How to track AI traffic in Google Analytics 4 (and test whether it converts better)

During Chile's CyberMonday sale in October 2025, a kitchenware and small-appliance store saw the biggest spike of visits from ChatGPT and Gemini it had ever recorded. Its agency, Metrix.digital, only caught it because it had set up a separate AI channel in GA4 a few weeks earlier. Without it, those sessions would have been scattered across Referral and Unassigned. This guide shows how to track AI traffic in GA4 the same way, which regex to use as of September 2026, and how to tell whether the conversion rate you see is real or just noise from a small sample.

The main source is a SearchBrand micro-training with Gonzalo Aburto (Head of Digital) and Alejandro Momberg (SEO Manager) from Metrix.digital, a Chilean agency. Metrix uses SearchBrand with its clients, so weigh their comments on the platform with that in mind. The case studies and the walkthrough are theirs. The math corrections and the calculator are ours. The session was in Spanish; quotes below are our translation.

Why GA4 doesn't show AI traffic out of the box

GA4 assigns every session to a channel based on source and medium. A visit from chatgpt.com arrives with medium referral, so it lands in Referral next to every blog that links to you. When ChatGPT appends utm_source=chatgpt.com and there's no medium, the session can end up in Unassigned. Gonzalo put it plainly: without a dedicated channel, this traffic "goes to referral, sometimes to unassigned" (27:03).

A custom channel pulls it into one row. Know up front what it will miss, because it's a lot.

How the visitor arrives	What GA4 records	Caught by the AI channel?
Clicks a link in ChatGPT, Perplexity or Claude in a browser	chatgpt.com / referral, perplexity.ai / referral	Yes
Link tagged utm_source=chatgpt.com	chatgpt.com / (not set)	Yes, if the rule uses source
Clicks from a desktop or mobile app that sends no referrer	(direct) / (none)	No
Clicks an AI Overview or AI Mode result in Google	google / organic	No
Copilot inside Bing	bing.com / referral	Can't be split from search
Meta AI inside WhatsApp, Grok inside X	(direct) or t.co	No
The AI answers and nobody clicks	No session at all	No

The last three rows are why an AI channel always undercounts. For AI Overviews you need Search Console, not GA4.

Before you build anything, open your property's default channel group and check whether Google has added an AI channel of its own. We didn't assume either way when we reviewed this guide on September 24, 2026. If your property has one, compare its rules with the regex below before you duplicate the work.

Step by step: build an AI channel group in GA4

Alejandro walked through this live (27:03). Here it is with the current English menu names:

1. Go to Admin → Data display → Channel groups.
2. On the default channel group, open the three-dot menu and choose Copy to create new. Give it a clear name, such as "Channel group with AI".
3. Click Add new channel, name it "AI" and set the condition: Source · matches regex · the regex you generate below.
4. Save the channel and drag it above Referral. GA4 checks channels top to bottom and assigns each session to the first match. Leave "AI" at the bottom and Referral swallows everything, so your new channel reads zero.
5. Save the group. In Reports → Acquisition → Traffic acquisition, switch the table's channel group dropdown to the new one.
6. To see which assistant sends the most, build an exploration in Explore with Session source as the dimension, filtered to the AI channel.

Why copy the default group instead of starting blank? Alejandro's answer:

"The point of copying the one I already had and adding this traffic, instead of building it on its own, is that it lets me see the share I'm getting from that channel." — Alejandro Momberg, SEO Manager at Metrix.digital [30:05]

Custom channel groups apply to historical data too, so you'll see past months the day you create it. Still, write down the date in your reporting. Anyone comparing an old report with a new one needs to know why Referral dropped.

Generate your regex and test it against your own sources

In the webinar, Alejandro built the rule around six names (ChatGPT, OpenAI, Perplexity, Copilot, Gemini, Claude). A year later the list is longer and some domains have moved. ChatGPT went from chat.openai.com to chatgpt.com in 2024, and Bard was renamed Gemini that February. DeepSeek and Grok now send their own visits, as do Meta AI and Mistral's Le Chat.

^(.*\.)?(chatgpt\.com|chat\.openai\.com|perplexity\.ai|gemini\.google\.com|copilot\.microsoft\.com|claude\.ai|chat\.deepseek\.com|meta\.ai|grok\.com|chat\.mistral\.ai)$

With the default selection, the tool outputs this regex (168 characters):

^(.*\.)?(chatgpt\.com|chat\.openai\.com|perplexity\.ai|gemini\.google\.com|copilot\.microsoft\.com|claude\.ai|chat\.deepseek\.com|meta\.ai|grok\.com|chat\.mistral\.ai)$

The (.*\.)? prefix accepts subdomains like www.perplexity.ai. The closing $ blocks false positives like claude.ai.someothersite.com. We reviewed the domain list on September 24, 2026, and it changes every few months: a new assistant launches, a short domain appears, a product starts tagging links differently. That's why the tool asks you to paste your own source list from the last 90 days. Anything containing "chat", "gpt" or "copilot", or ending in ".ai", that the regex misses shows up as a candidate for you to judge. Some entries, like copilot.com or copilot.cloud.microsoft (Microsoft 365 Copilot), are flagged "verify" and left off by default. Turn them on only if they appear in your data.

If the regex runs past 256 characters, the tool splits it in two and tells you how to join them with an OR condition. That's the cap GA4 has applied to regex fields; check it in your property if you get an error on save.

If you'd rather start with what AI says about your brand than with GA4 settings, SearchBrand tracks mentions and share of voice in the assistants your buyers use, starting with ChatGPT and Gemini. Sign up with 10 free credits, no card required.

Worked example: set up the channel and check whether your lift is real

We'll use the numbers from Metrix's e-commerce client to run the full process. Sessions are estimated from the rates they showed; everything else comes from the webinar.

Step 1. Export your sources. In Traffic acquisition, switch the primary dimension to Session source, pick the last 90 days and download the CSV (or copy the table). Paste it into the tester.

Step 2. Generate the regex and review what doesn't match. If chat.deepseek.com shows up with 4 sessions, or an unknown domain with "gpt" in its name, look it up before you add it. A customer-support chat widget can have "chat" in its hostname and nothing to do with AI.

Step 3. Build the channel in GA4 using the steps above. The step people skip: moving it above Referral.

Step 4. Write down the numbers. The store had 20 purchases from AI over six months. At a 3.1% conversion rate, that's roughly 645 sessions. The site converted at 2.2% on about 20,000 transactions in the same half-year (Gonzalo cited 40,000 over 12 months).

Step 5. Run them through the calculator. Here's what comes out:

Metric	AI channel	Rest of site
Conversion rate	3.1%	2.2%
95% confidence interval	2.0% to 4.7%	2.17% to 2.23%
Difference	+0.9 pp (+41%)
p-value	0.12
Verdict	Sample too small

The AI channel's interval starts at 2.0%, below the site's 2.2%. With 20 purchases you can't claim that traffic converts better. At the same rates, you'd need about 40 purchases (around 1,300 AI sessions) before the gap becomes significant.

Step 6. Repeat monthly. The number to watch is when the channel crosses 30 or 40 conversions, not the lift in month one.

Step 7. Add what GA4 can't see. Mentions, share of voice and server logs, covered below.

What Metrix.digital's two cases actually show

Gonzalo presented two anonymous clients. The figures come from his slides, and they read better with the corrections below.

E-commerce during CyberMonday: the spike and the small-sample trap

Gonzalo described the first-day spike as "never seen before." During the previous sale in May, nothing had registered (22:20). He put the 20 purchases at "around 0.01%" of the year's 40,000. The actual figure is 0.05%, and it compares six months against twelve; over the same half-year it's closer to 0.1%. Still marginal. His reading of the rate (3.1% vs. 2.2%, "a lift of almost one percentage point") is accurate as a description. As proof that the channel converts better, it falls short, for the reasons in the worked example.

Healthcare retail: 833 bookings in 12 months

The second client is a Chilean healthcare retailer with far more volume (25:38). Over 12 months the AI channel drove 833 appointment bookings, and the trend was still rising. The channel converted at 26% against the site's 21% (Gonzalo first said 26% for the site, then corrected himself). He called it "three percentage points more." It's five.

With 833 conversions the picture changes. Even assuming 500,000 sessions for the rest of the site, the AI channel's interval runs from 24.5% to 27.5%, entirely above 21%. Here you can say it converts better.

"In volume it's still marginal. In conversion rate it stands out well above the other acquisition channels." — Gonzalo Aburto, Head of Digital at Metrix.digital [27:06]

Why it converts better, and why we wouldn't credit AI for all of it

Hussam Sufan, SearchBrand's co-founder, pointed to the funnel: the model sends people straight to the checkout or booking page, which cuts out "that exploratory funnel of comparing five sites, 10 different sites" (33:04).

A second reading fits alongside that one. Someone arriving from ChatGPT has already done the research inside the chat, so they show up with more intent. The lift measures those people, not what AI does to the same people. For budget decisions the distinction doesn't matter much. For telling your boss that doubling AI traffic will double sales, it does.

What GA4 can't see: contextual visibility

The AI channel counts visits. Most of what happens in ChatGPT or Gemini never becomes a visit: the model names your brand, compares it with two competitors, and the user decides right there. Gonzalo proposed three metrics for that part (7:10): share of voice (your share of presence in answers against competitors), mentions in LLMs, and search volume per AI platform.

"This doesn't generate traffic, you won't see it in GA4, you won't measure it with your traditional SEO tools, but you need to quantify it and optimize it over time." — Gonzalo Aburto, Head of Digital at Metrix.digital [9:06]

Gonzalo said Metrix concluded neither Semrush nor GA4 was enough for this, after trying SE Ranking and HubSpot's AI module on top of Semrush, which they already used. For a structured way to run that kind of check, see our AEO and GEO audit guide.

Two more things from the session that GA4 won't show you:

• What AI says about you. Hussam noted that ChatGPT already cites Instagram posts, and that cuts both ways: the model repeats what it finds. At an earlier event, an attendee described ChatGPT attributing a health risk to a mass-market consumer brand, one the brand says it can disprove. Gonzalo suggested tracking the sentiment of each mention, counting how often you appear and how often it's negative (45:59).
• When the bots show up. On llms.txt, Hussam said there are signs it works but it isn't a rule yet, and "between having it and not having it, better to have it." Alejandro suggested testing it with data:

"Implement it, and if you can get the server logs, check whether ChatGPT actually came in more often or not." — Alejandro Momberg, SEO Manager at Metrix.digital [39:04]

Server logs show visits from bots like GPTBot or PerplexityBot, which GA4 never records because they don't run JavaScript.

Growing that traffic: what's working for Metrix

Measurement only pays off if you change something afterward. This is what Gonzalo and Alejandro said they're applying across their SEO clients, about 30 of them according to Gonzalo.

Short blocks that stand on their own. One section, one topic. Gonzalo's reasoning (12:11) is that the model builds answers from fragments. It doesn't take the whole page; it "takes just one paragraph from your site and combines it with a paragraph from another site." A paragraph that depends on the one before it loses that contest.

FAQs and complete product pages. Alejandro said "a large share of the citations" they get in AI Overviews are FAQs. For an automotive client, the most-cited pages are vehicle spec sheets with full detail on technology and safety (51:00). If you run a catalog, start there before writing new articles.

One niche, not ten. Gonzalo compared it to domain authority: if your site covers a topic in depth, models treat you as a reference on it. Alejandro added that AI sorts you as a leader or a niche brand and pairs you with certain competitors. If you're a niche player chasing prompts where the leaders show up, "you may be aiming your effort in the wrong place."

Technical SEO basics. Asked whether traditional SEO and AI optimization conflict, Alejandro said they complement each other (34:22), because ChatGPT pulls results from Bing or Google before it answers. If your page shows up as an indexable search result, he said, you raise your odds of appearing in an LLM. He also warned that a single-page React site rendered on the client "will be impossible for an LLM to navigate," whatever its structure. In the same segment Hussam argued that technical work will matter less than content and topical authority. We side with Alejandro here: as long as assistants pull from Bing and Google, clean indexing is the price of entry.

What we don't recommend

• Reporting a lift with fewer than 30 conversions in the AI channel. The e-commerce case shows why: +41% makes a nice headline while the interval still overlaps the site rate.
• Building the channel group from scratch. You lose the comparison with the total, which is exactly what your boss will ask about.
• Using bare chat, ai or gpt in the regex. They catch support widgets, .ai domains that aren't assistants, and any site with "chat" in its name. Use full domains.
• Adding bing.com or google to the rule. You blend search with assistants and the channel stops meaning anything.
• Reusing a regex from 2024. Without chatgpt.com it misses nearly all ChatGPT traffic. In the properties Alejandro and Gonzalo reviewed, ChatGPT accounted for about 80% of AI channel sessions.
• Treating the AI channel as your AI visibility. It's only the part that clicks through.

The full micro-training

Recorded in October 2025 with Metrix.digital, in Spanish. Jump straight to:

• New metrics: share of voice and LLM mentions (7:10)
• Case 1: e-commerce during CyberMonday (22:20)
• Walkthrough: tracking AI traffic in GA4 (27:03)
• Brand reputation risk in AI (45:59)

A few things changed after the recording. The tool carries the updated domain list, and we checked the GA4 menu names as of September 2026.

Once your AI channel is running and you want to see which questions in your category AI answers without naming you, try SearchBrand with 10 free credits, no card required.

Sources

• SearchBrand and Metrix.digital, "Micro-Training #2: Cómo medir tu presencia en la IA," YouTube, published October 28, 2025.
• Google Analytics Help, "[GA4] Default channel group": https://support.google.com/analytics/answer/9756891
• Wilson interval and two-proportion z-test: the tool on this page, using the webinar figures.