How to track AI search traffic in GA4

If you want to know whether AI search is sending you real visitors, Google Analytics 4 is where you look. It is also where most people get a number that is quietly wrong.

GA4 will show you sessions from ChatGPT, Perplexity, Claude, Gemini, and Copilot if you set it up correctly. What it will not tell you, unless you know to ask, is that it undercounts those sessions badly and that its new built-in report only recognizes three of the five engines. This guide is the honest setup: how to see the AI traffic GA4 does capture, and how to account for the large chunk it misses.

What tracking AI search traffic in GA4 actually means

Tracking AI search traffic in GA4 means isolating the sessions that arrive from AI engines like ChatGPT and Perplexity, so you can see which pages they land on and what those visitors do. GA4 logs these as referral traffic by the source domain. The catch is that a large share of AI referrals arrive with no referrer and fall into Direct, so the number you see is a floor, not the full count.

Hold onto that last sentence. It is the difference between using GA4 honestly and fooling yourself with it.

First, the two things Google changed in 2026

Before the setup, two recent changes you need to know about, because half the guides online predate both.

1. GA4 now has a native AI Assistant channel (since 13 May 2026). Google added an “AI Assistant” row to the Default Channel Group. When a session arrives with a referrer Google recognizes as an AI tool, it gets filed there automatically, with the medium ai-assistant. Useful, and zero setup. The limitation: Google’s recognized list is ChatGPT, Gemini, and Claude. Perplexity and Copilot are not in it. If you rely on this channel alone, you are blind to two of the five engines that matter most for B2B.

2. AI traffic is classified as referral, not organic search. Because the click comes from another site’s domain (chatgpt.com, perplexity.ai), GA4 treats it like any other referral. It does not roll into your Organic Search numbers. So if you have been judging AI’s contribution by watching Organic Search, you have been looking at the wrong row the entire time.

The method, step by step

You can capture all five engines and your full history with one custom channel group. Here is the setup I use on every client account.

1. Confirm the native channel first. Go to Reports, then Acquisition, then Traffic acquisition. Set the primary dimension to Session default channel group. If you have had AI traffic since mid-May 2026, an “AI Assistant” row appears. This is your quick gut check, but remember it only counts three engines. Do not stop here.

2. Build a custom channel group for all five. In Admin, under Data display, open Channel groups and create a new group. Add a channel called “AI Search” defined by Session source matching a regex of the AI domains. This catches Perplexity and Copilot too, and it backfills your history rather than starting from mid-May.

3. Use this regex. Match on Session source or referrer:

Copy it as one line. Add engines as new ones appear. The point is that you, not Google, decide which sources count.

4. Cross-check with an exploration. Build a blank Exploration. Add the dimensions Page referrer and Landing page, and the metric Sessions. Filter Page referrer on the same regex. This shows you the exact landing pages AI engines send people to, which is the data that actually tells you which content is doing the work.

5. Mark a key event so you can see if it converts. Traffic that does not convert is a vanity number. Tag your real conversion (a demo request, a trial signup, an audit booking) as a key event, then segment it by your AI Search channel. Now you can answer the only question that matters: does AI search traffic turn into pipeline, or just pageviews.

The number is a floor, not the truth

Here is the part the tool-vendor guides skip past.

Between 35 and 70 percent of AI referral sessions arrive with no referrer information. The browser strips it, the click comes through an in-app webview, or the engine routes it in a way that drops the source. GA4 cannot file what it cannot see, so those sessions land in Direct alongside your bookmarks and your dark-social traffic.

What that means in practice: the AI Search number in your report is real, but it is the visible tip. The true figure could be double. So treat GA4 as a directional floor. If it shows AI search climbing month over month, that trend is trustworthy even though the absolute number is undercounted. What you cannot do is quote the raw session count as “all the traffic AI sent us,” because it is not.

There is a partial tell. Watch your Direct traffic to high-intent pages, the comparison pages and product pages a person rarely types in by hand. If Direct to those pages rises in step with your known AI Search sessions, a chunk of that Direct is almost certainly unattributed AI traffic. It is an inference, not a measurement, but it is a more honest read than pretending the Direct bucket is all bookmarks.

Why this only measures half of AI search

GA4 measures clicks. It tells you who arrived after an AI engine sent them. It is silent on the larger event happening upstream: whether the engines are citing you at all.

Most AI answers do not produce a click. The engine names you, summarizes your point, and the user reads it inside ChatGPT without ever visiting your site. That is still a win, brand presence at the exact moment of research, and GA4 records none of it. So analytics is the downstream proof, the clicks that did happen. It is not a measure of your AI visibility.

For that you need the upstream view: running your buyer questions through each engine and scoring how often you are cited, which is a separate discipline from anything in GA4. The two fit together. GA4 confirms the clicks are real and converting. The citation protocol confirms you are present in the answers whether or not they click. Use one without the other and you are half blind. The full method for the upstream half is in the guide to measuring AI search visibility, and the free tools that read both signals are in the GEO tools I actually use.

What did not work

A few attribution mistakes I made before settling on the setup above.

Trusting the native channel as complete. When Google shipped the AI Assistant channel I used it as-is for a few weeks and reported off it. Then I cross-checked against a referrer exploration and found a stack of Perplexity sessions sitting outside the channel entirely. The native row was undercounting an already undercounted number. The custom channel group fixed it.

Reading AI traffic in the Organic Search row. Early on I expected AI clicks to show up under Organic Search, watched that row, and concluded AI was sending almost nothing. It was sending plenty. It was all sitting in referral, because that is how GA4 classifies it. I had been reading the wrong line for weeks.

Quoting the session count as gospel. I once put a clean AI-traffic number in a client report without the Direct-bucket caveat. A sharper client asked why their server logs showed more ChatGPT user-agent hits than my GA4 number. They were right. The honest report says “at least this many, likely more,” not a single confident figure that the data cannot support.

Why this matters

If you are investing in AI search, you need to know it is producing more than citations you can screenshot. GA4 is where you prove that real people are arriving and converting. Set up correctly, it is the closest thing to a profit-and-loss statement for your AI search work.

Just hold it honestly. It captures a floor, not a ceiling. It sees three engines natively and five only if you build the channel yourself. And it measures the clicks, never the citations that drove them. Read it as the downstream half of a two-part picture and it is the most useful free tool you have. Read it as the whole picture and it will undercount your wins and miss the biggest one entirely.

If you want this set up and read against your own buyer journey, that is part of a paid audit, and the full approach lives on the methodology page.