How to Track AI Traffic in GA4: The AI Assistant Channel, Its Blind Spots, and the Pipeline Report

Search for how to track AI traffic in GA4 and every tutorial ends at the same screenshot: Google’s new AI Assistant channel sitting in the traffic acquisition report, problem apparently solved. It is not solved. The native channel is real progress and you should use it - but by Google’s own documentation it excludes Google’s AI Overviews and AI Mode entirely, and by independent testing it scatters traffic from a single AI source across three different channels at once.

This guide covers all three layers: what the default AI Assistant channel actually counts (and a limitations table of what it does not), the custom channel group with a copy-paste regex that catches more, and the AI-influenced traffic no analytics tool can see - which is where pipeline reporting has to take over from channel reporting. Every Google claim below was checked against Google’s own documentation on July 28, 2026.

The GA4 AI Assistant channel: what it counts and what it quietly skips

On May 13, 2026, Google added a dedicated AI channel to GA4’s default channel group. Per the Google Analytics release notes: “Google Analytics now provides a dedicated way to measure and analyze traffic originating from popular AI assistants.” The mechanics, per Google’s channel definitions page: when a session’s referrer matches Google’s list of AI assistants, GA4 sets the medium to ai-assistant, the campaign to (ai-assistant), and the session lands in the AI Assistant channel.

There is no setup step. It appears automatically in your default channel group reports - which is exactly why most tutorials stop here. (Semrush’s coverage noted the rollout was gradual and the channel “may not be visible in all accounts yet” as of its May 2026 write-up.) One naming note: Google’s definitions table calls it “AI Assistant” while the descriptive text on the same page says “AI Assistants.” Your reports show AI Assistant, so that is what we use here.

Now the part the tutorials skip. Checked against Google’s documentation on July 28, 2026:

The third row deserves a second look. Google’s definitions page reads: “AI Assistants 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.” That exclusion is the big one. Clicks from AI Overviews and AI Mode are classified as Organic Search, per the same Google page - and as our AI Overviews traffic data roundup shows, AI Overviews now trigger on the overwhelming majority of tracked B2B technology keywords. In other words: the AI surface most likely to be reshaping your organic line is, by design, not in your AI channel. Google is not hiding this; it is in the definition. But it means the AI Assistant number describes chatbot referrals, not AI-mediated discovery.

The custom channel group that catches the rest

Even within chatbot referrals, the default channel undercounts. Search Engine Journal’s Itamar Blauer tested it in July 2026 and found traffic from chatgpt.com landing in AI Assistant (chatgpt.com / ai-assistant), Referral (chatgpt.com / referral), and Unassigned (chatgpt.com / not set) simultaneously.

The fix is a custom channel group that classifies by source domain instead of trusting the medium. Three documented properties make this worth doing, per Google’s custom channel groups documentation: standard GA4 properties get two custom channel groups beyond the default (360 properties get five), the conditions support a matches-regex operator on Source, and - the clincher - “Custom channel groups can be applied to your reports retroactively.” Google documents no such retroactivity for the default AI Assistant channel, so the custom group is also your only documented route to classifying sessions from before May 2026.

Google’s doc even includes its own “Example: AI assistants” section with a sample regex. Use its existence as permission, not its pattern as a template: several of its alternatives are bare substring matches like .*gpt.* and .*bard.*, which will sweep in any source containing those strings anywhere in its name. A tighter, domain-anchored pattern comes from SEJ’s July 2026 article, where every domain below appears verbatim, drawn from documented referrer strings:

.*(^|[/.:@?&=])(chatgpt\.com|chat-gpt\.org|openai\.com|perplexity|gemini\.google\.com|copilot\.microsoft\.com|edgepilot|edgeservices|claude\.ai|deepseek\.com|grok\.com|you\.com|nimble\.ai|iask\.ai|aitastic\.app|bnngpt\.com|writesonic\.com|copy\.ai)([/.:@?&#=]|$).*

Setup notes that matter:

1. Build it alongside the default group, not instead of it. You get two custom groups on a standard property, so this spends one - but keeping the default intact lets you compare the native AI Assistant count against your custom count. The gap between them is your first, property-specific estimate of the undercounting.
2. Order the new AI channel above Referral. Google’s own example advises exactly this reordering - in a channel group, position decides which matching rule wins.
3. Treat the domain list as a living document. It is SEJ’s list as of July 15, 2026. Prune domains you never see, and when a new assistant shows up in your referral report, add it - retroactivity means your history gets reclassified too, per Google’s doc.
4. Watch for tagged links. PPC Land reported in June 2025 that ChatGPT appends utm_source=chatgpt to links, extending UTM tagging from citation links to its “More” section links - a partial mitigation from OpenAI’s side that helps tagged clicks surface even when referrer handling is messy.

The AI traffic GA4 can never see - and the layer that catches it

Here is the honest ceiling on all of the above. Both the native channel and your custom regex depend on one fragile thing: a referrer (or a UTM tag) surviving the click. Plenty of AI-driven visits arrive without either. Per SEJ’s analysis: “The biggest blind spot is AI traffic with no referrer at all, which lands in Direct.” In-app browsers, mobile apps, and links a buyer copies out of a chat and pastes into a browser all strip the trail, per SEJ’s launch coverage.

You will find confident percentages online for how much AI traffic hides in Direct. We looked for a named, independent study behind any of them on July 28, 2026 and did not find one - the numeric claims in circulation trace back to vendor benchmarks on self-selected samples. So this post states the mechanism and declines to invent the magnitude. AI traffic GA4 undercounting is structural and real; its exact size, on your site, is not knowable from the outside.

And the sessions that do get labeled are only the visible end of the behavior. The measurement is post-click, per Semrush - the assistant conversation that briefed the buyer never appears anywhere. What the labeled traffic shows is telling, though: SE Ranking’s July 2026 study of 101,574 websites found ChatGPT referral traffic grew 36.7% month over month in May 2026 while still totaling just 0.32% of tracked traffic, and 60% of AI-referred visits land on homepages versus 17% for organic search. Those are buyers who arrive already knowing who you are. Ahrefs’ quality study (about 82,000 websites, mid-2025) found AI-referred visitors view fewer pages (4 versus 5.2) and bounce more (67.8% versus 63.7%) but spend slightly longer on site (86 versus 78 seconds) - they came for a specific answer. And Seer Interactive’s single-site case study measured ChatGPT referrals converting at 15.9% versus 1.76% for Google organic - one site, so a signal rather than a benchmark.

Small, fast-growing, pre-briefed, and partly invisible to channel reports. That combination points to two measurement layers that do not depend on referrers at all:

Self-reported attribution. Put “how did you hear about us?” on every form and have sales ask it on every first call. When a buyer types “asked ChatGPT for options,” you have evidence no channel report contains - lossy and fuzzy, but real. Our piece on traced versus modeled attribution frames this layer precisely: it is demand you can see but not trace, and it belongs alongside your traced numbers, not instead of them.

Conversation-level attribution. GA4 classifies how a session arrived. A conversation records what the session became. When the pre-briefed visitor lands on your homepage - referrer stripped, channel label reading Direct - and opens a chat, asks a pricing question, leaves an email, and books a demo, every link in that chain is a recorded event. The origin label was lost in transit; the outcome record is intact, and the outcome is the part revenue reporting actually needs.

That second layer is the lane Marqeable is built for, so judge accordingly. Our AI website chat is grounded in your business info and answers the buyer’s questions in the moment, captures and qualifies the lead, and books the demo - a one-snippet install with bot and spam protection. Revenue attribution then ties dollars to the exact message that started the chain. To be clear about what this does and does not do: it will not tell you which assistant stripped the referrer - nothing can - but it makes the visit’s outcome traceable even when its origin is unlabeled, which is the difference between “Direct went up” and “this conversation became this deal.” We are in private beta with a small early cohort. Get early access.

How to report AI traffic to the board

Now the question that probably sent you here: someone on the board asked how much of the pipeline comes from AI, and the AI Assistant channel row is the only number you have. Resist putting it on slide one. As our board report template for SaaS argues, website traffic is context at best in a board deck - and the AI Assistant number is a traffic metric with documented undercounting on top. Presented as “our AI number,” it will be either dismissed or, worse, believed.

Report AI traffic to the board as three layers instead, in ascending order of trust:

The framing writes itself once the layers are separated. The GA4 trend says AI-mediated discovery is growing, on a measure you openly label as a floor of visibility. The self-reported count says buyers themselves are naming assistants unprompted. And the traced line - the only one that belongs in the headline - says what it closed. That is the same discipline behind traced attribution generally: bring the smaller number you can defend line by line, not the bigger one you have to caveat. A board that watches you label your own metric as understated will trust the two lines above it more, not less.

One preemptive answer worth having ready: if a director asks why the AI channel number looks small against the AI headlines, the honest reply is that the channel measures chatbot referrals only - AI Overviews clicks sit inside Organic Search per Google’s own definitions, referrer-less visits sit inside Direct per SEJ, and nobody has a defensible way to size either. Saying “the true number is higher but unknowable, so we track the floor” is a stronger position than any precise-sounding estimate.

Frequently asked questions

What is the AI Assistant channel in GA4?

It is a default channel group value Google added on May 13, 2026. When a session’s referrer matches Google’s unpublished list of AI assistants, GA4 sets the medium to ai-assistant and the campaign to (ai-assistant), and the session is classified as AI Assistant. Google names ChatGPT, Gemini, Deepseek, Copilot, and Grok as examples, and it excludes Google’s AI Overviews and AI Mode by design - those clicks stay in Organic Search.

Why does ChatGPT traffic show up as Direct or Unassigned in GA4?

Classification depends on the referrer and medium surviving the click. Search Engine Journal’s July 2026 testing found sessions from the single source chatgpt.com split across three GA4 channels at once: AI Assistant, Referral, and Unassigned - the last being sessions with no medium, mainly from in-app browsers that strip referrer data. AI visits that arrive with no referrer at all land in Direct, and as of July 2026 we could not find a named, independent study behind any published percentage for how much traffic that hides.

Does the GA4 AI Assistant channel include Google AI Overviews?

No. Google’s own documentation classifies clicks from AI Overviews and AI Mode as Organic Search and explicitly keeps them out of the AI Assistant channel. Per Search Engine Journal, AI Mode visibility is tracked through Search Console instead. So a major AI-mediated discovery surface is excluded from the AI channel by design.

How should I report AI traffic to the board?

As three layers, in ascending order of trust. The GA4 AI Assistant channel trend is an operating signal - label it as understated and keep it in the appendix, because traffic metrics never lead a board report. Self-reported answers to “how did you hear about us” that name an AI assistant are demand you can see but not trace. Closed revenue with a recorded chain from conversation to meeting to deal is the floor you lead with: revenue you can prove.

The bottom line

The AI Assistant channel is already in your GA4 default channel group if the rollout has reached you, and it is worth reading - as chatbot referrals, nothing more. Build the custom channel group with the domain-anchored regex, because it catches what the default scatters into Referral and Unassigned, and it is the only documented way to classify your pre-May-2026 history. Then accept what neither can do: AI Overviews clicks stay in Organic Search by Google’s design, referrer-stripped visits stay in Direct, and the assistant conversation that pre-briefed your buyer was never measurable in the first place. Measure with GA4, but report with records - a self-reported answer from the buyer, and a conversation-level chain from first question to closed deal. The channel report tells you AI discovery is growing. Only the records tell you what it is worth.

See it live: Marqeable’s AI website chat answers the pre-briefed buyer and captures the lead, the conversations inbox is where your SMS replies land, and attribution ties the dollars back to the exact message.

Marqeable runs your campaigns, answers every visitor, text, and email in seconds, and turns them into booked jobs and meetings - even at 9pm on a Saturday. We’re in private beta with a small early cohort. Get early access