How to Track AI Referral Traffic in GA4 (2026 Setup + the Undercount Fix)

Start with GA4's native AI Assistant channel

Before you touch a regex, check what GA4 already gives you. In May 2026 Google shipped an AI Assistant channel inside the Default Channel Group. It requires zero configuration. Any GA4 property has it now.

The mechanism is simple. When a session arrives with a referrer that matches Google's recognized list of AI assistants, GA4 sets the session medium to ai-assistant and the campaign to (ai-assistant), then groups it under the AI Assistant channel in your standard reports. You'll see it in Reports → Acquisition → Traffic acquisition, right alongside Organic Search and Referral.

Per Google's Default channel group documentation, the recognized sources are ChatGPT, Gemini, Deepseek, Copilot, and Grok.

For those five platforms, the clicks that carry a referrer are now labeled for you with no maintenance. Everything else is untouched. Perplexity clicks still fall into Referral. Meta AI, Poe, You.com, and the long tail of smaller assistants stay unlabeled. The channel also excludes Google's own AI Overviews and AI Mode, which count as Organic Search. That gap is what the next step closes.

Add a custom channel group for everything the native channel misses

The regex hasn't gone away. Its job just narrowed. Instead of being your entire AI-tracking setup, it's now the net that catches every assistant GA4 doesn't label natively.

A custom channel group sits on top of your reporting without altering your raw data, so you can add it, reorder it, and change it later with no risk. Setup:

1. Go to Admin → Data display → Channel groups.
2. Click Create new channel group. Name it something like AI + Assistants.
3. Click Add new channel. Name it AI Assistants (manual).
4. Set the condition to Source matches regex, and paste the regex below.
5. Drag this channel above Referral and Organic Search in the channel order. GA4 assigns each session to the first channel it matches, so if AI Assistants sits below Referral, Perplexity traffic gets claimed by Referral first and you'll see nothing.
6. Save. The group applies to historical data in most reports, so you can look back immediately.

chatgpt\.com|chat\.openai\.com|openai\.com|claude\.ai|anthropic\.com|gemini\.google\.com|bard\.google\.com|copilot\.microsoft\.com|copilot\.com|perplexity\.ai|meta\.ai|mistral\.ai|chat\.mistral\.ai|deepseek\.com|character\.ai|grok\.com|grok\.x\.com|x\.ai|you\.com|phind\.com|copy\.ai|writesonic\.com|poe\.com|quillbot\.com|jasper\.ai|cohere\.ai|huggingface\.co|pi\.ai|exa\.ai|blackbox\.ai|edgeservices\.bing\.com|nimble\.ai|iask\.ai|aitastic\.app|bnngpt\.com|neeva\.com

That's ~35 domains, covering the near-universal set (ChatGPT, Claude, Gemini, Copilot, Perplexity, Meta AI, Mistral, DeepSeek, Grok) plus the second tier most guides skip (Poe, Quillbot, Cohere, Hugging Face, Pi, You.com, Phind, and the Edge Copilot sidebar). Some now overlap with the native channel. That's harmless, and it means you still catch them if Google ever drops one from its list.

Keep one owner for this regex and revisit it quarterly. New assistants launch constantly, and a channel group nobody maintains slowly drifts back into undercounting.

The quick filter, when you just need a fast look

Sometimes you don't need a permanent channel. You need one number, now. Skip the channel group and filter a report directly.

1. Open Reports → Acquisition → Traffic acquisition.
2. Set the primary dimension to Session source / medium.
3. Click the search box above the table and type a fragment such as chatgpt, perplexity, or .ai.
4. Read sessions, engagement rate, and conversions for what matches.

For a more flexible one-off, build a free-form Exploration, add Session source / medium as a row, and apply a filter that matches regex using the same string above. That lets you break AI traffic down by landing page or event without committing a channel group. Use the filter for a spot check; use the channel group for anything you report on repeatedly.

How to read the data (the part most guides skip)

Now the honest part. Whatever number you just pulled is the measurable slice of AI-driven demand, not the total. Three mechanisms guarantee GA4 undercounts real AI influence, and the May 2026 native channel fixes none of them.

Most AI answers resolve with no click. When ChatGPT or an AI Overview answers a question using your content, the user often gets what they need in the answer itself and never visits. That influence is real. It leaves no session to count.

Referrers get stripped to Direct. Traffic from native mobile apps (the ChatGPT and Perplexity apps, for example) and from privacy features like Safari's tracking prevention frequently arrives with no referrer at all. GA4 has nothing to match, so it files the visit under Direct, invisible to both the native channel and your regex.

AI-influenced buyers convert later through branded search. An assistant names your company without a live link. The person searches your brand on Google a day later and converts. GA4 credits Organic Search. The AI touch that started the journey is nowhere in the report.

I'm stating these directly rather than re-arguing them. For the full mechanism and why it changes how you value AI as a channel, read the pillar: B2B content attribution in the AI era. The point for setup is narrow: the native channel improved the labeling of the visible slice. It did not grow the slice. The share of AI influence your analytics can never see is the same today as it was in April.

Diagnostic: estimate your own undercount

Rival guides quote a generic "you're seeing 30 to 40% of real AI traffic" figure. That's someone else's site, not yours. You can bound your own gap with data you already have access to. It takes about fifteen minutes.

The idea: compare how often an AI surface showed your content against how many AI sessions GA4 actually recorded over the same window. The difference is exposure your analytics never converted into a counted visit, which is a lower-bound estimate of what GA4 is missing.

• Input A (measured AI sessions, GA4): Reports → Acquisition → Traffic acquisition, last 28 days. Read the AI Assistant channel sessions, then add your custom AI channel sessions. Sum them.
• Input B (AI exposure, Google side): Search Console → Performance → Search results, same 28 days. Add a filter → Search appearance → AI Overviews (where your account has it) and read Impressions. If that filter isn't available, use Bing Webmaster Tools → Copilot citations, or run 20 to 30 of your real buyer questions through ChatGPT and Perplexity and count how many cite you.
• How to compare: Look at the gap between B and A, or the ratio A ÷ B.
• How to read it: Input B counts times an AI surface put your content in front of someone. Input A counts people who actually landed with an AI referrer. Because AI Overviews pass no referrer and most AI answers resolve with no click, almost none of B shows up in A. The portion of B that produced no measured session is demand your reports can't attribute to AI.

Why this is a lower bound, not the true number: it counts only one surface (Google AI Overviews, or one engine's citations) and only the exposure Google can see. It ignores the entire zero-click universe on ChatGPT, Claude, and Perplexity that neither tool records. Your real undercount is larger. The point is that the gap is specific to your site instead of borrowed from a blog.

Run this once a quarter. When the gap widens while your measured sessions stay flat, AI influence is growing faster than your analytics can see, which is exactly the moment to stop reporting the GA4 number as if it were the whole story.

• AEO & SEO Solutions: we build the tracking and the content that gets you cited in the first place, then operate both.
• Fractional Growth Partner: senior marketing leadership to turn AI-channel signal into a revenue story your board trusts.

FAQ