How to Track AI Search Traffic in GA4 (ChatGPT, Perplexity, Gemini)

Summary

Traffic from ChatGPT, Perplexity, Gemini, Copilot, Claude and Mistral is already in your GA4 property, but it is spread across the referral report and partly hidden inside Direct. This guide covers the exact channel group configuration that isolates it, which assistants actually pass a referrer and which never will, why AI sessions convert far above site average despite low volume, and the structural blind spot that means referral data alone cannot tell you whether your brand is being recommended at all.

The Direct Answer: How to Isolate AI Traffic in GA4

Every marketing team eventually asks the same question: how much traffic are we actually getting from ChatGPT and its cousins? The answer is sitting in GA4 already, but it is scattered across the referral report in a way that makes it look like noise. Pulling it into one place takes about fifteen minutes.

• Open Admin, then Data display, then Channel groups, and create a new group based on the default one so you do not lose your existing channels.
• Add a new channel at the top of the list — order matters, because GA4 assigns a session to the first rule it matches.
• Name it AI Assistants and set the condition to Session source matches regex, using a pattern that covers the assistant hostnames listed in the next section.
• Save, then wait. Custom channel groups are not retroactive in the standard reports, so the data starts accumulating from the moment you save. Explorations can still segment historical sessions by source.
• Build one exploration with Session source / medium as the dimension and sessions, engaged sessions, engagement rate and conversions as metrics, filtered to your regex. This is your working view while the channel group fills up.

A regex that covers the current landscape looks roughly like this: chatgpt|openai|perplexity|gemini\.google|copilot|claude\.ai|anthropic|mistral|you\.com|phind. Keep it in a document somewhere, because it will need updating every few months as new interfaces launch and old ones change hostnames.

Which AI Referrers Actually Show Up

Not every assistant sends identifiable traffic, and the ones that do are not consistent about it. Here is what tends to land in your reports.

• ChatGPT sends chatgpt.com and, on older sessions, chat.openai.com. It also appends a utm_source parameter on some link formats, which is helpful because it survives even when the referrer is stripped.
• Perplexity sends perplexity.ai and is generally the most generous citer of the major assistants, so it often accounts for the largest slice of identifiable AI referrals.
• Google Gemini sends gemini.google.com. Note that this is distinct from AI Overviews, which do not.
• Microsoft Copilot sends copilot.microsoft.com, and sometimes bing.com when the answer is served inside Bing rather than the standalone app.
• Claude sends claude.ai, though only when a user clicks a citation link rather than reading the summarised answer.
• Mistral sends chat.mistral.ai, at low but growing volume in French-speaking markets.

The important omission is Google AI Overviews. Clicks from an AI Overview arrive as ordinary Google organic traffic, with no marker that distinguishes them from a blue-link click. There is no GA4 configuration that separates them. Search Console is slightly more useful here, since impressions from AI Overviews are folded into your web search data and a sudden impression rise with a flat click count is often the fingerprint of an Overview appearing above your result.

The second omission is everything that reaches a user through a mobile app or a desktop client that does not pass a referrer. Those sessions land in Direct, indistinguishable from someone typing your URL. If your Direct traffic has been climbing without an obvious cause, some portion of it is almost certainly AI-mediated.

Building an AI Assistants Channel Group

The channel group is worth the extra step over a saved filter for one reason: it makes AI traffic appear as a first-class row in the standard acquisition reports, which is where the rest of your organisation actually looks. A saved exploration is something only you will open.

Two configuration details cause most of the frustration. First, custom channel groups apply to reports going forward, not backwards, so create it the day you decide to care rather than the day before the board meeting. Second, your new channel must sit above Referral in the ordering, otherwise Referral catches everything first and your AI channel stays permanently at zero.

If you also want conversion attribution rather than session counts, add the same regex as an audience condition and mark the audience for export to Google Ads or your BI tool. Sessions tell you about exposure; conversions tell you whether that exposure is worth the effort of optimising for.

What AI Traffic Looks Like Once You Can See It

The first reaction to a working AI channel is usually disappointment. Volumes are small — for most B2B sites, somewhere between a fraction of a percent and three percent of total sessions, depending on category. Then the engagement numbers arrive and the picture changes.

AI referral traffic consistently behaves like the best segment in the report. Sessions run longer, bounce rates run lower, and conversion rates frequently land at two to four times the site average. The reason is structural rather than magical: the assistant has already done the filtering. Someone who clicks through from an AI answer has read a description of what you do, decided it matches their problem, and chosen to verify it. That is a very different visitor from someone who landed on a blog post from a broad query.

This is why judging AI search purely on session volume is the wrong call. A channel producing one percent of sessions and eight percent of qualified pipeline is not a rounding error. Report it on conversion contribution, not on traffic share, or you will talk yourself out of a channel that is quietly working.

The Blind Spot: What GA4 Will Never Show You

Here is the uncomfortable part. Referral traffic is the visible tip of AI exposure, and it is a small tip. When an assistant describes your product across three paragraphs, names your pricing model and compares you favourably to a competitor, the user very often has what they needed and never clicks anything. That answer shaped a purchase decision and left no trace in your analytics whatsoever.

The inverse is worse. If an assistant consistently recommends three competitors and omits you entirely, GA4 shows you nothing at all — not a decline, not a warning, just an absence that looks identical to a channel that was never going to send traffic. You cannot diagnose a visibility problem from a tool that only measures clicks.

Reporting Mistakes That Inflate or Hide AI Traffic

A few recurring errors make the number wrong in both directions.

• Counting bot traffic as visits. AI crawlers such as GPTBot, ClaudeBot and PerplexityBot hit your server constantly, but they do not execute JavaScript and so do not appear in GA4. If your AI numbers come from server logs instead, you are measuring crawls, not humans. Keep the two reports separate and label them clearly.
• Treating a referrer spike as a visibility win. One viral thread or a single well-placed citation can double AI referrals for a week and then vanish. Look at a rolling four-week average before drawing conclusions.
• Forgetting that Direct absorbs the overflow. App-based sessions and stripped referrers land in Direct, so your measured AI traffic is a floor, never a ceiling.
• Reporting AI traffic and organic traffic as competing lines. They are not substitutes. The pages that earn AI citations are usually the same pages that rank, and cannibalising your SEO budget to chase GEO tends to damage both.
• Letting the regex go stale. New assistant interfaces appear regularly and hostnames change. Review the pattern quarterly, or you will silently stop counting a channel you spent effort building.

Turning the Number Into Something Your Team Can Act On

Once the channel exists, the useful reporting habit is a monthly three-line view: AI sessions, AI conversion rate against site average, and the top five landing pages receiving AI referrals. That last line is the one that changes behaviour, because it tells you which pages the models trust enough to send people to.

Almost always those pages share characteristics — a direct answer near the top, a clear comparison table, specific numbers rather than adjectives, and a structure that can be quoted without surrounding context. When you find the pattern in your own top five, apply it to the pages you want cited next. That feedback loop is the practical payoff of setting the channel up in the first place.

Set it up once, review it monthly, and pair it with real visibility tracking so you can tell the difference between not being clicked and not being mentioned. Those two failures need completely different fixes, and GA4 alone will never tell them apart.