Linguistic QA is expensive, so teams sample. Sampling is fine. Convenience sampling is how a language becomes “clean” while the unreviewed tail holds the defects that will hit the store screenshot.
We see three recurring biases. First, vendors nominate files they already internally checked. Second, reviewers pick short UI strings because they are fast, leaving help articles untouched. Third, severity rubrics treat every preference as a blocker, which paradoxically makes teams stop sampling at all because the board is unusable.
A workable method from the lab: draw at random from the frozen set, record a small severity list that product has agreed in advance, and publish the sample size next to the score. A 2% defect rate on eight strings is not the same artefact as a 2% rate on eighty. Localization Performance Analytics without a denominator is theatre.
Random does not mean reckless. You may stratify — UI versus legal versus marketing — as long as the strata are declared. What you may not do is let last week’s “we already looked at onboarding” become a permanent exemption.
None of this replaces in-country review for regulated copy. It stops you from claiming a health score you did not earn.