Analytics tools handling large volumes often compute reports from a portion of the available data rather than all of it. The estimates that result behave in a specific and predictable way.
Sampling is a response to query cost
Processing every recorded event for every report would be slow and expensive, so platforms select a subset once a query exceeds a threshold.
The threshold depends on the date range, the complexity of the segments applied and the account's service level, which is why the same report samples on some days and not others.
Most tools indicate when sampling has occurred, though the notice is frequently subtle enough to be missed by people reading the numbers.
Accuracy depends on how common the thing is
A metric describing a large share of traffic is estimated accurately from a modest sample, because the subset closely resembles the whole.
A metric describing something rare is estimated from very few underlying records, and the margin of error around it can exceed the value itself.
This is why sampled reports are dependable for overall traffic and unreliable for narrow segments, which is exactly where analysts tend to look.
Extrapolation makes small numbers look solid
Sampled figures are scaled up to represent the full population, so the report presents a whole number rather than an estimate with uncertainty attached.
A conversion count derived from a handful of sampled events appears in the interface with the same authority as one derived from millions.
Decisions made on those scaled figures can therefore rest on a few records without anyone involved being aware of it.
Segmentation multiplies the problem
Each filter applied reduces the number of sampled records supporting the result, and several filters combined can leave almost nothing behind the number shown.
Comparisons between two heavily filtered segments are especially unstable, because both sides carry wide uncertainty that the interface does not display.
Differences that appear meaningful in such comparisons frequently disappear when the same query is run unsampled.
The workarounds trade convenience for reliability
Shortening the date range often brings a query under the threshold, at the cost of losing the longer view the analysis needed.
Exporting raw event data to a separate warehouse removes sampling entirely and moves the processing cost onto the organisation instead.
The judgement is about which reports carry consequences. Routine monitoring tolerates estimation comfortably, while decisions about budget allocation deserve figures computed from everything available.