Why GA4 Cannot See Your AI Traffic

The first thing everyone does is open GA4 and look for ChatGPT. Some traffic is there — a few hundred sessions, maybe a few thousand — and it looks reassuringly like a channel being measured.

That number is wrong. Not slightly, and not because of a setting somebody forgot. It is wrong because of four specific mechanics in how GA4 assigns traffic to channels, three of which you can partially work around and one of which you cannot touch at all. Worse, the fourth mechanic does not merely lose the traffic — it hands the credit to a different channel, which is a strictly worse outcome than reporting nothing.

Mechanic one: no referrer, no channel

GA4 determines traffic source from two inputs: campaign parameters in the URL, and the HTTP referrer. If a visit has neither, GA4 has nothing to work with and assigns it to Direct.

This is correct behaviour. GA4 is not guessing, and it should not guess — it reports what the browser told it.

The problem is that the browser usually tells it nothing for AI referrals. Referrer policy defaults, protocol downgrades, rel="noreferrer", and above all mobile in-app webview handoffs mean most AI-sent visits arrive with no referrer at all.

So GA4's AI number is not a measurement of AI traffic. It is a measurement of the minority of AI traffic that happened to preserve a referrer or carry a campaign tag — and, as covered in the benchmarks audit, that minority is not a random sample.

Mechanic two: the credit goes somewhere else

This is the mechanic that turns a gap into an error, and almost nobody knows about it.

GA4's default conversion attribution does not assign credit to Direct. When a conversion's touchpoint chain ends in Direct, GA4 looks back through the lookback window for the most recent non-direct channel and assigns credit there instead. This is the long-standing "last non-direct click" convention, inherited from Universal Analytics, and for the world it was designed for it is sensible: someone who clicked your ad on Monday and typed your URL on Thursday was still acquired by the ad.

Now run an AI-discovered purchase through it.

The result is not a missing number. It is a wrong one in a specific direction: organic search, email and paid campaigns are systematically over-credited by exactly the amount that AI referral is under-credited.

That is considerably more dangerous than a blind spot. A blind spot makes you uncertain. This makes you confidently wrong, and it argues for spending more on whichever channel happened to touch the visitor first.

Mechanic three: the source list has to know the source

Even when a referrer is present, GA4 has to decide what channel it represents. Channel grouping works by matching the source against Google's maintained categorisation — this host is a search engine, that one is a social network, this one is a shopping site.

AI assistants do not have a settled place in that taxonomy. Depending on the host and how the list has been updated, an AI referral with an intact referrer may land in Referral, in Organic Search, in Unassigned, or in a custom group somebody built by hand.

This produces a distinctive symptom: your AI numbers are not merely low, they are inconsistently low, moving between channels as the classification shifts underneath you without any change on your side.

Mechanic four: the window is too short

AI-influenced purchases have long consideration cycles. The discovery moment happens when someone asks for a recommendation; the purchase happens after more research, comparison, and often an internal approval process.

Two GA4 limits bite here.

Session timeout. Sessions end after 30 minutes of inactivity by default. Someone who reads your pricing page, closes the laptop, and returns that evening generates two sessions, and only the first carries any source information at all.

Data retention. Standard GA4 retains event-level data for a maximum of 14 months, with a default of two months that many properties never change. Exploration reports cannot reach past the retention window. If your sales cycle plus your reporting lag exceeds it, the touchpoint is simply not in the dataset any more.

The two-month default is the one to check first. It is a live setting on a great many properties, and under it the entire question of long-window AI attribution is unanswerable regardless of everything else.

What you can do inside GA4

Worth doing, none of it sufficient:

Raise data retention to 14 months. Settings, one toggle, applies only going forward. Do it today — the cost of not doing it is unrecoverable.

Build a custom channel group for AI sources. Match known assistant hosts and campaign parameters. This gets you a clean view of the observable subset, which is genuinely useful as long as you label it honestly.

Report Direct as a trend, not a channel. Direct is your error term. Watch its share over time. If it is climbing without a brand campaign, app launch, or email programme to explain it, that movement is your best available proxy for the traffic you cannot see.

Look at a data-driven attribution model. It handles the Direct problem differently from last-non-direct-click. It does not solve the underlying issue — an unidentified touchpoint is still unidentified — but it stops the mechanical hand-off of credit to whatever came first.

Segment your Direct traffic by landing page. Direct visits arriving on deep pages — pricing, comparison, specific product pages — are behaviourally implausible as typed URLs. Nobody types a URL with three path segments from memory. That segment is where your dark AI traffic lives, and it is the single most revealing thing you can look at without new tooling.

Where GA4's limits actually stop

To be fair about scope: GA4 is a good tool at the job it was built for. Referrer-based channel attribution was the correct architecture for a web where referrers were sent. Nothing about its design is careless.

What it cannot do is infer a source that the browser did not transmit. That requires collecting signals GA4 does not collect — client characteristics, navigation context, request-level attributes — and reasoning about them probabilistically, which is a fundamentally different kind of system. It also requires being explicit about uncertainty, and GA4's data model has no place to put a confidence score. A channel is a string; there is no field for "we are 60% sure."

That is the real boundary. Not that GA4 is configured wrong, but that AI traffic attribution asks a question GA4's data model cannot express. What such a system needs to do instead is a different architecture, not a different setting.