How to track AI referral traffic in GA4: ChatGPT, Gemini, Perplexity, Claude and Copilot

What counts as AI referral traffic in GA4?

AI referral traffic is any session that starts when someone clicks a link inside an AI assistant's answer and lands on your site. ChatGPT, Gemini, Perplexity, Claude, Microsoft Copilot, DeepSeek and Grok all produce it. GA4 records it like any other visit: a source (where the click came from), a medium (what kind of click it was) and a channel (the bucket those two values fall into).

Three kinds of AI exposure don't show up as AI referral traffic, and it helps to separate them from the start. The first is a mention with no click: an assistant names your brand, the reader remembers it, and they type your URL a day later. GA4 records that visit as Direct. The second is Google's own AI features. A click from an AI Overview or an AI Mode answer comes from a Google search page, so it's counted as Organic Search. The third is a crawler. GPTBot, ClaudeBot and PerplexityBot fetch your pages but don't run JavaScript, so they never trigger the GA4 tag at all. If you want to know whether those bots reach your pages, that's a server log question, covered in our list of AI crawler user agents.

So "AI traffic in GA4" really means one narrow thing: human clicks from standalone assistants that carried enough information (a referrer or a UTM tag) for GA4 to know where they came from. It's the most measurable slice of AI search, and also the smallest.

Does GA4 track ChatGPT traffic automatically now?

Yes, for most of it. Since May 13, 2026, GA4 automatically assigns the medium "ai-assistant" when a session's referrer matches Google's list of recognized AI assistants, and those sessions appear in a new AI Assistant channel in the Default Channel Group reports. You don't have to change any settings, and older custom setups keep working alongside it.

Google's Analytics release notes for May 13, 2026 list three changes. The Medium dimension gets a new value, "ai-assistant", whenever the referrer matches a recognized AI assistant. The Default Channel Group gets a new channel, AI Assistant. And the Campaign dimension shows "(ai-assistant)" for those sessions. The note names ChatGPT, Gemini and Claude as examples.

The default channel group documentation adds two details that matter for reporting. It describes AI Assistant as the channel for users who arrive "from sources like ChatGPT, Gemini, Deepseek, Copilot, or Grok", and it says the channel excludes Google's AI Overviews and AI Mode. In the rule list on the same page, AI Assistant sits after Organic Search and before Referral, and its definition is simply that the medium exactly matches "ai-assistant". Google doesn't publish the full list of assistants it recognizes, so you can't check coverage in advance. You can only check what your own property received.

Before May 13, the same visits usually showed up as "chatgpt.com / referral" in the Referral channel, which is why so many guides still tell you to dig through Referral. Several of the top Google results for this question still say GA4 files ChatGPT visits under Referral by default. That was true for a long time. It isn't the whole picture anymore, and a report built on that assumption now undercounts, because AI visits have moved out of the channel it filters.

What AI traffic looks like in a real GA4 property

In AI Ranked's own GA4 property, over the 28 days from September 24 to October 21, 2026, the AI Assistant channel recorded 86 sessions. ChatGPT sent 44, Gemini 27 and Perplexity 15. Claude and Copilot sent none we could see. AI visits were 5.9% of the size of Organic Search and engaged less often than search visitors did.

We're a young publication, so these are small numbers, and we're showing them for a different reason: they're real, unfiltered GA4 output from a property with no custom AI setup at all, which makes them a clean check on how the new channel behaves. Every number below comes from the standard Traffic acquisition report, read on October 22, 2026.

Four things in that table are worth a closer look, because none of the older GA4 guides could have shown them.

Perplexity is in the AI Assistant channel. Google's channel description doesn't name Perplexity, and some write-ups of the launch assumed it was left out. In our property, perplexity.ai sessions arrived with the medium "ai-assistant", so GA4 recognizes it even though the help page doesn't list it.

ChatGPT visits are tagged and still classified. ChatGPT adds utm_source=chatgpt.com to many of the links it shows, a behavior Search Engine Roundtable covered when a bug briefly removed the tag in October 2026. In our data, those sessions still came through as "chatgpt.com / ai-assistant", not "chatgpt.com / (not set)", so the tag doesn't knock them out of the channel.

Gemini shows up under two source names. Most Gemini sessions carried the source "gemini.google.com", but 4 carried the plain source "gemini" and were still put in AI Assistant. If you filter on the exact string "gemini.google.com", you'll miss them.

Some AI visits fall through. The 2 "gemini / (not set)" sessions and the 1 "perplexity / (not set)" session arrived with a source value and no medium. They're not in the AI Assistant total of 86, which matches the four ai-assistant rows to the session. That's only 3 sessions here, but the pattern is the important part: a tagged visit with an empty medium doesn't reach the AI channel by default.

One caveat on Direct. GA4's own insight panel flagged that Direct jumped to 879 sessions on September 22 against an expected 67, alongside a spike from Singapore, which looks like automated traffic. That inflates the Direct total and drags its engagement rate down, so don't read AI Assistant's 1.98% share of all sessions as meaningful. Comparing AI Assistant with Organic Search (86 against 1,470, or 5.9%) gives a fairer sense of scale for a site like ours.

How do you see AI traffic in Google Analytics?

Open Reports, then Acquisition, then Traffic acquisition. The table's default dimension is "Session primary channel group (Default Channel Group)", and AI Assistant appears as a row once your property has received at least one recognized AI visit. Switch the dimension to "Session source / medium" to see each assistant on its own line.

The full sequence, with the settings we use:

• Reports, then Acquisition, then Traffic acquisition. Set the date range to at least 28 days; daily AI numbers on most sites are too small to read.
• Click the dimension drop-down above the first column and pick "Session source / medium". Type "ai-assistant" in the table's search box to keep only AI rows.
• To check landing pages, click the plus icon next to the dimension name and add "Landing page + query string" as a secondary dimension.
• To check conversions, pick your key event in the "Key events" column header. Session key event rate is the number to compare across channels.
• Save a copy of the report with your filter in place (Save changes as a new report) so you don't rebuild it every month.

If the AI Assistant row doesn't appear at all, look at the date range first. The channel only applies to sessions after it rolled out, and on a small site a 7-day range may simply contain no AI clicks. If the range is long and the row is still missing, search the source / medium table for "chatgpt", "gemini" and "perplexity". Rows with "referral" next to them predate the change; rows with "(not set)" are the tagged-without-medium case covered below.

For a quick picture of how AI fits into your wider reporting, our guide on how to measure AI visibility puts this number next to the metrics that don't depend on clicks.

When do you still need a custom AI channel group?

You still need one in three cases: AI visits arrive tagged with a source but no medium, you want assistants Google doesn't recognize (smaller or regional ones), or you need a history that starts before May 2026. A custom channel group fixes all three, applies retroactively to reporting, and doesn't change your raw data.

Custom channel groups are defined in Admin, under Data display, then Channel groups. Google's custom channel group help page describes the two rules that trip people up. You start from a copy of the default group, and traffic is assigned to the first channel whose definition it matches, in the order the channels are listed. Put an AI channel below Referral and every chatgpt.com / referral session lands in Referral first.

The setup we'd use:

• Admin, then Data display, then Channel groups, then Create new channel group. Name it something like "Channels with AI".
• Click Add new channel and name it "AI (all)". Add one condition: Session source matches regex, using the pattern in the next section.
• Add a second condition group with OR: Session medium exactly matches "ai-assistant". This keeps everything Google already recognizes.
• Click Reorder and drag "AI (all)" above Direct, Organic Search and Referral, so it claims AI sessions before they fall into anything else.
• Save, then pick the new group from the dimension drop-down in Traffic acquisition ("Session primary channel group (Channels with AI)").

Placing it above Organic Search is a choice, not a requirement. Google's AI Overviews and AI Mode clicks come from google.com either way, and nothing in the source value separates them from a normal search click, so moving AI above Organic Search won't steal those visits. It does guarantee that a source like "gemini" can't be swept into some other rule further up the list.

Custom groups apply to past dates in reports, which is the main reason to build one even after Google's channel arrived. If you want to compare AI clicks this quarter with the same quarter last year, the default AI Assistant channel can't give you that, because it didn't exist yet and earlier AI sessions carry the medium "referral".

Which regex catches AI traffic in GA4?

Use a pattern that matches the whole source value, because GA4's "matches regex" condition is a full match. This one covers the assistants and source names we've seen or verified, and it passed every test case below:

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

Google's regex help page for Analytics spells out the trap with its own example. City matches regex "San Francisco" returns only "San Francisco", while ".*San Francisco.*" also returns "South San Francisco". Applied to AI sources, a pattern like chatgpt|perplexity|claude matches only a source that is exactly "chatgpt", "perplexity" or "claude". It misses "chatgpt.com", "perplexity.ai" and "claude.ai", which are the values GA4 actually stores. Many copied patterns fail silently this way: the report looks empty, and people conclude they have no AI traffic.

The opposite fix, wrapping everything in .*, overshoots. .*gemini.* also matches a jewelry store called geminiwatches.com, and .*claude.* matches a restaurant site. That's why the pattern above lists each domain whole. We tested it in Python with full-match semantics against 16 source values, 11 that should match and 5 that shouldn't.

Our pattern matched all 11 AI sources and none of the 5 decoys. The short pattern matched 2 of 11. The domains in it are the current product addresses; we checked each one, and chat.openai.com now redirects to chatgpt.com, bard.google.com to gemini.google.com, and perplexity.ai to www.perplexity.ai. GA4 tends to store Perplexity without the "www", which is what we saw, but the optional group covers both. Leave openai.com out on purpose: visits from OpenAI's documentation or blog aren't answer clicks.

GA4 uses RE2 syntax, which doesn't support lookaheads or backreferences. The pattern above uses only groups, alternation and the optional "?", so it works in channel groups, report filters, explorations and Looker Studio.

How each AI engine shows up in GA4

Each assistant leaves a different trail, and the default channel catches some better than others. ChatGPT, Gemini and Perplexity reached the AI Assistant channel in our property; Google says Copilot, DeepSeek and Grok are covered too; Google's own AI Overviews and AI Mode never are.

Microsoft's own analytics tool draws the same line. Microsoft Clarity's AI channel groups, AIPlatform and PaidAIPlatform, cover Claude, Copilot, Gemini, ChatGPT and Perplexity, and the documentation says the numbers reflect use of those tools on their standalone websites. AI features built into search engines or office tools aren't included. If you run Clarity next to GA4, it's a useful second count of the same visits.

What can't GA4 see in AI search?

GA4 can't separate clicks from Google's AI Overviews or AI Mode from regular Google clicks, can't see AI mentions that never produce a click, and loses any visit that arrives with no referrer and no UTM tag. For most sites, those blind spots are larger than the AI Assistant channel itself.

Start with Google's AI features. Google's guidance on AI features and your website says sites appearing in AI Overviews and AI Mode are included in the overall search traffic in Search Console, in the Performance report under the "Web" search type. The Performance report has no filter for them, and Search Console's separate generative AI report shows impressions, not clicks. In GA4, those clicks arrive from google.com and land in Organic Search with every other Google click. Google's same page says clicks from results pages with AI Overviews tend to be "higher quality", with visitors spending more time on the site, but it doesn't give you a way to isolate them.

Then the mentions without clicks. When Pew Research Center tracked the browsing of 900 US adults in March 2025, users clicked a regular result on 8% of visits to a Google page with an AI summary, against 15% on pages without one. They clicked a link inside the AI summary on just 1% of visits. The same gap applies to standalone assistants: a ChatGPT answer that recommends you, without a link the reader clicks, is invisible to GA4.

Then the lost referrers. A referrer is what the browser sends to tell your site where a visitor came from, and modern browsers trim it by default. MDN's Referrer-Policy reference notes that the default policy, strict-origin-when-cross-origin, sends only the origin for cross-site requests, and nothing at all from HTTPS to HTTP. AI apps add their own gaps: links opened from a desktop or mobile app, copied into a new tab or shared in a message often arrive with no referrer. Without a referrer or a UTM tag, GA4 has nothing to classify, and the session is recorded as Direct.

The October 2026 ChatGPT bug showed how big that gap can get. For about a week after a model update on October 7, links clicked in ChatGPT opened with no utm_source tag and no referrer. Search Engine Roundtable reported Adobe's Cedric Huesler counting roughly 31,000 fewer ChatGPT visits a day across about 500 pages for October 9 to 13, while direct traffic on the same pages rose by about 26,000 a day. On October 15 he called it a bug and said ChatGPT was tagging links again. We logged it in our AI search news at the time. If your ChatGPT numbers dipped and Direct rose in that window, that's tracking, not demand.

Finally, the crawlers and agents. Fetches by GPTBot, OAI-SearchBot, ChatGPT-User, PerplexityBot and ClaudeBot happen on your server and never run the GA4 tag. Agentic browsers that load pages on a user's behalf may run it, and when they do, there's no reliable way in GA4 to tell an agent's session from a human's. Server logs or your CDN's bot reports are the only place to count those.

How to build an AI traffic exploration with landing pages and key events

Build a free-form exploration with Session source / medium, Landing page and your key events, filtered to AI sources. It answers the questions the standard report can't: which pages AI assistants send people to, and whether those visitors do anything once they arrive.

• Explore, then Blank. Under Dimensions, import "Session source / medium", "Session primary channel group (Default Channel Group)", "Landing page + query string" and "Date".
• Under Metrics, import "Sessions", "Engaged sessions", "Engagement rate", "Key events" and "Session key event rate".
• Drag "Landing page + query string" into Rows and "Session primary channel group" into Columns. Add the metrics to Values.
• Add a filter: Session source matches regex, using the pattern above. Or filter on Session medium exactly matches "ai-assistant" if you only want what Google already classifies.
• Switch the visualization to a line chart with "Date" as the dimension to see whether AI sessions are rising, flat or tied to one page.

Keep the "query string" version of the landing page dimension. ChatGPT's utm_source tag is part of the landing URL, so you can spot tagged visits directly in this view, and you'll see when an assistant links to a parameter version of a page you'd rather it didn't.

The landing page view is usually the most useful output of the whole exercise. On most sites AI assistants link to a handful of pages: a pricing page, a comparison, a statistics page, a definition. Those are the pages AI engines treat as citable, and they're where a stronger answer, a clearer table or a better call to action will pay off first. In our study of 742 AI Overview citations, software vendors' own pages earned 44.1% of citations, so it's worth knowing which of yours already get the clicks.

How much traffic do AI assistants send, and which ones?

Not much yet, compared with Google, but it's growing fast and spreading across more assistants. Similarweb counted about 1.13 billion AI referrals to the top 1,000 websites in June 2025, up 357% from a year earlier, while Google Search sent 191 billion. ChatGPT's share of B2B AI referrals fell from 89.1% to 62.6% between mid-2025 and spring 2026 in one panel.

The Similarweb figures were reported by TechCrunch in July 2025. They cover the top 1,000 sites globally, so they describe the biggest destinations on the web, not a typical B2B site. Retail shows the same direction with sharper numbers: Adobe Analytics measured a 693.4% increase in generative AI traffic to US retail sites over the 2025 holiday season (November 1 to December 31) compared with the year before.

The engine mix matters more for GA4 setup than the totals, because it tells you which source values you'll actually see. Goodie's AI search traffic report measured 2,802,519 AI referral sessions across 41 mostly B2B brand sites from May to August 2025, then ran a second wave on a fresh panel for March and April 2026.

Goodie notes that the second-wave shares are averages of each brand's share, so they sum to 103.0%, not 100%. Read them as each engine's typical weight on a B2B site, not slices of one pie. Our own split looks different (no Claude at all, a larger Gemini share), which says more about who reads a GEO publication than about the market. Your own source / medium table is the only mix that matters for your reporting.

The Claude jump is the practical point. A regex written in 2024 for "chatgpt" and maybe "perplexity" would miss a sixth of B2B AI referrals in Goodie's spring 2026 panel. The default AI Assistant channel covers Claude, and the pattern above does too.

How to cross-check GA4 with Search Console, Bing and Clarity

Use each tool for the slice it can actually see. Search Console shows Google search clicks with AI Overviews and AI Mode blended in; Bing Webmaster Tools counts how often Copilot and Bing's AI summaries cite your pages; Clarity gives a second count of assistant clicks. Together they cover more than GA4 alone.

Search Console first. Because AI Overviews and AI Mode are counted inside the Web search type, a page whose clicks fall while its impressions and position hold steady may be losing clicks to an AI Overview. That's a pattern to investigate, not proof. Compare the page's Search Console clicks with its GA4 Organic Search sessions over the same 28 days; if both drop together, the cause is upstream of your site.

Bing Webmaster Tools reports AI citations directly. Its AI Performance report counts citations of your pages in Copilot, AI summaries in Bing and select partners, and in October 2026 Bing added a Topics Portfolio beta that groups the grounding queries where your site appears into themes. Citations aren't clicks, but they're the closest free view of the exposure GA4 never sees.

Clarity, as covered above, labels AI platform sessions on its own and is free. If GA4 and Clarity disagree by a wide margin for the same month, check consent settings first: GA4 under consent mode can model behavior for users who decline analytics cookies, and Clarity handles consent its own way.

Worked example: a monthly AI traffic line for a B2B site

This example is hypothetical, with round numbers chosen to show the arithmetic. Your own report should use the same steps with your own data.

A B2B software site reviews the last 30 days. The default channel report shows 12,000 Organic Search sessions with 180 demo requests (the key event), and 480 AI Assistant sessions with 12 demo requests. The custom channel group adds 40 more AI sessions that arrived as "chatgpt.com / (not set)" and "perplexity / (not set)", with no demo requests among them.

• AI sessions as a share of Organic Search: 480 ÷ 12,000 = 4.0% from the default channel, or 520 ÷ 12,000 = 4.3% with the custom group.
• Demo request rate: AI Assistant 12 ÷ 480 = 2.5%; Organic Search 180 ÷ 12,000 = 1.5%.
• AI share of demo requests across the two channels: 12 ÷ (12 + 180) = 6.25%, about one in sixteen.

Written as a report line: "AI assistants sent 520 sessions (4.3% of organic search volume) and 12 demo requests (6.25% of the two channels' total), converting at 2.5% against 1.5% for search." That's one sentence an executive can read, and every number in it can be traced to a GA4 table. With 12 conversions, the rate difference could easily be noise, so the report should say so rather than celebrate it. Track it for three months before you change budgets around it. Bigger samples point the same way on volume: our AI SEO agency guide sets three recent traffic studies against what a retainer costs.

The same template works on a small site. In our own 28 days, 86 AI Assistant sessions against 1,470 from Organic Search is 5.9%, and with no key events configured there's no conversion line to write, which is itself a finding: you can't judge AI traffic's value without a key event to measure it against.

Mistakes that make AI traffic look smaller or bigger than it is

The common errors all come from treating one GA4 view as the full picture. Each one below either hides real AI visits or inflates them.

• Filtering only the Referral channel. Since May 13, 2026, recognized assistants carry the medium "ai-assistant", so a Referral-only report now shows mostly old sessions and misses current ones.
• Copying a regex without testing it. Because GA4 regex conditions are full-match, short patterns like chatgpt|claude silently match nothing. Test any pattern against your real source values before you trust an empty report.
• Wrapping every term in wildcards. .*gemini.* and .*claude.* pull in unrelated domains that happen to contain those words.
• Putting a custom AI channel below Referral. GA4 assigns each session to the first matching channel, so a lower AI channel never sees referral-tagged AI visits.
• Counting AI Overviews as AI Assistant traffic. Google excludes AI Overviews and AI Mode from the channel, and nothing in GA4 separates them from other Google clicks.
• Comparing AI with Direct during a bot surge. Automated traffic usually lands in Direct, which distorts any share-of-total figure. Compare AI with Organic Search instead.
• Reading short date ranges. AI sessions are small and lumpy on most sites; a single shared link can double a week. Use 28 days or more.
• Calling every Direct increase "dark AI traffic". Some of it is, as the October ChatGPT bug showed, but Direct also absorbs bookmarks, untagged email and bots. Treat it as a question, not an answer.

For definitions of the terms used here, our glossary entry on AI referral traffic is the short version.

Frequently asked questions