How to Track AI Traffic in GA4: The Channel, the Regex and the Blind Spots

The AI Assistant channel Google switched on in May 2026

On May 13, 2026 GA4 gained a default channel called AI Assistant. When a visit's referrer matches an AI assistant Google recognizes, GA4 sets the medium to ai-assistant and the campaign to (ai-assistant), and the session gets its own row in the Traffic acquisition report. Google's channel definition names ChatGPT, Gemini, DeepSeek, Copilot and Grok as examples, and the launch note adds Claude. There is no tag to change and nothing to install. If you have GA4, you already have it.

Two details matter more than the launch. First, the channel deliberately excludes Google's own AI Overviews and AI Mode. Clicks from those stay in Organic Search, blended with every ordinary blue-link visit. Second, Google has not published the full list of assistants it recognizes. Perplexity is the obvious question mark: the analytics vendor Terminus reports it is not confirmed as part of the channel, and its visits can arrive as both perplexity.ai and www.perplexity.ai. In June 2026 GA4 also added a Source Group dimension that bundles messy source values, with built-in groups for ChatGPT (OpenAI) and Perplexity. Google says it is populated retroactively, so it works on older data too.

A custom channel that catches the stragglers

The default channel is fine for a glance. For a report you would hand a business partner, build your own channel group, so you decide exactly which sources count as AI and Perplexity sits next to ChatGPT where it belongs. Custom channel groups apply to past data as well as new data, so the moment you save one, last quarter is reclassified in your reports. Standard GA4 properties allow two custom groups beside the predefined one, each with up to 50 channels, and you need the Editor role or higher on the property.

1. 1

   Open the channel groups screen

   In GA4, go to Admin, then Data display, then Channel groups, and choose to create a new channel group. Starting from a copy of the default keeps every existing channel intact, so only the AI traffic moves.

2. 2

   Add a channel named AI referrals

   Click Add new channel, give it a name, then add a condition group: dimension Source, operator matches regex. Paste the pattern from the box below as a single line.

3. 3

   Check the pattern against your data

   Before saving, open Traffic acquisition, switch the dimension to Session source and search for each hostname in the pattern. If a source you expected never appears, it simply has not sent you a visit yet.

4. 4

   Move it above Referral

   GA4 files each session in the first channel it matches, top to bottom. Click Reorder, drag AI referrals above Referral (above Organic Search too, to be safe) and apply. Left below Referral, the new channel stays empty forever.

5. 5

   Save and switch the report

   Save the group. In Reports, open Traffic acquisition and change the primary dimension from Default channel group to your new group. Look at landing pages and key events for the AI referrals row, not just session counts.

6. 6

   Revisit it monthly

   Once a month, scan the Referral channel's source list for new AI hostnames. Engines launch, rename and move domains. Add what appears, and because the group is retroactive, the history updates with it.

A word on loose patterns. The GA4 help page for channel groups includes its own sample AI regex built from fragments such as .*gpt.* and .*copilot.*, and it opens with ^.*ai. Fragments like those are quick to write and easy to trip: any source whose name happens to contain the letters can land in the AI row, and a newsletter or a software vendor becomes an AI engine overnight. A list of exact hostnames is slower to type and much harder to fool. Whichever you use, test it against your real source list before trusting the totals.

The visits GA4 was never built to see

Everything above counts clicks, and most AI answers do not produce one. Pew Research Center found that in March 2025 US users clicked a link inside a Google AI summary on just 1% of visits to pages that showed one. Cloudflare estimated in July 2025, from crawl-to-referral ratios, that getting a visit from OpenAI was 750 times harder than getting one from Google search used to be. The customer who reads your name in an answer and then dials your number, or drives over, leaves no row in any analytics tool.

That second column is why analytics is only one leg of AI visibility monitoring. Someone also has to read the answers, on a schedule, for the questions customers ask. The free AI visibility check does one round of that: it puts a customer's question to ChatGPT and other engines through their APIs and shows whether you were named and who was named instead. API answers can differ a little from the consumer apps, so treat it as a sample, not a verdict.

Search Console and Bing cover the zero-click half

Google keeps AI Overview and AI Mode activity out of GA4's AI channel, but it does count it in Search Console. Clicks and impressions from both features sit inside the Performance report under the Web search type, where an AI Overview takes a single position and every link in it shares that position. Since August 31, 2026 every site also has a Generative AI performance report, which shows impressions only: how often links to your site appeared in AI Overviews and AI Mode, by page, country, device and date.

Lay the four side by side each month and a pattern appears. Impressions and citations show reach, the GA4 channel shows the trickle of visits, and key events show whether the trickle is worth anything. Our guide to what a GEO report should show puts those sources on one page in a sensible order.

Reading small numbers without lying to yourself

For most local businesses the AI channel will be small. Say a roofer in Boerne sees 14 sessions from AI assistants in September and 22 in October. That is not 57% growth worth announcing at the Monday meeting. It is eight visits, and one homeowner reloading the page on a phone could be three of them. Look at the trend across a quarter, and look at what the visits did.

• Landing pages first. AI visitors often land deep, on a service or pricing page that an answer cited. Those pages deserve your clearest phone number and your best offer.
• Key events over sessions. One tracked call or form fill from AI referrals outweighs a hundred visits that bounce. If calls are not recorded as key events yet, fix that first; our conversion tracking setup exists for this job.
• Watch Direct. A climb in Direct visits to deep pages, with no campaign to explain it, can be AI app traffic with the referrer stripped. You cannot prove it, but you can note it in the report.
• Compare with Organic Search. If AI visitors stay longer and convert better, their landing pages are doing their job. If they leave fast, the answer may be sending them for the wrong reason, which is worth checking against what the engine actually says.

Analytics counts the customers who clicked. The ones who just called are in your phone log, not your dashboard.

When to hand the plumbing to someone else

Building the channel takes ten minutes. Keeping the whole measurement picture honest takes longer: call tracking that feeds GA4, a linked Business Profile, Search Console and Bing Webmaster Tools verified, a fixed list of customer questions checked every month, and someone who reads the answers. If your tracking is already a tangle, the tracking problems page walks through the usual culprits, and our generative engine optimization guide explains how each engine chooses what to cite.

DataCram sets all of this up inside its monthly SEO programs, where GEO work is included. Programs start at $625 a month, most local businesses land between $625 and $2,250, and there is no long contract; the tier is set after a free GEO audit. GEO pricing explains what moves the number. If your question is which customer questions to watch, start with prompt tracking.