AI referral traffic is a small share of most sites' total visits, but it converts at a noticeably higher rate than classic Google organic traffic. GA4 will not show it to you properly by default: you need a custom AI channel with a session-source regex, ordered above Referral, plus a landing-page report to see which of your pages the engines are actually sending people to. Setup takes about twenty minutes. Here is every step, and every caveat.
A recommendation you cannot measure is a rumor. This is the plumbing that turns "ChatGPT mentioned me" into a number your accountant would accept.
Why track a stream that is only 1% of traffic?
Because a small AI referral stream is not like the rest of your traffic. AI referral traffic is a modest share of total visits for most sites, but the conversion behavior of that slice looks meaningfully better than an average organic visitor, and industry trackers that publish AI-versus-organic conversion comparisons consistently find the same pattern even though their exact figures differ by methodology. A visitor who arrives from an AI answer was not browsing, they were referred. The engine already framed you as the answer before the click happened.
Concentration makes the measurement problem tractable: ChatGPT is by a wide margin the largest single source of AI referral traffic for most sites, reflecting how large its active user base has grown relative to other assistants. Get ChatGPT attribution right and you have captured most of the story. And for anyone doing PEO the stream is even more personal: these clicks are frequently someone checking out a name a machine just gave them, which is why the pages they land on are usually your About page, your knowledge base, or the exact article that made you citable. Which of your questions deserve this treatment is the domain of Money Queries.
How does AI traffic actually show up in GA4?
Three arrival patterns, three very different visibilities:
- Tagged referrals. ChatGPT appends a utm_source=chatgpt.com parameter to many outbound links from its search-backed answers, so those sessions arrive cleanly labeled with session source chatgpt.com. This is the best-behaved slice of the stream.
- Plain referrals. Perplexity, Gemini, Copilot and most others usually pass an ordinary referrer header (perplexity.ai, gemini.google.com, copilot.microsoft.com) with no UTM tagging. GA4 files these under the generic Referral channel, mixed in with every blog that ever linked you.
- Dark arrivals. Clicks from mobile apps and some desktop clients ship no referrer at all. GA4 calls them Direct. This slice is real, material and invisible, which means every number your AI channel shows you is a floor, not a total.
GA4's default grouping has been catching up, and some AI sources now land in their own bucket on newer properties. Do not rely on it. Build the custom channel anyway: you control the definition, the ordering and the update cadence, and you can see exactly what it catches.
The GA4 setup: a custom AI channel in six steps
- Open the channel editor. In GA4: Admin, then under Data display choose Channel groups. Do not edit the default group; click Create new channel group so the original stays intact for comparison. Google's reference for this feature is the custom channel groups documentation.
- Create the channel. Inside your new group, select Add new channel and name it AI Traffic.
- Define the condition. Set the rule to: Session source matches regex, then paste the pattern below. Session source (not medium) is where both the UTM-tagged and plain-referrer arrivals surface.
- Order it above Referral. GA4 evaluates channel rules top to bottom, and an AI session that hits the Referral definition first stays a Referral forever. Drag AI Traffic above Referral (and above Organic Search) in the channel list. This single step is the one most setups get wrong.
- Save and wait. Custom channel groups apply to how data is displayed, so your new channel appears as a selectable dimension in Traffic acquisition reports. Give it a full week of data before judging volumes.
- Annotate the start date. Note the day you shipped the channel, so nobody later mistakes the appearance of AI Traffic for the appearance of AI traffic.
The regex directory
The matching pattern, current as of July 2026:
chatgpt\.com|chat\.openai\.com|perplexity\.ai|claude\.ai|gemini\.google\.com|copilot\.microsoft\.com|grok\.com|meta\.ai|deepseek\.com|you\.com
What each cluster covers:
- chatgpt.com, chat.openai.com: ChatGPT's current and legacy domains. Expect the bulk of your volume here.
- perplexity.ai: Perplexity's answers and its Discover feed. Typically your second-largest AI source.
- claude.ai: Claude's web client, including links surfaced in web-search-backed replies.
- gemini.google.com: the Gemini app. Note that Google AI Overviews clicks are not in this bucket; they arrive as ordinary google organic and cannot be separated inside GA4.
- copilot.microsoft.com: Microsoft Copilot on web and in Windows.
- grok.com, meta.ai, deepseek.com, you.com: small today, cheap to include, occasionally spiky.
Review the list quarterly. Assistants launch, rebrand and change domains constantly, and a regex nobody maintains quietly rots into undercounting.
Which reports turn mentions into money?
Landing pages by AI channel
In Reports, open Traffic acquisition, switch the primary dimension to your new channel group, then drill into AI Traffic by landing page. This report answers the operational question: which of your pages are the engines actually citing and sending people through? If your About page and one or two cornerstone articles dominate, the engines have effectively told you what they consider your canonical proof. Feed that knowledge back into your publishing plan.
The exploration for deeper cuts
For anything the standard reports cannot answer, build a free-form exploration: Explore, new blank exploration, add Session source and Landing page as dimensions, Sessions and Key events as metrics, then filter Session source against the same regex you used for the channel. The exploration works on your historical data too, so it is also how you reconstruct what AI traffic looked like before the day you built the channel. Save it once, revisit monthly.
Key events by channel
Mark your consequential actions as key events (contact form, booking, newsletter signup) and compare completion rates for AI Traffic against Organic Search. This is where the conversion-quality claim stops being an industry statistic and becomes your number, defensible in any budget conversation.
Connecting spikes to causes
Numbers without a changelog are trivia. Keep a simple placement calendar: every article published, every podcast aired, every mention earned, with dates. When the AI channel jumps, you want to look left on the timeline and see the probable cause sitting there. This is also how you learn which kinds of proof move the needle for your name: some people find one well-transcribed podcast outdrives five guest posts, others the reverse. The correlation is never courtroom-grade, but over two or three quarters the pattern gets hard to miss, and it should redirect your effort accordingly.
The quarterly triangulation
GA4 tells you about clicks, but plenty of AI influence never clicks: the buyer hears your name, then Googles you, or types your URL. So triangulate quarterly: your AI channel trend, your branded search impressions in Search Console, and a standing prompt audit of what the engines say about your name. The prompt side of that loop is documented in The 25-Prompt Audit and the tooling landscape in AI Visibility Tools Compared. When all three lines move together after a placement or a content push, you have attribution no single dashboard could give you.
UTM parameters you control: building your own AI attribution net
Waiting for an engine to tag its own outbound links is only half the strategy, and it puts you entirely at the mercy of whether that particular assistant happens to append a UTM parameter this month. The other half is engineering your own attribution into every link you place somewhere an AI answer might pull from. When you give a podcast host a specific link to put in the show notes, or when you build a machine-readable about page or a llms.txt-referenced resource, append your own UTM parameters to it, something like utm_source=podcast-name and utm_medium=referral. If an AI answer later cites that exact page and a reader clicks through, plain referrer data alone might file the session under Referral, but a link you tagged yourself survives even if the assistant strips or ignores its own tagging.
A more aggressive version of this technique is building a dedicated landing page for content you only ever place in AI-adjacent contexts, a knowledge base entry or an FAQ page you never link from your own navigation and never promote through ordinary channels. Traffic to that page has almost nowhere else to come from, so a spike there is close to unambiguous evidence of AI-driven discovery, even before you check the session source. This same logic underlies the broader practice of building pages specifically for machine consumption, covered in Build a Knowledge Base AI Will Cite.
AI crawler hits versus AI referral clicks: two different logs entirely
It is easy to conflate two things that are actually separate systems measuring separate events. An AI crawler, the kind that fetches your pages to build a retrieval index or to satisfy a live browsing request during someone else's chat session, hits your server directly and reads the raw HTML. Most of these fetches never execute your GA4 JavaScript tag at all, because that tag only fires in a real browser rendering the page, and a crawler is not a browser. So crawler activity shows up, if anywhere, in your raw server or CDN logs, filtered for known bot user agents, and it tells you whether your content is even reachable and being read by the systems that might cite you later. A full list of the named crawlers worth watching for is in the AI Crawler Directory.
GA4's AI referral channel, by contrast, only ever captures the second event: a human being clicking a link inside an answer and landing on your site with a browser that executes your tag. A healthy pattern looks like crawler access rising first, followed weeks or months later by a rise in GA4 referral sessions, once whatever the crawler indexed starts getting cited in live answers. If your server logs show heavy AI crawler traffic but your GA4 AI channel stays flat for months, that gap itself is informative, and it usually means your content is reachable but not yet considered citable, which is a different problem to solve than an access problem.
| Signal | What it captures | Where you find it | What it misses |
|---|---|---|---|
| AI crawler access logs | Bots fetching your pages for indexing or live browsing | Raw server or CDN logs, filtered by user agent | Whether that fetch ever turns into a cited answer |
| GA4 AI referral channel | Human clicks from AI answers with a referrer or UTM | GA4 Traffic acquisition, custom channel | Unlabeled clicks that GA4 files under Direct |
| GA4 Direct traffic to AI-favored pages | A rough proxy for dark AI referrals | GA4, Traffic acquisition filtered to Direct | Cannot confirm the source was actually an AI answer |
| Branded search impressions | People who heard your name and later searched it | Search Console, Performance report | Never shows the AI mention that prompted the search |
Four different logs, four different blind spots. Use all four before drawing a conclusion from any one of them.
GA4 data retention and why a spreadsheet still matters
Custom channel groups in GA4 apply going forward and can be applied retroactively within the platform's own event data retention window, which is limited by default and easy to forget about until the day you need eighteen months of history and only have a fraction of it. If AI referral traffic is going to matter to your reporting for years, do not rely on GA4's interface alone to hold that history. Export raw events to BigQuery on a schedule, or at minimum export your AI channel's monthly numbers into a spreadsheet the same day you review them. The habit costs a few minutes a month and it is the only way to guarantee you can still show a three-year trend line when a client or a partner eventually asks for one.
The dark funnel: when the mention never produces a click at all
Some of the most valuable AI influence never touches your analytics, because it never produces a session in the first place. A buyer hears an assistant name you, forms an impression, and later mentions you in a meeting or types your company name directly into a browser without ever clicking a link an assistant generated. GA4 cannot see any of this, and no amount of channel-group engineering will make it visible, because there was no referral event to capture. This is the same dark funnel that shows up across zero-click search generally, and the mechanics and coping strategies for it are covered in Zero-Click Engineering. The practical response is not to chase a metric that does not exist, but to widen what you triangulate: branded search volume, direct traffic to your name-bearing pages, and inbound conversations where someone says they heard about you somewhere, all become weak but real evidence of influence that clicks alone will never show you.
What GA4 cannot show you (and what to do about it)
Be honest about the blind spots. GA4 cannot see the prompts people typed, cannot separate AI Overviews from ordinary Google clicks, and cannot count the dark-referrer sessions hiding in Direct. It measures the click, and the whole point of AI answers is that many recommendations convert without one.
You can still put a rough boundary around the dark slice. Pull Direct traffic by landing page and watch the pages your AI channel already favors: when Direct sessions to those same deep pages climb in step with your AI channel, you are almost certainly looking at unlabeled AI referrals, because humans do not type a long blog URL from memory. Some teams sanity-check this by comparing the ratio month over month; if labeled AI traffic doubles and deep-page Direct doubles alongside it, attribute with appropriate humility and move on.
Treat the AI channel as your floor, watch Direct for unexplained growth to AI-cited landing pages, and keep asking new leads the oldest attribution question in the business: how did you hear about me? "ChatGPT told me" is showing up in that answer far more often than it did a year ago. The broader measurement stack, beyond GA4, is covered in How to Track Your AI Visibility, and if you want the whole loop built and monitored for you, that is what our services exist for.
FAQ
Why does most of my ChatGPT traffic show up as Direct? +
What regex should I use for an AI channel in GA4? +
Is roughly 1% of traffic even worth tracking? +
Can I use UTM parameters to track AI traffic myself, instead of waiting for the assistant to tag its links? +
Do AI crawlers show up in my GA4 reports? +
What if an AI mention never produces a click at all? +
Find out what AI says about you today.
Start with a baseline. See the exact words the engines return about your name, then decide.
Claim your name →