Manage localized copy across many A/B tests

WHAT THIS TESTS: It evaluates whether you can architect a translation-management system and reason about statistical validity when per-locale samples are wildly unequal and sometimes tiny.

A GOOD ANSWER COVERS: Represent copy as keys mapping to per-locale, versioned strings in a translation-management system. Code references stable keys, never literal text, so variants and locales are data rather than code. The translation workflow tracks each string's state per language, untranslated, machine-translated, human-translated, reviewed, with fallback to a base language when a translation is missing, plus versioning so you can roll back and audit changes. For the experiment layer, assign variants centrally and log the locale on every event. The hard part is statistics: high-traffic languages reach significance quickly, but small locales may never accumulate enough samples for an independent test. Address this by pooling related locales, or better, hierarchical or Bayesian partial-pooling models that share information across languages so a small locale borrows strength from the global effect while still allowing genuine per-locale deviation.

COMMON WRONG ANSWERS: Hardcoding translated strings directly into templates. Treating each language as a fully independent test, so rare languages produce noisy results you nonetheless act on. No versioning or fallback for missing translations, so users see broken pages. Ignoring multiple-comparisons inflation across hundreds of concurrent tests.

LIKELY FOLLOW-UPS: How exactly does partial pooling help small samples. How do you handle a missing translation at serve time. How do you control the false-positive rate across hundreds of simultaneous tests. How do you version and safely roll back copy that turned out worse.

ONE CONCRETE EXAMPLE: A headline test runs in thirty languages. English reaches significance in days, but Finnish has too few visitors to decide alone. A hierarchical model estimates the global lift and shrinks Finnish's noisy estimate toward it, yielding a stable directional read, while English still gets its own confident verdict. Any missing translation falls back to English copy at serve time, so no user ever sees an empty or broken string while the test runs.