Quality
Unsampled data
Also called 100 percent data.
Unsampled data is a report built from every collected hit in the query. It is slower and exact.
How it is measured
The query reads all rows in the period instead of a share. Counts are exact, up to what the collector received.
Tarsier reads from Cloudflare Analytics Engine, so counts are exact until write volume gets high enough for the store to sample. Check this on very busy sites.
Worked example
A marketing manager needs the exact number of visitors from a podcast link on one Thursday. The sampled report, built from 9 percent of sessions, estimates 480. The unsampled query over every session returns 412.
The 68-visitor gap is the sampling error on a small segment. She reports 412 to the sponsor and notes that sampled figures for slices this narrow can be off by double digits.
How it differs
Unsampled data uses every hit. Sampling uses a share. A hit is the unit. Unsampled does not mean complete, because blocked and lost hits are still missing.
Common errors
Assuming unsampled equals truth. Mixing with sampled numbers. Ignoring blockers. Waiting for slow queries. Using it for tiny segments without need.
In practice
Use unsampled for small segments and for anything you report to someone paying.