere else

This is the part that causes most of the confusion. A click from an AI Overview or from AI Mode does not end up in the AI Assistant channel. It lands in Organic Search. Google says so in the same documentation: Organic Search is the channel by which users arrive via non-ad links in organic search results, “including Google’s AI Overviews and AI Mode”.

That makes sense, because Google is still the one sending the click. The consequence, though, is that GA4 cannot show you which part of your organic traffic came from an AI answer and which part came from the blue links. That split does not exist in the data.

GA4 splits AI traffic across two channels. Only the left side counts as AI Assistant.

Search Console works the same way. Google writes that sites appearing in AI features are included in overall search traffic, and reported in the performance report within the “Web” search type. If you want your pages to land there well, schema markup for AI Overviews helps Google understand your content.

In short: GA4 knows two kinds of AI traffic, and they sit in two different channels. AI assistants are a channel of their own. Google itself stays organic.

What the Search Console report adds, and what it does not

On 3 June 2026 Google announced the Search Generative AI performance reports, with a separate version for Search and for Discover. The announcement now carries a note: “As of August 31, 2026, we’ve rolled out these insights to all websites worldwide.”

What you get:

Item	What is included (report for Search)
Metric	Impressions only. No clicks, no CTR, no position
Features	AI Overviews and AI Mode
Dimensions	Pages, countries, devices and dates
Granularity	Hourly, daily, weekly and monthly
Queries	Not available in this report

Discover has a report and a documentation page of its own. Google notes that the devices dimension is available for Search results, so do not assume it appears in the Discover report as well.

Two details you need before you do anything with it.

First, the way impressions are counted. Google: “For example, if two results from the same site appeared in a generative AI search results feature, they count as a single impression in the chart total.” Two of your pages in the same AI answer is therefore one impression in the total, not two.

Second, this is not extra traffic on top of your regular report. The same data is already in the performance report total. The new report is a separate view, not a separate source. Adding them up double counts.

Search Labs experiments are not included. Google: “Search Console doesn’t include data from experiments in Search Labs, as these experiments are still in active development.” If your report is empty, Google names three possible reasons: too few impressions in those features, your site is excluded from generative AI features, or your property does not have the report yet.

So for clicks from AI Overviews and AI Mode you still depend on the regular performance report, where they sit in one figure together with the rest of your organic clicks. According to the announcement, Google is working on more metrics.

Source Group: the field that does look back

On 11 June 2026 Google added one more thing: the Source Group field. It consolidates loose source values into one value per platform, so facebook, fb and Meta-facebook no longer sit apart in your report. For AI traffic this is the relevant line from the announcement: the field includes “built-in grouping for emerging traffic sources like ChatGPT (OpenAI) and Perplexity”.

And this is in there too: “As Source Group is populated retroactively, you can use dimensions to analyze your historical data to take action and optimize advertising.” So this is the only part Google itself describes as populated retroactively. If you want to compare AI traffic over a longer period, look at this field rather than at the channel.

Four gaps you have to work around

This is where measurement breaks down in practice. I list them in the order in which they have cost me the most head-scratching.

1. You do not know which tools fall into the channel. The channel description names ChatGPT, Gemini, Deepseek, Copilot and Grok, the release notes name ChatGPT, Gemini and Claude, and the custom channel group example names Perplexity as well. Those are examples, not a list. Google does publish a file with the sources and categories the channel grouping uses, but it only contains the categories for search, social, shopping and video. There is no category for AI assistants in it. So which referrers the default channel recognises is not public. Google itself writes that the definitions may evolve as the market changes.

2. Utm parameters override the channel. If the link in an AI answer carries utm parameters, manual tagging wins. Google writes: “In instances where the GCLID or DCLID cannot be used as intended, if any UTM parameter is present on the URL, Google Analytics will derive all cross-channel traffic source dimension values from UTM parameters, exclusively.” A click from an AI answer carries no GCLID, so that condition is met. With only a utm_source and no utm_medium, the medium cannot be set to ai-assistant; Google also notes that missing utm parameters result in (not set) values in your reports. Whether this happens to you is visible in your own data.

3. Part of the traffic has no referrer. The channel works off the referrer. A visitor arriving without one lands in Direct, where the session can no longer be told apart from regular direct visits. Google names among the causes links from offline documents such as PDF and Word, missing utm parameters, redirects that strip those parameters, shortened links and ad blockers. Links from apps are not on that list. Whether that plays a role for you shows only in your own data.

4. Be careful with comparisons across May 2026. The channel only exists from that date, and whether older sessions were reclassified is not documented. Check that yourself first. For trends over a longer history, filter on source, or on the Source Group field, which Google does populate retroactively.

How I check this in my own property

You do not have to take any of the points above on my word. This is the check I run, and it takes about fifteen minutes.

See whether the channel is there. Go to Reports, Acquisition, Traffic acquisition and set the first dimension to Session default channel group. If AI Assistant is in the list, it is active for your property.
Find out what sits underneath it. Add Session source as the second dimension. You now see exactly which hosts Google puts in this channel. Write that list down, because it can change.
Check the tools you expect. In an exploration, filter on session source and look up chatgpt.com, perplexity.ai, copilot.microsoft.com, gemini.google.com and claude.ai separately. For each one, check the medium next to it. Anything that is not ai-assistant falls outside the channel, so you are missing it in your channel report.
Check the utm parameters. If one of those sources shows a campaign name other than (ai-assistant), manual tagging is involved. That explains sessions missing from the channel.
Compare two periods. Put a month before 13 May 2026 next to a month after it and see which channel the same source lands in. That tells you whether historical data was reclassified. If it was not, use the Source Group field or the source itself for your comparison.
Measure on key events, not sessions. Put the channel next to your most important event: a request, a download, a call. For most sites AI traffic is still small in volume. Session counts say little; behaviour says more.
Keep the two sources apart. Impressions in AI features come from Search Console. Clicks and behaviour come from GA4. Put them next to each other in your report, never on top of each other. How I track the impression side systematically is described in measuring citations in ChatGPT.

If you want a third measurement point, server logs or your CDN are the most honest source. If your site sits behind Cloudflare, AI Crawl Control shows which AI crawlers visit. The referral figures inside it, meaning which AI platforms send visitors your way, are only available on a paid plan.

What I do with these numbers

Impressions without clicks are not a measurement error. They are the normal outcome of an answer engine: the user gets the answer inside the conversation and does not click through. Your brand was still seen. That is a different kind of result than a visit, and it should be judged differently. That is exactly why GEO is an addition to SEO and not a replacement.

So with AI traffic I look at two things first. Which pages appear in AI features, from the Search Console report. And what the visitors who do click through actually do, from GA4, measured on key events rather than session counts.

And I draw no conclusion from a few dozen sessions. At small volumes, a difference of one week is noise, not a trend. In that case I wait until there is enough data, or I say that I do not know yet.

In short

Since 13 May 2026 GA4 has an AI Assistant channel. The medium is set to ai-assistant automatically when the referrer is recognised.
That channel excludes AI Overviews and AI Mode. Clicks from those land in Organic Search and cannot be isolated there.
Since June 2026 Search Console has a separate generative AI report, rolled out worldwide as of 31 August 2026. It gives impressions only, no clicks and no queries.
Those impressions are already in the regular performance report. Do not add them up.
Which referrers Google recognises as AI assistants is not public. Utm parameters, missing referrers and the channel’s start date make the measurement less complete still.
The Source Group field, from 11 June 2026, groups sources per platform and is populated retroactively according to Google. Use it for comparisons over a longer period.
Check in your own property which sources fall into the channel, and judge AI traffic on key events instead of sessions.

Frequently asked questions

Does traffic from AI Overviews land in the GA4 AI Assistant channel?

No. Google explicitly excludes AI Overviews and AI Mode from that channel. Those clicks land in Organic Search, together with regular search results. GA4 cannot separate the two.

Is Perplexity part of the AI Assistant channel?

Not in the channel description itself, which names ChatGPT, Gemini, Deepseek, Copilot and Grok. Google does name Perplexity in the custom channel group example, and in the release notes of 11 June 2026 under the new Source Group field. A list of the referrers the default channel recognises is not published. So check it in your own data: filter on session source and look at the medium next to it.

Can I see which queries led to an AI Overview featuring my site?

No. The generative AI report in Search Console has no query dimension. You see pages, countries, devices and dates. For queries you still depend on the regular performance report, where AI features and regular results sit together.

Why is there no data in my generative AI report?

Google names three possible reasons. Your site has too few impressions in those features, you have excluded your site from generative AI features, or your property does not have the report yet because the rollout is still running. That last reason is still in the documentation, next to the note that the report has been rolled out worldwide since 31 August 2026.

Do I still need a custom channel group for AI traffic?

Not for a headline figure; the default channel covers that. You do need one if you want to track tools separately, include sources Google does not recognise, or look back at data from before May 2026. According to Google, custom channel groups do apply retroactively to your historical data. Spell out the sources in your regex rather than copying Google’s example regex.

Sources

All sources consulted on 18 September 2026.

Want to know whether your site appears in AI answers? In a free GEO scan I look at your visibility in AI features, your measurement setup and the technical basis underneath it. You get the findings with the source next to each one, so you can verify them yourself.