The localisation project ran for two quarters and shipped eleven languages. Nine thousand strings, professionally translated, reviewed by native speakers, with a proper pluralisation library. It was competent work and I would defend every rupee of it.
A user who switches the app to Hindi now reads Hindi throughout — and then reaches the search box, and types in English letters, because that is what is on the keyboard he actually has and because transliteration is what his thumbs know.
We measured strings translated. That was the metric on the slide.
Input was somebody else's problem, in the sense that it belonged to no team. The keyboard is the operating system's. The transliteration engine is the keyboard's. What arrives at our search box is a string of Latin characters that is neither English nor Hindi but a phonetic guess, spelled four different ways by four different people, and our search index was built on properly spelled Devanagari because that is what the localisation project produced.
So the more thoroughly we localised the reading, the worse the matching got, and for about seven months the Hindi interface had lower search success than the English one. Nobody had a chart with that on it, because search success was owned by the search team and language was owned by the localisation programme, and the join between them was a user.
I have now seen versions of this at three companies and the pattern is always the same: localisation is scoped as a translation exercise because translation is the part that is countable. Strings are a unit. You can burn down a backlog of strings, report percentage complete, and finish. Input, sorting, name formats, address formats, dates, the fact that a name may have one word in it, and the way people actually type — none of those is a countable backlog, so none of them is in the project.
The counterargument, and I have made it in planning meetings myself, is that translation genuinely is the largest single win and that a partial job beats no job. Someone who can read the app in their own language will use it more than someone who cannot, and waiting to solve input before shipping any of it would have meant shipping none of it for another year. That is right, and it is why I would run the project again.
But the sequencing argument only holds if the rest arrives afterwards, and at all three companies it did not, because the programme closed. A programme with a metric closes when the metric is met. There was no phase two, not because anybody decided against it, but because the thing that would have justified a phase two — a number showing that Hindi users search worse — did not exist and could not be produced by either of the two teams that would have had to produce it together.
What eventually fixed ours was not a project. An engineer added transliteration matching to the search index over about three days because it annoyed him. Hindi search success moved by a large enough margin that somebody asked what had happened, and by then he had already shipped it.
His name is not on any slide about localisation, and if you look at the project retrospective the eleven languages are a success, which they are, and the thing that made them work is not in it.
