GA4’s estimates, and what’s actually counted.

When a visitor declines cookies, GA4 loses sight of them. Google's answer is behavioural modelling: estimate the people you couldn't measure from the people you could, and blend the estimate into your reports.

That isn't a scandal. Estimating is a reasonable thing to do. The problem is narrower and more practical: once the estimate is blended in, you can't point at a number and say which part was counted and which part was guessed. Here's how GA4's modelling works, from Google's own documentation.

What happens when someone says no

With consent mode, Google says: "When visitors deny consent, instead of storing cookies, tags send pings to Google." Those pings carry "a minimum of information about user activity". They're the raw material for the model.

If your banner blocks Google's tags completely until someone says yes, nothing is sent, and there's nothing to model from.

When GA4 starts estimating

Modelling isn't automatic for every property. Google lists the prerequisites:

• At least 1,000 events a day from visitors who said no to analytics, for at least 7 days.
• At least 1,000 daily users who said yes, on at least 7 of the previous 28 days.

Even then, Google says meeting these "doesn't guarantee eligibility", because the model has further criteria of its own, such as the ratio of new to returning users. Training "may take more than 7 days", and estimated data is only available "from the date the property became eligible".

So a small site may never be modelled at all. A large one may switch between modelled and unmodelled as traffic changes.

What gets modelled, and what doesn't

Google is specific about this, and it surprises people. In reports, "modeling is applied to metrics like users, sessions, and new users counts. However, modeling is not applied to event counts such as page_view, first_visit, and session_start."

So within one report, your user count can include estimates while your event counts don't. Two numbers side by side, built two different ways.

How you can tell

Modelling only shows up when your property uses what Google calls the blended reporting identity. Then a data-quality icon on a report tells you whether it's "Including estimated user data", or why it isn't.

That icon is at report level. It tells you the report contains estimates. It doesn't tell you which rows, which days or what share of a number was estimated.

Counted, observed, inferred

There's another way to lay the same data out, which keeps the three kinds apart:

Nothing in that table is cleverer than GA4's model. The difference is that every inferred number is labelled as inferred wherever it appears, so nobody mistakes an estimate for a count in a client meeting.

What this means for you

• Know which identity your property uses. Check the reporting identity setting before you quote a GA4 user count to anyone.
• Prefer counts for decisions about money. Event counts aren't modelled; user counts may be.
• Watch for the eligibility date. A jump in users on the day modelling switched on is the model arriving, not your marketing working.
• Don't drop your banner because GA4 models. Modelling fills gaps in reports. It doesn't change what the law requires before data goes to Google.