How sampling works
Sampling is when a tool answers your question by examining a portion of the data rather than all of it, then scaling the result up. It is a speed trade. Reading everything is slow and expensive, so beyond a certain size the system reads a slice and estimates the rest.
The estimate is usually close for large, common things and unreliable for small, rare ones. A sampled report can tell you confidently which channel is biggest, and can be badly wrong about a niche landing page that converted a handful of times. The narrower the question, the more the estimate wobbles.
You normally meet it in custom or ad-hoc analysis over long date ranges rather than in standard reports. Tools usually show an indicator when a report is sampled, and it is worth learning exactly where yours puts it.
Why sampling matters
Because a sampled number invites a decision it cannot support. Two campaigns that appear to differ may be identical once the full data is read, and reallocating budget on that basis costs real money for no reason.
It also breaks reconciliation. When a sampled analytics figure is compared against an unsampled ad platform figure, the difference looks like a tracking fault, and people go hunting for a bug that was never there.
Where sampling goes wrong
The commonest error is simply not noticing. The indicator is small and easy to skip, and screenshots pasted into a client report lose it entirely, so the caveat never reaches the person making the decision.
The second is drilling down inside a sampled report. Each extra condition shrinks the underlying slice, so the deeper you go the shakier the estimate becomes — precisely where you wanted more precision, not less.
The third is treating sampling as an accuracy problem to be fixed. It is not a bug; it is a documented trade-off. Confusing it with genuine data discrepancies between platforms sends the whole investigation in the wrong direction.
How to act on it
Check the sampling indicator before quoting any figure from an exploration or custom report, and note its presence in the report itself. Shortening the date range, removing unnecessary conditions, or using a standard report where one answers the question will often return an unsampled result.
For questions that genuinely need the full data — revenue reconciliation, rare high-value conversions, anything a client will challenge — export raw event data to a warehouse such as BigQuery and count it there, or take the figure from the system of record instead. Your ecommerce platform or CRM knows exactly how many orders there were; analytics is estimating.
And keep the stakes in proportion. For a direction-of-travel question on a busy channel, a sampled report is perfectly adequate, and waiting for perfect data usually costs more than the imprecision ever would.