How to Track AI Referral Traffic and Leads in GA4

Every guide on this topic ends at the same place: set Session source to chatgpt.com, read the number, feel something. That number is both wrong and useless. Wrong because GA4 systematically undercounts AI traffic. Useless because you do not sell sessions — you sell booked calls.

This is the setup that fixes both problems: the channel group, the Direct-traffic leak, and the hidden form field that carries the AI source into your CRM.

Why doesn't AI traffic show up properly in GA4 by default?

Because GA4 splits AI traffic across three buckets, and only one of them is labeled. Google's default channel group now includes an AI Assistants channel — but Google's documentation states it covers sources like ChatGPT, Gemini, Deepseek, Copilot and Grok and that it excludes Google's AI Overviews and AI Mode (Google Analytics Help).

Those AI Overview and AI Mode clicks are filed under Organic Search instead, indistinguishable from a blue-link click. Google Search Console does the same thing: its own docs confirm AI features are reported inside the overall 'Web' search type, with no separate breakout (Google Search Central).

So the three buckets are: AI Assistants (chatbots that pass a referrer), Organic Search (everything Google's AI produced), and Direct (every assistant that sent no referrer at all). Any number you report from one bucket alone is a fragment.

Which referrer hostnames do you need to group as AI traffic?

Ten hostnames cover effectively all trackable AI referrals today. Ahrefs studied 3,000 sites and found ChatGPT sent 50% of AI traffic, Perplexity just shy of a third and Gemini about 18% — 98% of AI traffic came from those three chatbots, with Claude, Copilot, Mistral and Jasper passing on less than 2% between them (Ahrefs, February 2025).

That is your priority order. Track the rest because they cost nothing to add, not because they will move your numbers.

Note what is missing from that list: google.com. Google's AI Mode does not send a distinct hostname. That traffic is Organic Search and it is not coming back — see the GEO complete guide for what to do about visibility you cannot count directly.

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

It takes about ten minutes in Admin > Data display > Channel groups, and standard GA4 properties get 2 custom channel groups with up to 50 channels each (Google Analytics Help). Copy the default group, add a channel called 'AI assistants', and set Source matches regex against a hostname-anchored pattern.

Use this: chatgpt\.com|chat\.openai\.com|perplexity\.ai|gemini\.google\.com|copilot\.microsoft\.com|claude\.ai|grok\.com|deepseek\.com|you\.com|mistral\.ai

Google publishes its own example regex on that same help page, and it opens with a fragment that matches any source containing the letters 'ai' anywhere in the string. That includes mail.google.com. Anchor on hostnames instead of loose letter matches, or you will book your email traffic as AI wins.

Two settings people miss. First, channel order matters: traffic is assigned to the first channel whose definition it matches, so drag 'AI assistants' above Referral. Second, custom channel groups are not available in the BigQuery export schema — if you warehouse your data, replicate the regex there too.

Why does ChatGPT traffic keep landing in Direct?

Because a browser only reports a referrer when the sending page tells it to, and several AI clients tell it not to. Ahrefs ran repeat click tests from Copilot and watched every visit register as Direct — same result for Mistral and Jasper — concluding that AI traffic is 'likely being vastly underestimated' (Ahrefs, February 2025).

GA4's own documentation lists the mechanics: missing referral information, redirects that strip parameters, URL shorteners, and ad blockers all dump sessions into (direct) / (none) (Google Analytics Help). A native app that opens a link in an in-app browser is the same class of problem.

You cannot fix this in GA4. You can only detect it: watch for Direct-traffic spikes on individual deep pages that nobody types from memory. A surge of Direct sessions on a 2,000-word service post is not somebody bookmarking you — it is an assistant that swallowed the referrer.

How do you capture the AI source into your CRM as a lead, not a session?

Read the referrer on the first landing, store it in a first-party cookie for 90 days, and write it into a hidden field on every form. That single hidden field turns an unattributable session into a CRM record you can close revenue against — and it survives the multi-page, multi-day path a $2,500 decision actually takes.

Add one plain-text 'How did you hear about us?' field as a backstop for the referrer-stripped share. It is the only instrument that catches the visitor who read your page inside ChatGPT, closed the app, and typed your name into Google two days later. Ahrefs reports collecting 14,000+ self-attributed new users from ChatGPT this way.

Then mark the form submit and the booked call as GA4 key events, not just page views. If your analytics stack cannot survive this, that is a stack problem — we cover the tradeoffs in GA4 vs Plausible vs Fathom.

How much AI referral traffic should a service business expect?

Small numbers. Across Ahrefs' 3,000-site sample, 0.17% of the average site's visitors came from AI chatbots, and 63% of sites received at least one AI visit (Ahrefs, February 2025). That study is dated, and Ahrefs itself argues the true figure is higher because of the Direct leak — but the order of magnitude is the point.

The encouraging cut: in that same study, the smallest sites (under 999 monthly visitors) took the greatest share of their total traffic from AI — proportionally more than the big sites did. Low traffic does not mean no AI traffic.

So if you run 4,000 sessions a month, single-digit AI sessions is a normal reading, not a broken tag. Do not let an agency sell you a GEO retainer on a promise of AI traffic volume. Judge it on leads. That is the same 90-day kill switch logic we apply to every channel: no qualified leads in 90 days, the channel gets cut.

Does AI traffic convert better or worse than organic search?

The best public number says better — Semrush found the average AI search visitor was 4.4 times as valuable as the average traditional organic search visitor, based on conversion rate (Semrush, 2025). Their explanation is that the assistant already did the comparison shopping, so the person landing on you is late-funnel.

Read the scope before you quote it: Semrush measured over 500 digital-marketing and SEO topics, tracked to non-Google AI sources. That is a B2B software audience, not a plumbing customer. Treat 4.4x as a directional signal that AI clicks are qualified, not as a number you can put in your own forecast.

Which is exactly why you build the CRM field first. Once you have 90 days of lead_source_detail data, you have your own conversion rate for AI traffic — and yours beats anyone's benchmark.

What can server logs tell you that GA4 cannot?

Two things GA4 structurally cannot see: requests from visitors who blocked the tag, and the AI crawlers themselves. GA4 fires from JavaScript. Bots do not run it. Your access logs record every request regardless.

OpenAI publishes distinct user agents for distinct jobs: GPTBot for model training, OAI-SearchBot for ChatGPT search indexing, and ChatGPT-User for a fetch triggered when a user asks ChatGPT a question (OpenAI). A ChatGPT-User hit in your logs means an assistant pulled your page into a live answer — a citation signal that will never appear in GA4.

Grep your logs weekly for those agents plus PerplexityBot and ClaudeBot. Which pages get fetched is the leading indicator; referral sessions are the lagging one. This is the downstream half of measurement — the upstream half (am I being mentioned at all?) is covered in measuring AI search visibility on a budget.

If your GA4 is reporting AI traffic you cannot trace to a single booked call, the tracking is not the deliverable — the pipeline is. We build the channel group, the CRM field, and the reporting that ends in revenue, on month-to-month terms with no lock-in. Get my free audit and we will tell you what your current setup is losing.