Marketing in Indian languages: why translation fails

Populr writes marketing in 32 languages, 12 of them Indian. The easy way to build that is to write everything in English and run it through translation. We chose not to, and the reasons are most of what we know about marketing in Indian languages.

Translated copy reads as translated

Translation is faithful to words. Marketing has to be faithful to how people talk, and those are different targets. A translated post keeps English sentence order, swaps idioms for their literal meaning, and lands in a register nobody uses with a shopkeeper they know.

Register is what gives it away fastest. Hindi has three levels of "you" — tu, tum and aap — and choosing one is choosing a relationship. A bakery writing to its regulars and a bank writing to account holders should not sound alike, and an English source sentence carries no information about which one it meant. Translation has to guess. It usually guesses formal, which is how a neighbourhood shop ends up sounding like an insurance circular.

People do not write in one language

Much of Indian social media is written in more than one language at once: Hindi in Latin script, English nouns inside Marathi sentences, a Tamil post with an English call to action. That is not an error to be cleaned up. It is how the audience writes, and copy that refuses to mix reads like a brochure.

Script is a decision too. Some audiences read Devanagari comfortably on a phone but search in Latin letters; others are the reverse. There is no universal right answer, which is exactly why a pipeline that silently picks one will be wrong for someone.

The limits are not where you expect

A platform that limits characters counts the characters that are stored, not the letters you see. A Devanagari word that looks like three syllables can be seven characters to a counter. क्षेत्र (kshetra, "area") is stored as क, ्, ष, े, त, ्, र: consonants joined by a virama, plus a vowel sign, each counted on its own.

So posts in Indian scripts reach length limits sooner than their visible length suggests, and copy trimmed by eye gets cut mid-word by the platform. Count the way the machine counts.

What writing natively actually took

Writing natively meant using a model trained on Indian languages, and that model taught us something we did not expect. Our first attempt handed it the same detailed brief our English writer gets — around two thousand tokens of rules, examples and structure. It came back with nothing at all.

It was a reasoning model: it spends part of its output budget thinking before it writes, and a long, dense brief gave it a great deal to think about. In our measurements it needed roughly six thousand tokens of output budget to finish, and we had allowed four. It ran out while still reasoning and never wrote the post.

Two changes fixed it, and both were needed. A compact brief of about 280 tokens that keeps the language, audience, brand voice, platform limits and the rule against inventing facts, and drops the long craft guidance. And a larger output budget for that model alone. The English path was left byte-for-byte unchanged, because it was working.

The general lesson outlasts the specific bug: the instructions that make one model better can make another fail outright. A multilingual product cannot have one prompt. It needs a brief per model, tested per model.

If you are writing in an Indian language yourself

• Write in the language from the first word. Translating a finished English post is where the stiffness comes from.
• Choose the register on purpose — who is speaking, and to whom — before the first sentence.
• Match the script your audience types and searches in, not the one that looks most formal.
• Check length in the platform's own composer rather than by eye.
• Have a native speaker read it aloud. Where they stumble, your audience will too.

Common questions