GA4 Data Sampling: When It Happens and How to Avoid It

Sampling is GA4's answer to an expensive question: what happens when you ask for an analysis across more events than the system wants to scan? Instead of refusing, GA4 scans a subset, multiplies the results up, and shows you an estimate. Often the estimate is fine. Sometimes it is quietly wrong in exactly the rows you care about.

When GA4 samples and when it doesn't

The dividing line runs between the two halves of the product. Standard reports, the pre-built ones in the left navigation, are served from aggregated tables GA4 prepares in advance, so they are not subject to query-time sampling. They have their own caveats, such as cardinality limits that collapse long tails into an "(other)" row, but the numbers are computed from all the data.

Explorations are computed on demand from raw events, and that is where the limit lives: on a standard property, a query touching more than 10 million events gets sampled. Note that the limit is about events read, not the size of the result. A simple question over a long date range on a busy site crosses the line easily, because every event in the range counts against the quota whether or not it appears in your output.

Analytics 360 raises the ceiling to roughly a billion events per query, which is one of the clearer reasons enterprises pay for it.

How to spot it

Every exploration shows a data-quality indicator in the top corner. Green means the result used 100 percent of available data. When sampling kicked in, the indicator changes color and reports the share of events actually used, which can be a large fraction or a startlingly small one depending on how far over the limit your query went.

It is worth building the habit of checking, because nothing else about a sampled exploration looks different. The charts are just as confident either way.

What sampling distorts

Sampling error is not spread evenly. The total sessions for your whole site, estimated from a 40 percent sample, will land close to the truth, because large numbers estimated from large samples behave well.

The damage concentrates in small segments. A landing page with 90 conversions in reality might contribute a handful of events to the sample; multiply the sample back up and the estimate can miss badly in either direction. Run the same exploration twice and sampled long-tail rows can visibly change between runs. Funnels compound the problem, since each step filters the sample further, so the later steps of a multi-step funnel rest on very few observed events. The rows most likely to drive a decision, the surprising campaign, the underperforming step, are precisely the rows sampling estimates worst.

Reducing and avoiding it

The workarounds all reduce the events a query must read.

Shorten the date range: if 12 months samples, four monthly queries may each stay under the limit, at the cost of stitching results together yourself. Simplify the query, since fewer segments and filters sometimes keep evaluation cheaper. Prefer a standard report when one answers the question, because the aggregated tables are unsampled. And check whether your property has options for higher-fidelity results on individual explorations, which trade speed for accuracy where available.

The complete answer is the BigQuery export. Raw events in BigQuery are never sampled; a query scans everything you ask it to and charges you for the privilege. For a site large enough to hit GA4 sampling regularly, the export plus SQL is the only way to get exact numbers for arbitrary questions, with the real cost being the engineering time to reconstruct sessions, attribution, and funnels that the GA4 interface otherwise handles.

Analyse computes on all of your data

Analyse does not sample. Every report, funnel, and breakdown is computed from the complete event stream, whatever the date range, because the storage model is built for aggregation queries rather than retrofitted around an events limit. A funnel over a year of traffic in Analyse's conversion funnels counts every visit at every step, and the number does not change when you run it again. For most sites under tens of millions of pageviews a month this is not a heroic claim, just a different set of priorities: exact answers in the dashboard rather than estimates with an export escape hatch. The GA4 comparison lays out where each tool's limits actually sit.