Our US product copy was genuinely good. Punchy, conversational, a little irreverent — the kind of voice a startup spends two years and one expensive agency finding. When we launched in Mexico, we translated it. We kept the structure, swapped the words. 'Game changer' became 'cambiador de juego.' 'Knock it out of the park' became 'sacarla del parque.' Both technically correct. Both used by exactly nobody who actually speaks Mexican Spanish. We weren't offensive. We were worse than offensive — we were invisible. Nothing landed because nothing sounded like a human.
Translation asks 'what does this mean.' Localization asks 'what would they say.'
Those are completely different questions and I didn't understand that until our Mexico City contractor read our copy and made a face I could feel through the screen. 'Knock it out of the park' works in the US because baseball is culturally loaded — the metaphor carries decades of shared feeling. In Mexico, fútbol is the loaded thing. The equivalent gut-punch metaphor involves scoring a goal, not hitting a home run, and a literal translation of the baseball line just reads as a baffling sports reference from a country that isn't yours.
So I ran our launch copy through the Localize Idioms and Slang for a Target Culture prompt with the source set as 'American casual startup voice', the target as Mexican Spanish, and the content type as social ads. I used Grok 4.3 deliberately — this is a job about what people actually say right now, casual and current, and Grok's punchy, contemporary voice is well-suited to slang and informal register where the dictionary-correct answer is the wrong answer.
The idiom analysis is a hit list of everything I got wrong
Twelve expressions came back flagged. 'Touch base' — common in US business English, sounds strange translated into Mexican Spanish, better as 'ponernos al tanto.' 'Low-hanging fruit' — understood literally, lands as nonsense idiomatically. 'Game changer' had a natural equivalent that wasn't 'cambiador de juego', which it turns out is something only a translation engine has ever produced. Each one came with what it means in our culture and what a real person would say in theirs. But the part I now read first is the CULTURAL INSIGHT at the bottom: it flagged that Mexican business Spanish tends toward warmer, more relational framing than US copy's transactional efficiency — that our 'let's get straight to it' energy reads as slightly cold, and a touch more warmth up front would do more than any single phrase swap.
- Idiom-by-idiom analysis tells you exactly what to change and why, so you're not guessing
- The CULTURAL INSIGHT section is usually the most valuable output — it changes the whole tone, not one line
- The WATCH OUT flags caught one reference that was neutral for us and genuinely off-putting there
- A literal translation of a culturally loaded metaphor doesn't offend — it just makes you invisible, which is worse
- Specifying the content type ('social ad') gets you slang that fits the channel, not a generic register
Why Grok, and why 'what people actually say today' is a real instruction
I picked Grok 4.3 for this and I want to explain why, because it's not arbitrary. Slang has a shelf life. The 'natural equivalent' of an idiom from five years ago can already sound like a substitute teacher trying to be cool, and a model that leans textbook will confidently hand you a phrase that's technically idiomatic and socially dead. The prompt explicitly tells the model to favor what people actually say today over the dictionary form, and Grok's punchy, current register is the right instrument for that note. When I'm doing careful, register-preserving translation of something formal, I reach for a different model entirely — but for the question 'what would a 28-year-old in Mexico City actually say in an Instagram caption,' you want the model that's least precious about formality.
The one place I'm careful: I always run the localized output past a human native speaker before it ships, because slang is also the highest-risk register. The WATCH OUT section flags the obvious landmines — a phrase that's neutral for us and loaded for them — but a model can't know that a particular word blew up on local social media last month. The prompt gets me 90% of the way and shows me exactly where to point the human reviewer, which is the same triage logic that makes the whole approach worth the money.
What the numbers did
Engagement on the relocalized Mexican copy ran meaningfully higher than on the literal translation — enough that we stopped treating localization as a translation line item and started treating it as its own pass. We ran the same process for Colombia and Argentina, and here's the lesson that surprised me most: the cultural insights were different every single time. Argentine Spanish wanted a different register again; the Colombian version flagged a slang term that was fine in Mexico and weird in Bogotá. When I just need a faithful, register-preserving translation rather than a full slang adaptation, I reach for my Translate and Preserve the Tone, Not Just the Words prompt instead — but for copy that lives or dies on sounding like a local, this is the one.
Grab the Localize Idioms and Slang for a Target Culture prompt on Prompt Dock, run your splashiest, most idiom-heavy ad through it, and read the cultural insight twice. The thing you'll change isn't a phrase. It's the temperature.