Which Languages Should You Localize First?
A method instead of a list, with the two halves most advice skips. Opens by naming where every top-10 ranking comes from — the T-Index, whose public table is labelled 2018 with a 2021 projection and whose inputs put United States internet penetration at 76.2% and China’s at 53.2%, so the articles quoting it as current are passing along numbers nobody has refreshed in years. Then the ratio that is the entire business case: English is the content language of 49.5% of websites and the language of roughly 26% of internet users, which is why competition thins out sharply the moment you leave it. The centre of the article is a demand model you run against your own traffic — parse Accept-Language with q-weights rather than splitting on commas, compare each language’s conversion rate against your source-language baseline, and rank by sessions x conversion gap x order value. Executed against an illustrative English-first SaaS dataset it puts German seventh despite German being the second-biggest market by revenue, because German speakers already convert at 86% of baseline; it puts Japanese fourth on 2.2% of sessions, because the worst conversion index in the table multiplied by the highest order value beats languages with four times the traffic; and it values Dutch and Swedish at a combined $663, since both already convert at 98% of baseline. The second half prices the engineering nobody quotes: CLDR 49 cardinal plural categories per language (1 for Japanese and Chinese, 4 for Russian and Polish, 6 for Arabic and Welsh — and 3 for Spanish and French, which stopped being two-form languages when CLDR added many), the W3C expansion table where the shortest strings expand most, real English→German UI strings where Undo becomes Rückgängig machen at +325%, RTL as a UI project rather than a translation one, and executed Node output showing why Turkish breaks case conversion. Plus the variant decisions that must happen before the first word is translated, a four-tier commitment model, five documented ways this goes wrong, and the argument for making your first language a deliberately cheap one because the first launch is a rehearsal for a pipeline that does not exist yet.