Multi-Language Instagram Inboxes: Keywords Are Literal

There is a comfortable assumption behind the word localization, and it is worth breaking early: that somewhere in the system there is a translation layer, and that a trigger written in English will quietly understand the same question asked in Hindi.

There is not. A keyword is a string. Matching compares strings. price matches the letters p-r-i-c-e, and it will sit there unmoved while somebody types कीमत, விலை, ధర, or kitna.

That single mechanical fact is the whole of this post, because almost every complaint that begins "the automation is not firing for half my audience" ends at it.

What the matcher does and does not fold away

Before comparing, the text is normalised. Three things get folded, and it is worth knowing exactly which three because people over-plan for these and under-plan for everything else.

What is not folded: script, language, spelling, and word endings. Those four are yours to enumerate.

So the list is long, and you write it by hand

A price trigger for an account with a Hindi, Tamil and English audience is not one keyword. It is a single any-of list containing every string that means the same thing:

price, cost, rate, कीमत, दाम, விலை, kitna, kitne, kitni, kitne ka, daam, how much

Notice what is in there. The scripts, obviously. But also the transliterations, which are the part people forget, and which on a lot of Indian accounts outnumber both the English and the Devanagari. Somebody typing Hindi on an English keyboard is not typing Hindi as far as the matcher is concerned. kitna is a Roman-alphabet string and must be listed as one.

Two details that save an afternoon:

Keywords match whole words. A keyword of price will not fire on pricing, and kitne will not fire on kitnee. Word endings are variants like any other and go in the list as their own entries.

Write Indic keywords exactly as people type them. Because vowel signs are preserved, कीमत matches कीमत and nothing else. That is the behaviour you want, and it means the spelling in your list has to be the spelling in the comment — including the matra. If your audience types the same word two ways, both go in the list.

One list to fire, several templates to answer

Getting the trigger to fire is half the job. The other half is that a Tamil speaker should not get an English answer.

There is no language detection and no automatic reply-language switching. What there is: automations are cheap, so build one per language. Same post, same intent, different keyword list, different reply text. The Hindi automation lists the Hindi and transliterated strings and answers in Hindi. The English one answers in English.

Two things make this less laborious than it sounds. Templated replies are free and unlimited, so a fourth language costs nothing to serve. And negative keywords are per-automation, so an intent that means one thing in one language and something else in another can be excluded in exactly the place it causes trouble.

Test each language automation on yourself with a comment typed in that language. Not an approximation of it — the actual string, from the actual keyboard, with the actual spelling your audience uses. The test result names which keyword matched, which turns a vague "it did not work" into a specific missing entry.

The escape hatch nobody uses

If a post's comments are overwhelmingly one intent — a giveaway, a "drop a word for the link" reel, a single-product launch — you can leave the keyword list empty. An automation with no keywords replies to every comment on the posts it is scoped to.

That is a blunt instrument and it is exactly right for a small number of posts. It sidesteps the entire language problem, because it never has to recognise anything. Scope it to specific posts, keep the negative keywords, and it will out-perform a carefully tuned multi-script list on the one reel where everybody is asking the same thing.

What this does not solve

Voice notes have no words to match, so nothing fires on them. Photographs likewise. If a meaningful part of your inbox arrives as audio or images, automation is covering less of your workload than the reply count suggests.

And the register still has to be yours in each language. A generated reply is shaped by the replies the account owner actually wrote, so a language you have never written in has nothing to learn from. For those, write the template yourself and leave it templated — which is free, and better.

For the mechanics of triggers in general, the keyword post covers them properly. For how messily people type even in one language, this one is about that. And the India post covers the rest of what changes when your audience is here.