How To Track ChatGPT, Perplexity & Gemini Referral Traffic In GA4

Build a custom channel group for ChatGPT, Perplexity and Gemini in GA4 rather than relying on the native AI Assistant channel, because Google names ChatGPT, Gemini and Claude as examples and does not confirm Perplexity. Expect an undercount either way. Visits without a referrer land in Direct, and no GA4 setting recovers them.

This applies to any standard GA4 property where you can edit channel groups under Admin, and it counts the visits that arrive with a recognisable source.

GA4 shows clicks only, so it cannot show mentions that never produced a visit, and the number it gives is a floor, not a total.

If someone has asked you what AI search is worth, the rest of this guide gives you the setup, the split by engine and a one-page report you can hand over.

Does GA4 Track ChatGPT Traffic Automatically?

Partly, and not well enough to trust on its own. GA4 added an AI Assistant channel to the Default Channel Group on 13 May 2026. Google names ChatGPT, Gemini and Claude as examples and gives no full list, and its documentation does not name Perplexity. So the native channel exists, but you cannot assume it covers all three engines you care about.

Check what your own property shows before you build anything. These steps follow the current GA4 interface.

1. Reports, Acquisition, Traffic acquisition. Open the report and set the date range to the last 90 days. You should see a table of channels.
2. Look for an AI Assistant row. It sits in the Session default channel group dimension. If the row is there, your property has the native channel. If it is missing, you have no native AI number at all.
3. Add Session source as a secondary dimension. Read which hosts sit inside the AI Assistant row. If perplexity.ai is not among them, Perplexity visits are landing in another channel.

That last check matters. A channel you can't audit is a channel you can't defend in front of leadership. Keep the native channel as a cross-check against the number you build in the next sections, and don't make it the report itself.

Native Channel Or Custom Channel Group: Which One Do You Need?

You have four ways to isolate AI traffic in GA4. They don't do the same job.

Use a custom channel group. It applies to past data in reports and explorations and it can include Perplexity, while audiences apply only going forward. Keep the native channel as a cross-check.

One rule to remember. GA4 puts traffic into the first channel it matches in the group's order, according to Google's channel group documentation, so a channel placed above Referral takes matching visits first. Get the order wrong and your AI channel stays empty.

How To Build An AI Referral Channel Group In GA4

Every guide gives you this method. Here is the version that holds up when someone checks it.

1. Admin, Data display, Channel groups. Custom channel groups are built here. Create a new group by copying the default group, and name it something you will recognise. You now have an editable set of channels.
2. Add a channel named AI Referral. Use the add channel option and give it exactly that name. You now have an empty channel at the bottom of the list.
3. Set the condition. Choose Source, then matches regex, and paste the regex below. The channel now has a rule.
4. Drag the channel above Referral. If you also want one combined number with the native channel, drag it above AI Assistant too. Matching visits now fall into your channel first.
5. Save and wait. A new custom channel group can take a day or two to show data. After that, rows appear in your reports.

Here is the regex.

chatgpt\.com|perplexity\.ai|gemini\.google\.com

It names three hosts, chatgpt.com, perplexity.ai and gemini.google.com. Each part does one job.

• ChatGPT: chatgpt\.com matches visits whose source is chatgpt.com.
• Perplexity: perplexity\.ai matches visits whose source is perplexity.ai.
• Gemini: gemini\.google\.com matches visits whose source is gemini.google.com.
• The backslash: \. makes the dot a literal dot, so the pattern can't match a different character in its place.
• The pipe: | means or, so a visit only needs to match one of the three.

Don't copy a long host list from another site. Open Traffic acquisition with Session source as the dimension and add any other AI host you actually see in your own report. A host you have never seen in your data is a guess, and guesses make the channel harder to defend.

Two limits to know. A group allows up to 50 channels per group, which you won't hit here. Standard properties allow two custom channel groups, so don't burn one on a test. And because a custom channel group applies to reports and explorations retroactively, your history fills in as soon as data appears. An audience wouldn't do that.

How To Split ChatGPT, Perplexity And Gemini Into Separate Rows

One channel gives the leadership headline. A Session source exploration gives the per-engine breakdown, with Sessions, Engaged sessions and Key events as the three values.

Build it once you have the AI Referral channel.

1. Explore, then a blank free-form exploration. Name it something like AI Referral By Engine. You now have an empty canvas.
2. Rows, Session source. Import the dimension and drop it into Rows. Each source gets its own line.
3. Values, Sessions, Engaged sessions and Key events. Import all three and add them to Values. Each engine row now carries three numbers.
4. Filter to the AI Referral channel. Add a filter on the channel group dimension and set it to AI Referral. Choose your custom group's version of the dimension if GA4 lists it. The table now shows only the engines you matched.

Why bother with both views? The single channel is the headline you give leadership, because it moves less and is easier to explain. The three rows are for you. They tell you which engine is moving when the headline changes, and they show you whether one engine is carrying everything.

Clicks are only half the picture. If you need to know whether Perplexity mentions you at all, not just whether it sends visits, look at Perplexity visibility tracking tools.

Why ChatGPT Visits Show Up As Direct Or (Not Set)

Every number in this article rests on one uncomfortable fact. GA4 files a visit with no referrer and no source tag as Direct. Google's definition of Direct is exact. The source is (direct) and the medium is (not set) or (none). Anything that removes the referrer on the way to your site, such as a browser or app that doesn't pass it on, sends the visit there.

ChatGPT has had its own version of this problem. A reported ChatGPT link bug in October 2026 left some answer-text links untagged, so those visits went to Direct, and a fix began rolling out on 15 October 2026. Clickport's dataset shows identified ChatGPT visits falling 57% between 9 and 15 October 2026 against the prior baseline. That is one vendor's data, not a measure of the whole web. It does show how fast a real visit can vanish from the channel you built.

Here is the part people miss. A channel group regroups visits GA4 recorded with a source. It cannot restore source data that was never recorded. Building a better regex won't bring back a visit that arrived with nothing to match.

So the AI Referral figure is a floor. Report it that way, as at least this many visits, and put the reason in a footnote. A floor you explain is credible. A total you can't defend is the number that gets picked apart in the meeting.

Tag The Links You Control And Keep Crawlers Out Of The Count

Two practical points keep the number cleaner.

Tag the links you place yourself. UTM tags only help on links you place yourself. Add them wherever you control the link, so the visit carries a source whatever the referrer does. That covers:

• Partner and directory pages that link to you.
• Newsletters and email signatures.
• Profile links on social and community sites.
• Links you share into an assistant while testing.

Use the three standard parameters, utm_source, utm_medium and utm_campaign, and keep the values consistent so one campaign doesn't appear under four spellings. Don't assume an engine adds tags of its own. Read your report and see what actually arrives.

Separate people from bots. A crawler fetch is a bot reading your page. It is recorded in server logs and crawler analytics, not as a person clicking through in GA4, so a crawler spike is not a traffic gain. If your server logs show heavy AI crawler activity and your AI Referral channel is flat, those are two different facts. Don't blend them in a report.

To see crawler activity properly, compare AI crawler analytics tools. If you need to check that those crawlers can reach your pages in the first place, that is AI crawler access and technical readiness work.

Build A Landing Page View For AI Visits

Totals tell you whether AI visits exist. Landing pages tell you what to do about them.

Build another free-form exploration. Set Rows to Landing page plus query string, filter the channel to AI Referral, and set Values to Sessions, Engaged sessions and Key events.

1. Explore, then a blank free-form exploration. Name it AI Landing Pages.
2. Rows, Landing page plus query string. Each page appears as a row.
3. Filter to AI Referral. Only AI visits remain.
4. Values, Sessions, Engaged sessions and Key events. Sort by Sessions.

Look for two patterns. First, a few pages carrying most of the visits. Those are the pages engines send people to, so keep them current and give them a clear next step. Second, pages that get visits and no key events. Those are mismatches between what the visitor expected and what the page offers, and they're the cheapest fixes on the list.

This view shows where visitors landed, not what the engines cited. If you want that audit run across your whole site, see the GA4 and Search Console content audit tools.

What To Report To Leadership: Sessions, Conversions And Share Of Total

This is the page leadership will actually read. One page, four blocks, no invented numbers.

Share of total is AI Referral sessions divided by all sessions. Use it instead of an absolute count in the value view, because a share survives traffic swings. If the whole site has a bad month, the count drops with it. The share tells you whether AI is gaining or losing ground inside the traffic you do have.

Block 4 is the one people skip, and the one that protects you. State that the figure is a floor, and say what you left out.

Where key events end and deals begin. GA4 and your CRM own different halves of the story.

• GA4 owns: The visit, its source, and the key events it produced on your site.
• The CRM owns: The lead, the deal and the revenue that followed.

A lead can only be tied to ChatGPT when the visit arrived with a ChatGPT source. A visit filed under Direct can't be tied to any engine, so don't promise leadership a clean line from prompt to deal.

When you're ready to make this live, read about leadership dashboards for AI visibility. If you're choosing which numbers belong beside it, the AI visibility metrics to track give you the wider framework.

So what is AI search worth? In terms of this report, it is worth at least the sessions, key events and share of total in the AI Referral row, and the footnote tells leadership how much more may be sitting in Direct.

When Is An AI Referral Number Too Small To Act On?

Most of the time, early on. The rule below is this article's rule of thumb, not a benchmark.

The rule: Report counts and the trend, not a rate, until the channel has at least ten key events in the period. A rate built on a handful of conversions swings too widely to mean anything.

Here is the arithmetic, so you can check it yourself. Take 2 conversions from 40 AI sessions. That is a 5% rate, because 2 divided by 40 is 0.05. One more conversion makes it 3 from 40, which is 7.5%. A single person moved the rate by half again, from 5% to 7.5%. Nothing about your channel changed. One visitor did.

What to do below the line:

• Keep tagging the links you control.
• Keep the channel in the monthly report, as counts.
• Use the landing page view for content decisions.
• Make no budget decision on the channel.

What to do above it:

• Compare engagement and key events with organic search.
• Let that comparison guide spend, not the headline count.

Most readers will sit below the line for a while. That's normal. To see how much traffic these tools usually send, read how much traffic these tools send without expecting your own number to match. And for what a channel above the threshold can look like, the ChatGPT referral growth case study shows the other end. If your channel is too small and you want to grow it, that is generative engine optimization work.

What GA4 Cannot Tell You About AI Search

GA4 records visits and not mentions or prompts. That is the limit, and it's worth stating once and plainly.

• Mentions: GA4 can't see an engine naming your brand if nobody clicks.
• Prompts: It can't show the question behind a visit.
• Uncited and unclicked: A citation nobody clicks is invisible to it.
• Direct: Some real AI visits are hiding there, and no setting brings them back.

Treat the GA4 number as one half of the picture, and track mentions separately for the other half. GA4 only sees visits that click through. To see whether ChatGPT, Perplexity and Gemini mention your brand when nobody clicks, we pair it with AI visibility tracking in Semrush, which covers the mentions that never reach your site.

Two places to start on the mentions side. You can track brand visibility in ChatGPT yourself, and you can run an AI visibility audit to see where you stand across all three engines. If you'd rather not talk to anyone yet, the free AI Visibility Report puts referral sessions next to mention data without a call.