How to Track AI Traffic in GA4 Without the Gaps

How do you track AI traffic in GA4?

Track AI traffic in Google Analytics 4 in two layers. Plenty of guides give you only one. Older ones tell you to build a regex channel from scratch, because they predate the AI Assistant channel Google shipped in May 2026. Newer ones say the native channel is enough. It isn't.

1. Read the native channel. Open Reports → Acquisition → Traffic acquisition, with Session default channel group as the primary dimension, and find the AI Assistant row.
2. Look for what it missed. In the same GA4 report, switch the primary dimension to Session source / medium and search for chatgpt, perplexity, claude, gemini and copilot. Any row not labelled ai-assistant is AI traffic sitting in another channel.
3. Build a custom channel group. In Admin, under Data display, create an AI channel that matches session source only, and order it above Referral.
4. Add the landing page. Use Landing page as a secondary dimension to see which pages assistants send people to.
5. Compare outcomes. Set up key events and compare the AI channel's key event rate with Organic Search over at least 90 days.

When we built the AI-traffic reporting in RankX AI, we checked where AI visits actually landed on three live GA4 properties, on 25 September 2026. Visits from the same assistants turned up in four places: the AI Assistant channel, Referral, Unassigned and Organic Search. Steps 1 and 2 take five minutes. Step 3 fixes the numbers, so it gets its own section below.

What counts as AI traffic in GA4?

AI traffic in GA4 means human visits from an AI platform: someone reads an answer in ChatGPT, Perplexity, Claude, Gemini or Copilot and clicks through to your site. That's AI referral traffic, and it's the only kind of AI traffic Google Analytics is built to attribute.

Two other things get called AI traffic, and neither belongs in this report. AI crawlers such as GPTBot and ClaudeBot fetch pages without running JavaScript, so they never reach GA4; you read them in server logs. Clicks from Google's AI Overviews and AI Mode are AI search traffic, but they arrive inside your organic traffic from Google and GA4 can't split them out.

Mix the three and you'll either undercount AI referrals or credit ordinary organic search traffic to AI tools that never sent it.

What does GA4's AI Assistant channel count?

GA4's AI Assistant channel is a default channel Google announced on 13 May 2026 and rolled out to properties gradually. When a visit's referrer matches Google's list of AI assistants, GA4 sets the medium to ai-assistant and the campaign to (ai-assistant), and the session lands in that channel. The source keeps the hostname, such as chatgpt.com or perplexity.ai.

Google doesn't publish the list, and its two help pages give different examples. The launch note (opens in a new tab) names ChatGPT, Gemini and Claude. The channel definition names ChatGPT, Gemini, DeepSeek, Copilot and Grok, and adds that AI Overviews and AI Mode are excluded and stay in Organic Search.

Practitioner reports have filled the gap, and they disagree. Itamar Blauer, writing in Search Engine Journal on 15 July 2026, noted that Claude had dropped out of Google's documented examples and reported Perplexity landing in Referral. Our data differs on Perplexity: on three GA4 properties measured on 25 September 2026, visits from chatgpt.com, claude.ai, perplexity.ai, gemini.google.com, copilot.com and copilot.microsoft.com all carried the ai-assistant medium.

So don't trust anyone's list, ours included. Google can change an unpublished list whenever it likes, and the only list that matters is the one in your own Session source / medium report this month.

The AI Assistant channel isn't retroactive

The AI Assistant channel only labels visits from the date it reached your property. On the one property we measured with older history, the same AI hosts carried medium referral up to 10 June 2026 and ai-assistant from 11 June. The visits didn't change. Only the label did.

That catches people out on trend charts. Plot the default channel group across that date and AI traffic appears from nowhere in June, which is a relabelling rather than growth. If you've shown a client a sudden AI jump in May or June 2026, check this first.

Where does GA4 put the AI traffic the channel misses?

GA4 puts the AI traffic its AI Assistant channel misses into four other channels, and each needs a different fix. We call this the AI traffic leak map, and it's what we check first on any property. The source / medium values are the ones we actually found:

Unassigned is the one people skip, and it's where tagged AI visits without a referrer end up. OpenAI says ChatGPT adds utm_source=chatgpt.com to its referral links (opens in a new tab). When a click arrives with that tag and no referrer, the session keeps a source but has no medium. GA4 assigns the AI Assistant channel by referrer, so nothing claims it and that traffic from ChatGPT lands in Unassigned.

So a visit with no referrer isn't automatically lost. If the link carried a tag, the source survived, and a rule on source alone puts the visit back. Only a visit with neither a referrer nor a tag is genuinely invisible to GA4.

One more source to know: vertexaisearch.cloud.google.com. It's the redirect used when a third-party app grounds its answers on Google's Gemini models, and GA4 files it under Referral. Count it as AI traffic, but give it its own channel rather than crediting Gemini, because the visitor was in somebody else's app.

How do you build a custom AI channel group in GA4?

A custom AI channel group in GA4 is a copy of the default rules that sort your traffic sources into channels, with AI channels added near the top. In a standard GA4 setup it's the only fix that reaches every recoverable row of the leak map, and Google states that custom channel groups apply to your reports retroactively (opens in a new tab). You need the Editor role or above, and a standard property allows two custom groups.

1. In Admin, open Data display, then Channel groups, and create a new channel group. Call it something like Default with AI.
2. Add a channel called AI assistants with one condition: Source, matches regex, using the pattern below.
3. Don't add a medium condition. The medium is exactly what differs between the ai-assistant, referral and (not set) versions of the same visit, so matching on it splits ChatGPT across three channels again.
4. Add a second channel called AI grounding: Source exactly matches vertexaisearch.cloud.google.com.
5. If you buy ads inside assistants, add a Paid AI channel: Source matches the same pattern, and Medium matches (cpc|ppc|paid|paidsearch|paid_search|cpm|display|sponsored).
6. Choose Reorder. Put Paid AI first, then AI assistants and AI grounding, all above Referral and Organic Search. Save the group and switch the primary dimension in Traffic acquisition to it.

This is the pattern we use. It's built from the answer hosts and UTM values we've seen in real GA4 data, plus the other answer hosts Google's channel definition names:

(www\.)?(chatgpt\.com|chat\.openai\.com|chatgpt|openai|perplexity\.ai|perplexity|claude\.ai|claude|gemini\.google\.com|gemini|copilot\.microsoft\.com|copilot\.com|copilot|grok\.com|meta\.ai|chat\.deepseek\.com|chat\.mistral\.ai)([/?#:].*|https?:.*)?

It lists exact answer hosts, the bare UTM values assistants and tagged links actually send, and a tail that tolerates malformed sources. We found one in the wild: a page had concatenated URLs into its utm_source, producing a source that began chatgpt.comhttps://. The pattern is 249 characters, inside the 250 Google allows for regex in report filters, so it also fits a temporary report filter.

Google's regex documentation says matching is a full match by default, so the pattern needs no anchors. Case depends on where you paste it. Google says channel group definitions aren't case-sensitive, but report filters are, so a temporary filter using this pattern will miss a source tagged ChatGPT with a capital.

Which AI regex patterns go wrong in GA4?

Three mistakes recur in AI regex patterns shared online, and Google's own example makes two of them. We tested Google's example and a widely linked URL-style pattern against GA4's full-match rule, using 16 AI sources (nine found in real GA4 data, seven known answer hosts) and 17 sources that aren't assistants:

• Matching a vendor's whole domain. openai.com, anthropic.com and x.ai are corporate sites. platform.openai.com is a developer console and help.openai.com is documentation. A visit from either is a developer reading docs, not someone who followed an answer.
• Matching anything ending in ai. Google's own example pattern begins ^.*ai, which under full-match rules counts every source ending in those letters: writing tools such as jasper.ai and copy.ai, and meeting tools such as otter.ai. Its .*gemini.* also matches gemini.com, a cryptocurrency exchange. On our test set it matched 11 of the 17 non-assistant sources, two of them invented look-alikes, and it missed grok.com and chat.deepseek.com, two assistants Google's own channel definition names.
• Applying a URL pattern to Source. Patterns that begin ^https:\/\/ are correct on the Page referrer dimension, which holds a full URL. On Session source, which holds a bare hostname such as chatgpt.com, the same pattern matches nothing at all, and the channel stays silently empty.

Keep your list to the hosts answers are actually served from, and review it every quarter. New AI platforms launch, existing ones change domains, and GA4's own list moves without notice.

Why does GA4 show zero AI traffic?

When a GA4 AI channel shows zero AI traffic, check four things before you decide assistants send you nobody. Each one produces an empty channel with no error message:

• The report is still on the default group. A custom channel group only appears when you choose it as the primary dimension, unless you've made it the property's primary channel group.
• The date range predates the channel. The native channel shows nothing before it reached your property, and on a small site AI referrals can be days apart.
• The channel sits below Referral. Google states that traffic goes to the first channel whose rules it matches, in order, so an AI channel under Referral never receives a chatgpt.com / referral visit.
• The regex can never match. A pattern written for full URLs matches nothing on Session source, which holds a bare hostname.

How do you see which pages AI assistants send people to?

To see which pages AI assistants send people to, add Landing page as a secondary dimension in Traffic acquisition with your AI channel group as the primary dimension. For more control, build a free-form Exploration with Session source and Landing page as dimensions, sessions, engaged sessions and key events as metrics, and a filter on your custom channel group.

Keep Session source in the table, so you compare engagement rate assistant by assistant instead of averaging every AI platform into one figure. Explorations only reach back as far as your data retention setting, 2 or 14 months on a standard property, while standard reports aren't limited by it.

Expect the homepage to lead. ChatGPT replaced footnote-style citations with clickable brand names in May 2026, and Similarweb's data through late May showed homepage referrals growing several times faster than ChatGPT referrals overall. A homepage-heavy report reflects how ChatGPT now links, so judge your deeper pages on their own rows.

Then take product flows out. On properties whose product connects to ChatGPT or Claude, we found GA4 counting sessions that landed on OAuth consent, login and account pages as AI Assistant traffic. Those are users connecting an integration, not an answer recommending you. Exclude landing paths such as /login, /oauth and /account from the AI view before you report anything.

How do you measure conversions from AI traffic in GA4?

Measure conversions from AI traffic in GA4 by marking the actions that matter as key events, then comparing your AI channel's key event rate with Organic Search on the same report. Use at least 90 days. AI sessions are few on most sites, and on 30 sessions a single conversion moves the rate by more than three points.

Check which key events are real first. Every Google Analytics 4 property we inspected carried purchase as a default key event, so a lead-generation site can look configured while recording nothing. Add the events that represent your actual outcomes, such as a demo request or a trial sign-up, and report against those.

Keep paid clicks out of the comparison. Ads now run inside ChatGPT, and Similarweb's 2026 generative AI report, summarised on its blog in July 2026, puts roughly 26 per cent of ChatGPT responses as containing one, without stating the market. That's what the Paid AI channel in the build steps is for.

Be wary of published conversion multipliers. Most come from one company's own site or from modelled value rather than measured rates, and none of them is your funnel. The AI referral traffic glossary entry covers why volume alone misreads this channel. Your own 90-day comparison is the number worth reporting.

How much traffic are AI assistants sending?

A small share of total traffic, for most sites. Conductor's AEO and GEO benchmarks, drawn from 1,215 of its enterprise customers' domains and more than 3.3 billion US sessions between May and September 2025, put AI referrals at 1.08 per cent of website traffic (opens in a new tab), with Information Technology highest at 2.8 per cent. ChatGPT accounted for 87.4 per cent of those AI referrals.

ChatGPT still sends most of it. Statcounter measures page views across the more than 1 million sites running its tracking code, and in September 2026 it put ChatGPT at 80.48 per cent of its AI chatbot share. Google Gemini followed at 10.91, Perplexity at 5.85, Microsoft Copilot at 1.79 and Claude at 0.95 per cent. A setup that only matches chatgpt.com would leave out roughly a fifth of what Statcounter counts.

What can't GA4 tell you about AI traffic?

GA4 can't see four kinds of AI influence, and for three of them there's a better instrument than analytics.

Clicks from Google AI Overviews and AI Mode

Clicks from Google AI Overviews and AI Mode arrive as google / organic, indistinguishable from ordinary search clicks, and Google's channel definition excludes them from the AI Assistant channel. Google Search Console's generative AI performance report shows their impressions by page, country, device and date, with no clicks or queries. Our guide to Search Console impressions without clicks covers what that report can and can't show, and AI Mode versus AI Overviews explains why the two are reported together.

Visits that carry no source at all

A copied link, or a click from an app that sends no referrer and no tag, reaches GA4 as Direct. Cloudflare stated in July 2025 that traffic referred by Claude's native app carries no Referer header. No setting recovers these visits after the fact, because the information never reached your site.

You'll see a figure claiming 70.6 per cent of AI traffic lands in Direct. It comes from Loamly, an AI-traffic vendor, and in a February 2026 update describes 14,413 of 20,428 AI visits in its own database of 446,405 visits, classified by detection models whose weaker layers score confidence between 0.35 and 0.65. It's an estimate of what direct traffic contains rather than a measurement of GA4, and it should be reported as one.

By the mechanism above, a visit that keeps its utm_source lands in Unassigned rather than Direct, so the label overstates even on its own terms.

Two honest proxies work better than any single number. Watch new-user Direct sessions that land on deep informational pages, which nobody types from memory, and add a "How did you hear about us?" field to your highest-value form with ChatGPT and other assistants as options.

Sessions inside AI browsers

AI browsers don't announce themselves. Statcounter reports that Perplexity's Comet, The Browser Company's Dia and OpenAI's ChatGPT Atlas all identify as Chrome, so a session inside one of them looks like any other Chrome session in GA4's browser report. Nothing in GA4 separates them today.

Being named without a click

Most AI answers carry no link at all: Similarweb found a citation in about 6.8 per cent of US ChatGPT prompts in May 2026. Crawlers such as GPTBot and ClaudeBot never run your GA4 tag either. A brand can be recommended daily and show almost nothing in analytics, because an AI citation that nobody clicks leaves no trace there.

Tracking brand mentions in AI search covers the answers themselves, tracking AI bots covers the crawlers in your server logs, and measuring AI search visibility sets out how the layers fit together.

The AI traffic audit checklist for GA4

The AI traffic audit checklist below is what we run on a GA4 property before trusting any AI traffic data from it. It takes about 20 minutes once you know where to look:

1. Find the date the AI Assistant channel reached the property, and never chart the default channel group across it.
2. Search Session source / medium for every AI host and note each medium it appears with.
3. Open Unassigned and look for AI sources with a (not set) medium.
4. Create a custom channel group with a source-only AI rule, ordered above Referral and Organic Search.
5. Test the regex against your real Source values before you trust the channel.
6. Remove vendor corporate and developer domains from the pattern.
7. Give vertexaisearch.cloud.google.com its own channel.
8. Exclude login, OAuth and account landing pages from the AI view.
9. Split paid AI clicks into their own channel if you run ads inside assistants.
10. Replace the default purchase key event with the outcomes your business actually records.
11. Next to every AI traffic figure you report, note what GA4 can't see: AI Overviews, AI Mode, AI browsers and untagged Direct visits.

How RankX AI reports AI traffic

RankX AI reads your GA4 property and classifies AI visits with the rules this guide describes: exact answer hosts rather than vendor domains, bare UTM values, and malformed sources. It leaves out sign-in and account pages, reports paid AI clicks separately, and applies its rules each time you open the report, so an improved rule relabels your whole history instead of starting a new trend line.

RankX AI then puts each assistant's sessions next to how often that assistant mentions you across your tracked prompts. That answers the question GA4 can't: is being cited actually sending anyone? Sessions with no mentions usually mean an assistant is citing you for prompts you don't track yet. Mentions with no sessions mean you're cited and nobody clicks. Run the free AI visibility checker to see where you stand today, or see how AI visibility tracking in RankX AI works.