Skip to content
Measurement

Measuring AI referral traffic: reading ChatGPT, Perplexity, and Gemini in GA4

How to build an AI-assistant referral segment in GA4, and why the number it gives you undercounts your real influence in AI answers.

6 min read · Updated 2026-07-08

You can see AI-assistant referrals in GA4. Build a channel group or segment that matches source hosts such as chatgpt.com, perplexity.ai, and gemini.google.com, and the clicks appear. Treat that number as a floor, not the measure. AI answers are often zero-click, many in-answer citations pass no referrer, and some assistants strip the source, so GA4 records only the trickle that escapes the answer. This article covers that trickle and its limits. A companion article covers AI Presence (AI Share of Voice), your presence inside the answers, which is the primary signal referral traffic cannot capture.

Build the AI referral segment in GA4

GA4 sorts traffic by session source and medium. When an AI assistant links out to your site, the visit usually arrives as a referral with the assistant's hostname as the source. To isolate these visits, create a custom channel group in Admin under Data display, or build a segment or comparison that filters session source to the assistant hosts.

Match on hosts such as chatgpt.com and chat.openai.com, perplexity.ai, gemini.google.com, copilot.microsoft.com, claude.ai, and grok. Some arrive with a utm_source parameter set to the assistant, others as a plain referral, so match the source with a regular expression or a contains rule rather than an exact string. Keep the list current as new assistants and hostnames appear.

One host you cannot cleanly isolate is Google AI Overviews. Those answers render on the Google results page, so clicks from them arrive as google organic and blend into your regular search traffic. GA4 has no separate hostname to filter them by.

Why the number is a floor, not the measure

The purpose of an AI answer is to resolve the question in place. When the user reads the answer and stops, there is no click and no session. Your brand can be named, quoted, and recommended inside the answer without a single visit reaching GA4.

Even when someone clicks a citation, many assistants open the link without passing a referrer, or route it through a redirect that hides the origin. GA4 then has no source to attribute, so the visit lands under direct or unassigned instead of your AI-referral segment.

Some answers cite sources without linking to them, or name a brand with no link at all. Voice answers and in-app responses may produce no web session. In each case the influence is real and the referral record is missing.

The result is that your GA4 AI-referral count is genuine but partial. It captures only the visible clicks that survive zero-click behavior, missing referrers, and stripped sources. That is a small and biased slice of your real presence in AI answers, which is why it is a floor and not the measure.

Measure presence inside the answer first

The signal that matters is presence: whether the answer names you, quotes you, and cites you. Presence happens whether or not anyone clicks, so it captures the influence that referral traffic drops. AI Presence measures it: how often your brand appears across the answers to the questions your buyers actually ask.

Getting AI Presence means reading the answers themselves at scale. Run your real questions across ChatGPT, Gemini, Claude, Perplexity, Grok, and Google AI Overviews, and record whether each answer mentions you, in what position, and with which citation. Referral logs cannot supply this, because the answers that never produce a click never appear in them.

This is the line between the two measures. AI Presence reads the answer. Referral traffic reads the click that escaped the answer. The presence layer is both larger and closer to the buying decision, so it belongs first in your reporting.

Read referral traffic as corroboration

Referral counts still earn a place, as trend and corroboration. A rising AI-referral segment confirms that your presence in answers is turning into visits. Read the direction over time rather than the absolute number, which will always be understated.

Watch the signals next to it. Because much AI influence is zero-click, look for lift in branded search and direct traffic in the weeks after your presence in answers grows. New-visitor branded queries and assisted conversions are the downstream echo of being named inside answers.

Triangulate three reads. Presence in answers is the primary measure, the AI-referral segment is the visible click trickle, and branded and direct lift is the zero-click echo. None is complete alone. Together they show whether AI answers are moving people toward you.

In Ansyra

Measure the answer, not just the click

Ansyra connects to GA4 to segment AI-assistant referrals, and measures AI Presence inside the answers themselves, the presence that referral traffic structurally cannot capture.

Straight answers

Frequently asked

Can I see Google AI Overviews traffic separately in GA4?
Not cleanly. AI Overviews render on the Google results page, so clicks from them arrive as google organic and blend into your regular search traffic. There is no distinct hostname to filter them by, so GA4 cannot separate AI Overview clicks from normal organic clicks.
Why do some ChatGPT clicks show up as direct traffic?
When an assistant opens a link without passing a referrer, or routes it through a redirect, GA4 has no source to attribute and files the session under direct or unassigned. That is one of the reasons the AI-referral segment undercounts.
If referral traffic undercounts, why track it at all?
Because trend and corroboration still matter. A growing AI-referral segment confirms presence is producing visits, and pairing it with branded and direct lift surfaces the zero-click influence GA4 cannot see directly. Track it as a floor and a direction, not as the total.

See where you stand in AI answers

Run your real questions across ChatGPT, Gemini and Google's AI answers, free for 14 days. No connections required to start; all five AI engines on every plan.