Data sampling in GA4: when it happens and how to avoid it

Sampling means a tool doesn't process your whole dataset but takes a representative slice and estimates the total from it. That saves compute, which is why Google does it.

The benefit sits mostly with the vendor: cheaper to produce your report and faster to load. Your downside is simple: an estimate is less accurate than a count. On small variances you won't notice, on skewed distributions you will.

GA4 doesn't always sample. It kicks in on large datasets: from 500,000 sessions in GA4, and from 100 million sessions in GA360.

Worth knowing: standard reports are never sampled. Only more advanced analysis such as Explorations can be. If you run a smaller site and look at a few months at a time, this doesn't affect you.

GA360 users have more room, with a far higher threshold and a toggle between faster analysis with sampling and more accurate results without.

Look at the top of your report. A green check means no sampling. A yellow or red percentage means there is. Hover over it and you'll see what share of the data was used.

In GA360 the same spot holds the toggle between the two analysis modes.

Four routes, from simple to structural:

• Shorten the date range. Below 500,000 sessions nothing happens, so a shorter window often solves it on its own.
• Simplify the report. Fewer dimensions and segments means a smaller dataset and less chance of sampling.
• Use BigQuery. It never samples. You pay per query rather than nothing, but you get your full data back. GA4's admin screen has a BigQuery link under the property column; once it's set, your data flows there automatically.
• Move to GA360. Higher thresholds and a choice between fast and accurate, at a price.

Sampling is a useful technique that can give you a distorted picture of your users. As long as you know when it applies and how to recognise it, you can live with it.

Once it becomes structural, the BigQuery export is the answer. It costs tens of euros a month and gives you every hit back, unsampled.