Translating the ad is the easy part

When an ad goes into another language, four things break, and only one is the words. Translated copy overflows the character budgets the layout sets; image models may not draw the target script correctly; right-to-left markets need the composition mirrored, not just the text; and price, guarantee and health claims are regulated differently in each market.

Ask anyone who has taken a campaign international what the hard part was and almost nobody says the translation. The translation was fine. What broke was everything the translation touched.

We support 47 languages for ad creative, which has meant meeting each failure mode more than once. There are four, and only one of them is about words.

1. The copy no longer fits

Ad layouts run on character budgets, and they're tight. A headline gets a line or two; a benefit row gets a few words; a CTA button holds barely more than a phrase. Those limits aren't stylistic — they're what fits at a legible size in the space the design allows, and they're set by the layout, not by the writer.

English is unusually compact. German, Finnish and Polish routinely need a third more room to say the same thing; Romance languages sit somewhere in between. As a rough rule of thumb, the shorter the source string, the worse the expansion ratio — which is exactly backwards from what you want, because the shortest strings in an ad are the ones in the tightest boxes.

So "Get 20% off" fits your button beautifully. Its German equivalent doesn't, and now you have three bad options: shrink the type until it's illegible on a phone, let it wrap and break the button, or truncate mid-word.

The fix isn't translation, it's re-writing to a budget. The target-language copy has to be written to the same character count as the source, not translated and then squeezed. That means the constraint has to travel with the string — a translator (human or otherwise) handed "Get 20% off" with no budget attached will give you the accurate answer, not the one that fits.

2. The image model may not be able to draw the script

This is the one that surprises people, and it follows directly from the fact that image models don't really spell — they draw shapes that look like text.

That approximation degrades sharply outside the training distribution. Of the languages we support, 26 are Latin-script. The rest are not: four CJK, four Arabic-script, four Cyrillic, three Devanagari, and a tail of others. A model that renders English headlines almost perfectly can produce confident, well-kerned nonsense in Devanagari — and worse, you may not be able to tell.

Latin-script languages with diacritics are the sneaky middle case. Turkish, Czech, Vietnamese and Polish all look like "normal" text to a casual glance, and a dropped or wrong diacritic changes or destroys the word. It reads as fine to a non-speaker reviewing renders.

The practical consequence: the threshold for compositing real text instead of generating it drops as you move away from Latin script. For English you can often let the model draw the headline and check it. For Hindi or Japanese, draw the type with a font renderer and accept that it sits on top of the image.

And whatever you do, the reviewer has to be able to read the language. A check on the rendered text — automated or human — that can't read Arabic isn't checking Arabic.

3. Right-to-left isn't a text problem, it's a layout problem

Four of our languages are RTL: Arabic, Hebrew, Persian and Urdu. Teams tend to treat this as a rendering flag. It isn't.

In an RTL market the composition mirrors. Reading order runs right to left, so the visual hierarchy has to as well: the element you want seen first belongs on the right. A Before/After split has its "before" panel on the right. A product hero that sits left of a benefit list moves to the opposite side. Arrows, progress indicators and numbered steps all reverse direction.

What does not mirror: the product photograph, logos, and anything with embedded Latin text. Flip those and you get a mirrored logo, which reads as sloppy at best.

This is the failure mode where "just translate it" produces something that looks technically correct and feels wrong to every native reader. It needs a re-render, not a text swap.

4. The claim may not be legal there

Prices, guarantees, discount framing and health or finance claims are all regulated differently by market. "Money-back guarantee," "clinically proven," and a struck-through original price each carry rules that vary across the EU, let alone globally.

This is genuinely outside what any creative tool can solve for you. It's worth naming anyway, because it's the one that carries actual consequences and it's the easiest to forget when a nine-market rollout feels like a formatting exercise.

How should you localize ad creative?

The instinct is to treat localization as a post-production step: finish the ad, send the strings out, paste them back. That works for a webpage. It doesn't work for creative where the words are inside the image, because pasting strings back means re-rendering — and if you're re-rendering, the layout can and should adapt to the new copy rather than fighting it.

Which means the honest sequence is:

1. Lock the layout and the character budgets.
2. Write the target copy to those budgets, in-language.
3. Re-render the creative per language.
4. Have someone who reads the language check the render.

Step 4 is the one that gets cut, and it's the one that catches everything the other three miss.

The practical version

1. Translate to a character budget, not to a meaning. Send the constraint with the string.
2. Know which script you're asking for. Non-Latin scripts push you toward compositing real type rather than generating it.
3. Diacritics are a silent failure. Turkish and Czech look fine to people who don't read them.
4. RTL mirrors the layout, not just the text — and never the logo.
5. A reviewer who can't read the language isn't a reviewer.
6. Check the claim, not only the copy. Regulation varies by market.

None of this is a reason to skip localized creative. Running the same English ad everywhere is worse. It's a reason to budget for a re-render per market rather than a find-and-replace, which is a much smaller cost than it sounds like — and a much bigger one than most plans allow.

Sources

• Text size in translation — W3C Internationalization: average expansion by source-string length, highest for the shortest strings (accessed 2026-09-15)

---

Related: why AI image models still can't spell, and the sixteen layouts almost every performance ad uses — the layout is what sets the character budget this post keeps running into.

Terrads localizes into 47 languages by re-rendering the creative, not by pasting new strings over the old one.