Three corrections to every GA4 guide for tracking ChatGPT referrals

Google added an AI Assistant default channel group to GA4 on 13 May 2026, and it excludes Perplexity.

That one fact invalidates the screenshot in about a dozen near-identical tutorials, all of which tell you to open Acquisition, find the new channel, and report what it says.

The number it says is wrong in three separate directions, and none of them is a configuration mistake you can fix by being more careful.

Here is the part I got wrong first. We built a clean custom channel group at NovaTechRay in June, checked it against the raw source list, and were pleased with it for about six weeks. Then the landing page distribution stopped making sense, and the reason turned out to sit inside ChatGPT rather than inside GA4.

So this is the correction, not the tutorial.

Correction 1: the built-in channel is incomplete by design

Start by not trusting the default.

The AI Assistant channel is a convenience grouping, and a grouping that leaves out Perplexity is not a category, it is a subset.

If your reporting rolls up to that channel, you are comparing months during which the definition of the channel has changed underneath you.

Build a custom channel group instead, with a source match you control:

chatgpt.com|chat.openai.com|perplexity.ai|copilot.microsoft.com|
gemini.google.com|claude.ai|bing.com/chat|you.com|poe.com

Then keep the raw source list open beside it once a month and add what is new. Assistants ship new domains, new apps and new link formats faster than a default channel definition gets updated, so the list is maintenance rather than setup.

And label the custom group clearly enough that nobody confuses it with the built-in one in a shared report. Two channels with similar names and different membership is how a board deck ends up arguing about which of two true numbers is correct.

Correction 2: a third to two thirds of it never arrives labelled

This is the ceiling, and it is not yours to raise.

Similarweb's analysis, published 29 July 2026, puts the share of AI referral sessions that arrive with no referrer at 35 to 70 percent. A session with no referrer, no UTM parameters and no click identifier is Direct by definition, and nothing in GA4 can recover it afterwards. You cannot close it.

The usual reflex is UTM tags.

They do not apply here. You do not control the link inside an answer.

So whatever your AI channel reports, the true figure sits somewhere above it in a range you cannot narrow with better tracking. That is not a reason to skip the measurement. It is a reason to label it a floor every single time it appears in a report, and to stop treating month-on-month movement in a number with that much unmeasured mass as a result.

There is a weak signal worth watching alongside it.

Direct sessions landing on deep pages rather than the homepage were followed links, because nobody types a URL with three slashes in it from memory. We wrote that one up in the booking that came from ChatGPT and never showed up in your analytics. It is a hint about direction and never a count.

Correction 3: page-level attribution is broken at the source

This is the correction the tutorials cannot patch, and the one that cost us six weeks. It is structural.

Similarweb reports that after ChatGPT's May 2026 homepage-link change, roughly 62 to 63 percent of its referrals now land on homepages. Read what that does to your analysis. The landing page in your report is a product decision made by the assistant, not evidence about which of your pages was retrieved, quoted or cited.

So the standard GEO deliverable, a list of which pages are winning AI traffic, is measuring the wrong object entirely.

Two pages could be doing all the work in the answer and both show up as one homepage session. A page you have never optimised gets credit because it is the destination the interface chose. The correlation between cited page and landed page has been severed at the product layer, and no amount of event configuration reconnects it.

The right move is to stop producing that report. Page-level performance in AI search is measurable, but it is measurable from the answer side rather than the analytics side: run the prompts, record which URLs get cited, count those. The sampling method is in how to measure whether AI assistants are recommending you.

Analytics tells you about arrivals. The answer tells you about citations. They stopped being the same question in May.

One consequence worth planning for. If most AI arrivals now land on your homepage, the homepage has to carry the job the deep page used to: say what you are, where you are, what it costs and what to do next, above the fold and in plain sentences.

So a broken attribution model quietly turns into a copy brief.

That is the only useful thing to come out of this particular change.

Correction 4: what a defensible report contains

Four instruments, each with a stated limit, reported together.

Share of answer from a fixed prompt set, run enough times that the number survives the non-determinism. This is the only one that measures the thing you are buying, and it lives outside analytics entirely.

AI referral sessions from your custom channel group, labelled as a floor and never as a total.

Branded search volume and Direct sessions, trended against the same timeline. Someone reads a recommendation, never clicks, and searches your name three days later. Rising branded search with flat category rankings is the cleanest indirect evidence available.

And a how did you hear about us field with the assistants as named options on a fixed list, never a free-text box. It is crude and people report the last step rather than the first, so it also reads as a floor.

None of those four is good on its own.

Together they bracket the truth, which is the most any of us can claim right now, and it is what we put in front of a client at NovaTechRay instead of a single confident percentage.

The short version

• Replace the built-in AI Assistant channel with a custom group, because the default excludes Perplexity.
• Refresh your source regex monthly from the raw list, since assistants ship new domains faster than defaults update.
• Report AI referral sessions as a floor, given that 35 to 70 percent arrive with no referrer.
• Stop producing page-level AI attribution from analytics, because most ChatGPT referrals now land on the homepage.
• Measure page-level citation from the answer side, by recording which URLs get cited in your prompt runs.
• Pair the referral floor with share of answer, branded search, Direct trend and a fixed-option source question.
• Put the limit next to every number, in the same sentence.

Building that four-instrument report is a normal first month at novatechray.com.