How do I track traffic from ChatGPT in Google Analytics?

What does GA4 do on its own now?

Since 13 May 2026 the Default Channel Group has included an AI Assistant channel, and it needs no configuration.

Google's release note for that date reads: You can now identify how users are discovering your site through chatbots like ChatGPT, Gemini, and Claude via a new AI Assistant channel in your Default Channel Group reports. The channel documentation describes it as the channel by which users arrive at your site from sources like ChatGPT, Gemini, Deepseek, Copilot, or Grok, and states that the medium is set to ai-assistant if the referrer matches a list of AI Assistants.

Two consequences follow from that last clause, and they matter more than the feature itself:

• It is forward-only. Sessions before 13 May 2026 are not reclassified, so your earlier AI traffic stays buried in Referral.
• It keys off the referrer. A session that arrives without one cannot land in this channel, by definition. That is the source of almost all of the undercount described below.

Where to look: Reports, then Acquisition, then Traffic acquisition, with the dimension set to Session default channel group.

Sources: Google Analytics release notes, 13 May 2026. GA4 default channel group.

How do I build a custom channel group as a backstop?

Create one in Admin, add a channel matching session source against the assistant domains, and put it above Referral.

1. Admin, then Data display, then Channel groups, then Create new channel group.
2. Add a channel and name it something you will recognise later, such as AI assistants.
3. Set the condition on Source, matching any of chatgpt.com, chat.openai.com, perplexity.ai, gemini.google.com, copilot.microsoft.com, claude.ai and you.com. Add new hosts as you spot them in your referral report; the list changes every few months.
4. Move it above Referral. Google's documentation is explicit: Traffic is included in the first channel whose definition it matches given the current order of channels in the group. Left below Referral, your channel will never fire, because Referral will have claimed the session first. This is the step people get wrong.
5. Save, then switch the dimension in Traffic acquisition to your new group.

A standard GA4 property allows two custom channel groups and a 360 property allows five, so spend them carefully.

Worth building even though the native channel exists, for one reason: the custom group is yours to edit the day a new assistant appears, whereas the native channel updates when Google decides to update it.

Source: GA4 custom channel groups.

Why will the number always be too low?

Because a large share of AI-influenced visits never carry an AI referrer, and some never produce a visit at all.

• In-app web views. A link opened inside the ChatGPT mobile app frequently arrives with no referrer, so GA4 files it under Direct. Since the AI Assistant channel is defined by a referrer match, those sessions cannot reach it. Test this on your own site: open one of your pages from a ChatGPT answer on your phone, then watch where it lands in your realtime report.
• The name search. An assistant shortlists you, the buyer then searches your business name in Google, and the session arrives as Organic Search. No field anywhere records why they typed your name. This is the single largest gap, and it is the one we describe on is AI visibility worth it.
• No visit at all. Pew Research Center tracked the Google searches of about 900 US adults through March 2025 and found people who saw an AI summary clicked a traditional result in 8% of visits, against 15% without one, and clicked a link inside the summary in 1% of visits. That is US data about Google's AI Overviews rather than Australian data about ChatGPT, and the direction is the point.
• Privacy tooling. Consent settings, tracking prevention and referrer policies all strip referrers independently of any of the above.

Read the AI Assistant channel as a floor. Growth in it is meaningful. Its absolute value is not the size of your AI audience.

Source: Pew Research Center, 22 July 2025.

What should I measure instead of referrals?

Whether you are named, and by whom, on a fixed set of questions you re-ask on a schedule.

If the buyer never clicks, presence in the answer is the only thing that happened, so presence is the thing to count. Four numbers are worth keeping:

• Share of the answer. How many of your buyer questions name you at all.
• Position within the answer. First named and third named are not the same outcome.
• Who is named instead. The most actionable field on the whole list.
• Which sources the answer cites. That is your off-site target list, derived rather than guessed.

Two rules make the series usable. Keep the questions fixed, because changing them makes the trend meaningless. And re-ask on a schedule, because a single run is a sample of one, and answers move between runs for reasons built into how the models work. The two engines are also not interchangeable readings of the same thing: across the 137 businesses in our scan with complete records from both, Gemini scored the same business higher than ChatGPT for 105 of them.

Source: Answerable, The State of AI Visibility in Australia 2026. 300 Australian businesses, six industries, scanned 30 July 2026. Raw data published as CSV and JSON under CC BY 4.0.

How do I wire this to enquiries rather than sessions?

Put one open question on your enquiry form and read the answers.

1. Add How did you hear about us? as a free-text field, not a dropdown. A dropdown never contains ChatGPT told me you were the ones who do heritage roofs.
2. Tag every enquiry record with what they said, and count monthly. It will undercount, because most people do not report and many genuinely do not remember.
3. Mark your enquiry events as key events in GA4, so the AI Assistant channel can be read against enquiries rather than sessions. A channel with few sessions and a high enquiry rate is a different story from one with many sessions and none.
4. Ask on the phone. Sales conversations capture what no analytics property will.
5. Put the three series side by side: questions you are named in, AI Assistant sessions, and self-reported AI enquiries. If the first rises and the third follows with a lag, that is the closest thing to evidence this work produces.

None of that is attribution in the strict sense, and we would rather say so than sell a dashboard that implies otherwise. The honest framing is that you are triangulating a thing that cannot be measured directly, and that being named is the part you can actually count.

Measure being named, not just being clicked.

Referral numbers will always be a floor. A free scan counts the thing analytics cannot see: how often ChatGPT and Gemini name you when a buyer asks.