Ask most marketing teams how much traffic they get from ChatGPT, Perplexity or Gemini and you’ll get a shrug — not because the traffic isn’t there, but because the default analytics setup was never built to see it. AI assistants send visitors in ways that break the assumptions Google Analytics 4 was designed around, and the result is that AI-driven demand quietly leaks into buckets labelled Referral, Direct or Unassigned. If you want to measure AI referral traffic in GA4 properly, you need to layer several methods together. Here’s the methodology we use ourselves, described so you can replicate it on your own stack.

Why AI referral traffic is nearly invisible in GA4

Three separate mechanisms hide AI-driven visits, and they compound.

1. Many AI answers end with no click at all

When an assistant answers a question in full — including recommending your product by name — the user often never clicks through. There is no session, no pageview, nothing for any web analytics tool to record. The influence is real; the measurable event is absent. Users who were convinced by an AI answer frequently arrive later by typing your URL or searching your brand name, which lands them in Direct or Organic Search. We’ve written before about why zero-click AI search still builds brand visibility even when your analytics shows nothing — the measurement problem doesn’t make the channel imaginary, it just makes naive dashboards misleading.

2. Assistant apps strip or omit referrers

Even when a user does click, the referrer header often doesn’t survive the trip. Mobile assistant apps, in-app browsers and copy-pasted links routinely arrive with no referrer at all, and GA4 classifies a session with no referrer and no UTM parameters as Direct. Google’s own documentation for the newer AI Assistants channel acknowledges the boundary: classification depends on the referrer, so AI visits arriving without one still land in Direct. A meaningful slice of your “Direct” traffic is AI referral traffic wearing a disguise.

3. GA4’s default grouping buckets AI referrers generically

Until recently, a session that did arrive with chatgpt.com as its source was simply classed as Referral, mixed in with links from blogs, news sites and directories. Unless someone drilled into source/medium and knew which hostnames to look for, AI traffic was invisible at the channel level — the level where budget conversations actually happen.

The GA4 fix: a custom channel group with a source regex

The standard remedy is a custom channel group. In GA4, go to Admin → Data display → Channel groups, create a new group (or edit a copy of the default), and add a channel — call it something like “AI Assistants” — with a condition on session source matching a regex such as:

chatgpt\.com|chat\.openai\.com|perplexity\.ai|claude\.ai|gemini\.google\.com|copilot\.microsoft\.com|meta\.ai

Extend the pattern as new assistants appear — the list is a moving target, and reviewing your Referral source report monthly for unfamiliar AI hostnames is worth the five minutes.

The ordering gotcha that silently breaks it

Here’s the mistake that wastes the most hours: GA4 evaluates channel conditions in order, top to bottom, first match wins. If your custom AI channel sits below Referral in the list, every AI session matches Referral first and your shiny new channel records zero — forever, with no warning. Drag your AI channel above Referral before saving. If your custom channel is showing nothing despite visible AI sources in source/medium reports, ordering is almost always the culprit.

Two more caveats: custom channel groups apply from creation onwards in some reporting surfaces, so build yours early rather than waiting until someone asks; and the channel only catches sessions that arrive with a recognisable referrer — the Direct-leakage problem above still applies.

GA4’s native AI Assistants channel

Google has now added a native answer to part of this problem. GA4’s Default Channel Group documentation at support.google.com lists an AI Assistants channel, defined as traffic from “sources like ChatGPT, Gemini, Deepseek, Copilot, or Grok”, excluding Google’s AI Overviews and AI Mode. When GA4 recognises an AI assistant referrer, it sets the medium to ai-assistant and classifies the session automatically — no configuration needed. Search Engine Journal reported the rollout in May 2026.

So is the custom channel group obsolete? Not yet. Notice what Google’s documented source list omits: Perplexity — a high-intent AI referrer for many B2B sites — is not named, and Google hasn’t published the full referrer list. The native channel is also referrer-dependent, so the Direct leakage remains, and it can’t reclassify history from before it existed. Our advice: keep the custom channel group running in parallel. It’s your controllable, extensible layer; the native channel is the convenient baseline.

The server-log layer: watching the engines read you

Everything above measures humans arriving after an AI answer. There’s an earlier signal GA4 can never see: the AI engines themselves fetching your pages to build those answers. These fetchers don’t execute JavaScript and don’t carry cookies, so they’re invisible to GA4 by design — but they’re sitting in plain sight in your server access logs.

The user agents worth counting, per each operator’s own documentation:

• ChatGPT-User — OpenAI’s on-demand fetcher, sent when a ChatGPT user’s request causes a live page visit (OpenAI’s bot documentation). This is the closest thing to “an AI is reading my page for a user right now”.
• OAI-SearchBot — OpenAI’s crawler for surfacing sites in ChatGPT’s search features, distinct from GPTBot, which crawls for model training.
• PerplexityBot and Perplexity-User — per Perplexity’s crawler documentation, the former surfaces and links sites in Perplexity results, and the latter fetches pages when a user’s question triggers a live visit.

The counting method is deliberately boring. On an nginx box, something like:

grep -Ec 'ChatGPT-User|OAI-SearchBot|PerplexityBot|Perplexity-User' /var/log/nginx/access.log

run daily (split by UA and by path, and filtered to successful 2xx responses so you’re not counting blocked or broken requests) gives you a per-page, per-engine read on which content the answer engines are actually consuming. Rising on-demand fetches to a page are a leading indicator: engines are pulling that page into answers before any referral traffic shows up in GA4. In our experience running this as part of an answer engine optimisation programme for SaaS, log-level fetch counts move first, citations second, referral sessions third.

Two hygiene notes. First, verify suspicious traffic: plenty of scrapers spoof browser user agents, and some spoof bot user agents, so filter by response status and — if it matters for a decision — check requests against the IP ranges each operator publishes. Second, keep bot counts and human counts in separate columns forever. The moment a dashboard adds fetcher hits to human sessions, every number downstream is fiction.

Why bother: the traffic is small but disproportionately valuable

The best public first-party data here comes from Ahrefs. Patrick Stox, then product advisor at Ahrefs, published figures from Ahrefs’ own site, measured in Ahrefs Web Analytics: AI search was roughly 0.5% of visits over the 30 days studied, yet drove roughly 12.1% of signups — a conversion rate about 23 times higher than traditional organic search visitors, with most of the AI traffic coming from ChatGPT. That’s one company’s first-party data, not a universal law, but the direction matches what many teams see: a visitor who arrives after an AI assistant has already explained, compared and effectively pre-sold your product behaves like a warm referral, not a cold click. We unpack the mechanics of that in why AI search visitors convert better.

The practical implication: if you evaluate AI channels on session volume, you’ll conclude they’re a rounding error and underinvest. Evaluate them on conversions per session and pipeline contribution, and the picture inverts.

Comparison: what each measurement layer sees and misses

Layer	What it sees	What it misses
GA4 default setup	AI sessions with an intact referrer, but buried in Referral (or auto-classed by the newer native AI Assistants channel where the source is on Google’s list)	Zero-click influence; referrer-stripped visits (land in Direct); sources not on Google’s list, e.g. Perplexity; anything pre-dating the native channel
GA4 custom channel group (regex on source)	Any AI referrer you name in the regex, surfaced as its own channel with conversion and engagement metrics	Zero-click influence; referrer-stripped visits; new AI sources until you add them; nothing if the channel is ordered below Referral
Server logs (bot user agents)	AI engines fetching your pages (ChatGPT-User, OAI-SearchBot, PerplexityBot) — a leading indicator, per page, needing no JavaScript or cookies	Human visits and conversions entirely; whether a fetch became a citation; spoofed user agents unless you verify IPs
Referrer survey (“How did you hear about us?”)	Zero-click and referrer-stripped influence — the only layer that captures “ChatGPT recommended you” when no measurable click ever happened	Everyone who doesn’t answer; recall error; no volume or trend precision — directional, not quantitative

No single layer is sufficient. GA4 custom channels give you trend and conversion data for clicked visits; server logs tell you which pages engines are reading before clicks exist; a one-question survey on your signup form is the only net that catches zero-click influence. Run all three and triangulate.

FAQ

Does GA4 track AI referral traffic by default?

Partially, and only recently. Google’s Default Channel Group documentation now includes a native AI Assistants channel covering sources like ChatGPT, Gemini, Deepseek, Copilot and Grok, classified automatically via the referrer. But it excludes Google’s own AI Overviews and AI Mode, doesn’t name Perplexity, can’t reclassify historical sessions, and any AI visit arriving without a referrer still lands in Direct. A custom channel group with your own source regex remains worthwhile.

Do visitors from AI assistants really convert better?

The strongest public first-party evidence says yes. In Ahrefs’ study of its own site data, published by Patrick Stox using Ahrefs Web Analytics, AI search accounted for about 0.5% of visits over 30 days but about 12.1% of signups — roughly a 23x higher conversion rate than traditional organic search. It’s one company’s data, but the pattern is consistent with AI answers pre-qualifying visitors before they ever click.

Why does my custom AI channel in GA4 show zero traffic?

Almost always ordering. GA4 evaluates channel group conditions top to bottom and stops at the first match. If your AI channel sits below Referral, AI sessions match Referral first and your custom channel never fires. Move it above Referral. If it still shows nothing, confirm the sources actually appear in your session source/medium report and that your regex matches them exactly.

Which user agents indicate AI engines are reading my site?

The main on-demand fetchers are ChatGPT-User (OpenAI, fired when a user’s ChatGPT request triggers a live page visit), OAI-SearchBot (OpenAI’s search-surfacing crawler) and PerplexityBot plus Perplexity-User (Perplexity’s search and user-request fetchers). Count them in your server access logs with simple user-agent matching, filtered to successful responses, and track per page over time as a leading indicator of AI answer inclusion.

Can server logs replace GA4 for measuring AI traffic?

No — they answer a different question. Server logs count AI engines fetching your pages, which happens before and independently of any human visit; they contain no conversion or engagement data. GA4 counts humans who clicked through, but only when a referrer survives. You need both, plus a “how did you hear about us” field to catch zero-click influence neither tool can see.