AI traffic in GA4 — how to identify visitors from ChatGPT and Perplexity

Traffic from AI services shows up in GA4 as referral traffic, but in the default reports it disappears among everything else. When ChatGPT, Perplexity or Gemini links to your site and a user clicks, that visitor is recorded under sources such as chatgpt.com or perplexity.ai. They just need to be gathered into a channel of their own. Here is how to do it and what the numbers actually tell you.

Where AI traffic comes from and what it looks like

When an AI answer contains a link to your site and the user opens it, the browser passes the referrer as normal. In GA4 these visitors appear in the Traffic acquisition report as a source and medium pair, for example chatgpt.com / referral or perplexity.ai / referral. Typical source domains:

• chatgpt.com and chat.openai.com — ChatGPT
• perplexity.ai — Perplexity
• gemini.google.com — Google Gemini
• copilot.microsoft.com — Microsoft Copilot
• claude.ai — Claude

Two caveats about the numbers. Some AI clicks arrive without referrer information, from apps and certain browser environments, in which case they register as direct traffic. So you are measuring a floor, not the whole truth. And clicks from Google’s AI Overviews do not separate into a source of their own at all; they blend into google / organic.

Gather the AI sources into one channel

A one-off setting in GA4: Admin → Data display → Channel groups. Create a custom channel group and a channel in it, for example ”AI search”, with the condition: source matches regex

chatgpt|openai|perplexity|gemini|copilot|claude

After this, AI traffic appears in the reports as a single line alongside Organic Search and Paid Search, and you can follow its development, conversions and landing pages without manual work. The expression is worth extending as new services appear, because the list keeps changing.

A lighter alternative without any settings: build a saved report in Explore where the source is filtered with the same expression. It does the same job, but you have to go and look at the numbers separately.

What to look at in the numbers

Landing pages before volume. Total AI traffic is still small on most sites, and that is expected. What is more interesting is which pages it lands on. Those pages have become sources in the answers, and their pattern is worth repeating.

Conversion rate separately. Someone clicking from an AI answer has often already compared options and received a recommendation, so visitor quality can be high even when the volume is low. Compare the AI channel’s conversion rate against organic; that tells you the channel’s real value better than session count does.

Trend by month. Individual days swing wildly at small volumes. A month-on-month comparison shows whether the channel is growing, and after GEO work it should be.

Remember the blind spot in measurement

The most important limitation, stated honestly: a large part of the effect of AI search never appears in analytics at all. When a model answers a question mentioning your company but the user clicks nothing, or goes to your address directly later, the effect registers as direct traffic or a brand search. That is why it is worth measuring visibility itself alongside analytics: test the same buying-intent questions regularly across different models and record whether the company is mentioned. The traffic report tells you the clicks; the question tests tell you the mentions the clicks eventually come from.

Check first that measurement works at all

Before building channel groups, confirm two things. Otherwise you are chasing numbers that were in the wrong place to begin with.

Referral spam out. The GA4 referral report collects stray domains that are not real traffic. If you have not set your own domains as excluded referrers, payment flows and subdomains show up as referrals and distort the comparison.

Cookie consent affects the numbers. If the site has a consent banner and measurement starts only after acceptance, some AI visitors disappear from measurement entirely. This cannot be fixed with settings, but it has to be known: the figures are systematically low, and the shortfall grows with the banner’s rejection rate.

A third check concerns the time window. AI traffic is so small on most sites that a week’s sample says nothing. Set the comparison period to at least a month from the start.

Separate the bots from the humans

This gets mixed up often. AI services generate two kinds of traffic, and only one of them appears in analytics.

Crawlers such as GPTBot and PerplexityBot fetch your pages to train models or to answer questions in real time. They do not execute JavaScript the way a browser does, so as a rule they never register in GA4 at all. Bot traffic is visible in server logs, not in analytics.

Referral traffic, by contrast, is real people who read an answer and clicked a link. Only this shows in GA4. The practical implication matters: if the logs show busy bot traffic but GA4 shows no AI referrals, models are reading your pages without ending up recommending them. That is a different problem from not being reachable, and it is fixed with content rather than technology.

Judging the effect of AI Overviews

Because AI Overviews clicks blend into ordinary organic traffic, the effect has to be read indirectly from Search Console. Three signs indicate an answer box has appeared in your topic.

The first is impressions growing without clicks growing. The page appears as often or more often, but fewer people click, because the answer is read straight from the results. The second is average position improving while click-through rate falls. The third is the change concentrating in question-shaped searches: queries beginning with how, why and what react first.

This is interpretation rather than precise measurement. It is still worth doing, because the alternative is to assume a traffic drop is your own doing when the cause is a change in the shape of the results page.

Three mistakes that make the numbers misleading

Adding the channel mid-month. If you enable a custom channel group halfway through a month and then compare months, the first month looks like half. Enable it at a month boundary or note the date from which the figures are comparable.

Missing subdomains in the expression. ChatGPT uses both chatgpt.com and the older chat.openai.com, and Copilot appears in several forms. If the expression covers only some, part of the traffic falls outside the channel and it looks smaller than it is.

Internal traffic included. Your own team tests AI services and clicks its own links. A handful of sessions a month is enough to distort a small channel. Filter the office IP or use a separate test view.

What these share is that the error is invisible in the report. The number exists and looks plausible, but it measures something other than you think. Write measurement definitions down when you make them rather than trying to recall them six months later.

Build a report you will actually read

Measurement dies when nobody opens the report. Keep it to one page and four numbers: AI channel sessions, its conversions, the five most common landing pages and the volume of brand searches. Monthly review is enough, and each quarter it is worth checking whether the landing pages have changed.

If you use Looker Studio, the same fits into a single view with GA4 and Search Console side by side. The tool matters less than looking at the same numbers the same way every month. A shifting method produces shifting conclusions.

Frequently asked questions

How do I see traffic from ChatGPT in GA4?

In the Traffic acquisition report under the source chatgpt.com or chat.openai.com, with referral as the medium. The most practical approach is to gather all AI sources into a custom channel so they appear as one line.

Why does my AI traffic show as zero?

Either your site is not yet reaching the sources in answers, or the clicks are registering as direct traffic without a referrer. Start by checking with a few test questions whether the models mention your site at all.

Do clicks from Google AI Overviews show separately in analytics?

They do not. They register as ordinary organic Google traffic, so the effect of AI Overviews has to be assessed indirectly from changes in impressions and clicks for question searches in Search Console.

Is AI traffic more valuable than other traffic?

Measured by conversion rate, often yes, because the visitor arrives having received a recommendation and already done the comparison. Because of the small volumes, conclusions are still best drawn from months of data rather than days.

Do AI crawlers show up in GA4?

As a rule, no. Crawlers such as GPTBot do not run the measurement script, so they appear only in server logs. What GA4 shows is people who clicked a link inside an answer.

How often should AI traffic be reviewed?

Monthly is enough. The volumes are still small enough that at daily or weekly level random variation hides the real development and leads to wrong conclusions.

Measurement is one half of GEO work; the other is getting into the answers. You can start on that from the GEO service page, and see where you stand with the AI visibility test.