How to Track AI Traffic in GA4 (When ChatGPT Shows Up as “Direct”)

Updated: July 2026. If you open Google Analytics 4 and look for your ChatGPT, Perplexity, or Claude visitors, most of them aren’t where you expect. They’re hiding in Direct. Google added a native “AI Assistant” channel in May 2026, and it’s a genuine step forward — but it quietly misses a large slice of your AI traffic. This guide explains exactly why your GA4 undercounts AI visits, walks through the fix (a custom AI channel group you can build in ten minutes), shows you how to surface the AI sessions already buried in Direct, and closes with the harder truth: GA4 measures the click, while the thing that actually moved — the citation — often produces no click at all.

This is the evergreen how-to. For the announcement of the native channel itself — the launch date, the medium, the recognized engines — see our news write-up of GA4’s AI Assistant channel. That piece covers the news; this guide is the durable method for measuring AI traffic no matter how Google’s defaults shift.

Why Your GA4 Is Undercounting AI Traffic

The root cause is the referrer header. GA4 attributes a session by reading the referring domain the browser hands over when a visitor arrives. Follow a normal link from perplexity.ai in a desktop browser and GA4 sees “perplexity.ai” and can file the session under Referral. But a huge share of AI-assistant traffic never carries that header at all.

Referrer-less arrivals are the norm for AI, not the exception. When a visitor taps a link inside the ChatGPT mobile app, an in-app browser, ChatGPT’s Atlas browser, or simply copies and pastes a URL an assistant gave them, no referring domain reaches your server. GA4 has nothing to attribute the visit to — so it drops the session into Direct, the bucket for “we don’t know where this came from.” As Search Engine Journal summarized it: AI-assistant traffic that arrives without a referrer header still lands in Direct.

The practical effect is that your real AI traffic is split across at least three channels — some in Referral, some (post–May 2026) in the new AI Assistant channel, and a stubborn remainder in Direct — while your “AI Assistant” number looks small and reassuring. Industry estimates for how much AI traffic gets under-counted range widely, roughly 35% to 70% depending on the study, the site, and the device mix. That’s a wide band on purpose: treat it as a warning that the gap is large and site-specific, not as a precise figure. The only number you can trust is the one you measure on your own property.

In summary, GA4 undercounts AI traffic because a large share of AI visits arrive without a referrer header and get filed under Direct. The native AI Assistant channel classifies only the sessions Google can positively identify, so the rest stays scattered and invisible until you go find it.

What Google’s Native “AI Assistant” Channel Does — and Doesn’t

On May 13, 2026 (with a broader rollout around June 7), Google added a dedicated default channel to GA4. In Google’s own words: “Google Analytics now provides a dedicated way to measure and analyze traffic originating from popular AI assistants.” Under the hood it introduces a medium of ai-assistant, a new “AI Assistant” channel group, and a campaign value of (ai-assistant).

Per the live documentation as of June, the channel recognizes ChatGPT, Gemini, Deepseek, Copilot, and Grok. Notably, Perplexity is absent from that list and continues to show up under Referral — though this specific exclusion isn’t confirmed on a raw Google source, so treat it as provisional rather than settled. Either way, it’s a reminder that the native channel’s coverage is a moving target you shouldn’t depend on for a complete count.

The channel is real progress, but it has three holes you need to know about:

• It misses referrer-less traffic. Visits from mobile apps, in-app browsers, Atlas, and copy-paste have no referrer, so they fall into Direct — not AI Assistant. This is the single biggest source of undercounting.
• It ignores Google’s own AI surfaces. Traffic from AI Overviews and AI Mode is routed to Organic Search, not AI Assistant, because it originates on a Google search results page. Your Google-side AI visibility stays folded into organic.
• It isn’t retroactive. The channel only classifies traffic going forward. Every AI visit from before launch stays mislabeled — so year-over-year AI comparisons built on the native channel are broken by design.

In summary, the native AI Assistant channel is worth turning on and worth watching, but it is a floor, not a full count. It catches the cleanly-referred sessions and leaves the referrer-less majority, the Google-surface AI traffic, and all of your history unaccounted for.

Build a Custom AI Channel Group (the Fix)

The fix is a custom channel group that classifies AI traffic on your terms instead of waiting for Google’s defaults. It takes about ten minutes in Admin → Data display → Channel groups → Create new channel group. Three rules make it work:

1. Match on the Source domain with a regex

Create a channel called “AI Assistants” with a single condition: Source matches regex. Use a pattern that covers the assistant domains you care about:

Matching on the Source dimension catches referral-tagged AI sessions regardless of whether Google’s native rule fired for them. Extend the list as new assistants appear — this is a pattern you own and can edit, unlike the native channel.

2. Do NOT constrain the medium

This is the step most people get wrong. It’s tempting to add “and Medium = ai-assistant,” but AI referrals are medium-inconsistent. OpenAI stamps utm_source=chatgpt.com on desktop only, and a large share of AI sessions arrive as “chatgpt.com / (not set)” with no usable medium at all. If you require a specific medium, you re-inherit the native channel’s blind spots. Match the Source domain and leave the medium open.

3. Put the AI rule at the top of the group

Channel-rule order is not cosmetic. As MO Agency states plainly: “GA4 evaluates channel rules top to bottom — the first matching rule wins.” Because some AI domains would otherwise be swept up by a broad Referral or Organic rule, your AI Assistants rule must sit above them in the group. If it sits below Referral, an AI session that carries a referrer gets claimed by Referral first and never reaches your AI rule.

One caveat worth setting: a custom channel group reclassifies going forward and, in Explorations, can be applied to historical data — but it will never recover the referrer that was never sent. It moves correctly-referred AI sessions out of Referral and into your AI channel; it cannot pull referrer-less sessions out of Direct. That job belongs to the next section.

In summary, build a custom channel group that matches AI-assistant domains on the Source dimension, leaves the medium unconstrained, and sits at the top of the rule order. That single group recovers the referral-tagged AI traffic the native channel scatters — but not the referrer-less traffic hiding in Direct.

Find Your Hidden AI Traffic

The referrer-less AI sessions in Direct can’t be re-attributed — but they can be estimated with a “Direct-shadow” exploration. The logic: genuine Direct traffic is people who typed your URL or used a bookmark, so it clusters on your homepage and high-intent pages. AI referrals behave differently — they land deep, on the specific answer page the assistant pointed to. That behavioral fingerprint is how you isolate them.

Build the exploration like this:

• Open Explore → Free-form and add Session default channel group as a filter set to Direct.
• Break down by Landing page and Device category.
• Exclude your homepage and your direct-intent paths — /pricing, /contact, /login — which are where legitimate Direct traffic concentrates.
• Focus on deep content and answer pages that suddenly show Direct sessions with real engagement (long engagement time, scroll, conversions). Direct traffic landing mid-site on a niche article is the signature of a referrer-less AI referral.

You won’t get a perfect count — that’s the point of calling it a shadow. But comparing the “deep Direct” trend on your content pages before and after AI assistants started sending traffic gives you a defensible directional estimate of the AI sessions the native channel and even your custom group can’t catch. Pair it with the utm_source=chatgpt.com desktop signal to sanity-check magnitude.

In summary, surface hidden AI traffic by isolating Direct sessions that land deep on content pages with strong engagement, after excluding the homepage and direct-intent paths. It’s an estimate, not a re-attribution — but it turns the Direct black box into a defensible read on your true AI referral volume.

Beyond the Click: Measure Citations, Not Sessions

Here’s the ceiling on all of the above: GA4 can only ever measure the click. And AI search is systematically removing the click. When an assistant answers the question in-line and cites you as a source, the user often gets what they need without visiting — you shaped the answer, you earned the citation, and GA4 records nothing. This is the zero-click reality of AI search: visibility and traffic have decoupled.

The context is worth stating so you calibrate expectations. AI referral traffic is still a small slice of the total — Conductor’s measurement across 13,770 domains and 3.3 billion sessions put it at 1.08% of total traffic, growing roughly 1% per month, with ChatGPT at 87.4% of the AI referral share. Similarweb, measuring differently, puts global AI traffic at 0.15–0.25% of the total. The clicks are real, they convert well — Similarweb’s panel found ChatGPT referrals convert at 7.1%, and ChatGPT referrals jumped +157.7% week-over-week right after OpenAI added clickable links on May 7, 2026 — but they are a rounding error next to the volume of answers where you’re cited and no one clicks at all.

That’s the gap GA4 can’t close. If you optimize purely on AI sessions, you’ll under-invest in the surfaces where your influence is largest and your click count is smallest. The complete picture needs two instruments: GA4 for what AI clicks convert to, and a citation-readiness view for whether the engines are citing you when they answer — the core of GEO. One measures traffic; the other measures the visibility that never becomes traffic.

In summary, GA4 is necessary but not sufficient. Track AI traffic in GA4 with a custom channel group and a Direct-shadow exploration to see every click you can — then measure citations directly to see the far larger share of AI influence that produces no click at all.

The Verdict

Your GA4 is undercounting AI traffic, and the native “AI Assistant” channel — real as it is — only lifts the floor. Referrer-less visits fall into Direct, AI Overviews routes to Organic, and none of your history is reclassified. Fix the count with a custom channel group that matches AI-assistant domains on Source, leaves the medium open, and sits at the top of the rule order; then run a Direct-shadow exploration to estimate the referrer-less remainder. Do both and you’ll finally see your true AI click volume. But remember the ceiling: GA4 measures the click, and AI search keeps deleting the click. To see the visibility that never becomes a session, you have to measure the citation itself.

FAQ