Data sampling is a procedure in which a tool does not process all recorded data but calculates an estimate from a portion of it. It is used with large volumes and long periods so that a report is generated quickly. The consequence is that the same number can differ slightly on repeated runs, and for small segments the estimate tends to be significantly inaccurate – precisely where the cause of a problem is most often being sought. In practice, check with every report whether it is sampled, and if so, shorten the period or narrow the scope. For high-stakes decisions, use an unsampled export. And never compare a sampled figure with an unsampled one; the difference may not reflect any real change. For regular reports it is therefore better to set shorter periods and aggregate them than to pull a long range in one go.
See also: Data quality in analytics, Statistical significance of a test, Modelled conversions.
