What is the difference between translation and transcreation?

Translation moves the words. Transcreation moves the effect, which sometimes means the sentence in the new language shares almost nothing with the original. Break a leg translated literally is an insult. A transcreator writes whatever that market says to wish someone luck. Product documentation wants translation. Headlines, taglines and ad copy want transcreation, and they are priced differently because they are different work.

We already ran it through DeepL. Can you just check it?

Yes, and that is now the most common version of this job. It is called post-editing and it is priced below new translation, which is part of why the per-word range is as wide as it is. What you are paying for is the class of error a machine cannot see — idiom, register, cultural fit, and legal wording — rather than the class it handles well, which is ordinary sentences.

Why is a native review non-negotiable?

Because the failure mode is silent. Bad marketing translation does not look broken to anyone who cannot read the language, so nothing alerts you. The first signal is a customer telling you a line at checkout reads as rude, by which point months of traffic have seen it. A native reader of that market signs off before anything goes live, and it is priced into the work rather than offered as an upgrade.

Do you handle app strings and website files, not just documents?

Yes. JSON, PO and XLIFF files are handled in their own format with the keys and placeholders left intact, which matters because a translated placeholder breaks the build rather than the sentence. Website copy goes back into the CMS in the structure it came from.

What about legal text and terms of service?

Regulated and legal wording is looked at separately from marketing copy and is not covered at the marketing rate. Ad rules differ sharply by market — what is an ordinary comparative claim in one country is prohibited in another — and that review needs someone reading for compliance, not for tone.

How is the per-word rate calculated?

On the source word count, agreed before work starts, after the source text is frozen. Language pair, subject difficulty and whether it is new translation or post-editing decide where in the range it falls. Repeated content across a site is discounted through translation memory, so paying twice for the same product description is avoided.

Content, Social & Communications

Translation and transcreation

Marketing copy rebuilt so it lands the same way in the new market — with a glossary first, a separate pass for dates and currency, and a native reader before anything ships.

From: $0.04–$0.15/word
Typical time: ongoing
Usually bought by: Exporters

What this actually solves

A company that sells well at home ships its website into a new market and the numbers do not follow. The copy is grammatically correct. Nothing looks broken. It simply does not persuade anyone, because persuasion is made of idiom, rhythm and cultural assumption, and none of that survives a word-for-word move.

Sometimes it is worse than flat. A joke that works in one market is baffling in another. A comparative claim that is routine advertising in one country is a regulatory problem in the next. A politeness register that reads as friendly and direct at home reads as presumptuous or even rude somewhere else. The brand does not find out from analytics. It finds out from a customer, months later, if at all.

The honest description of this market in 2026 is that machine translation did not end the work, it changed what the work is. Clients now produce the first draft themselves in seconds. What they are buying is the second pass — the errors the machine cannot see, because the machine has no view on whether a sentence sounds like something a person in Munich or Riyadh would actually say, or whether the claim in it is legal there.

What we do, and in what order

Scope the source and freeze it. Source word count agreed, file formats identified, register and audience defined, and the source text locked. Copy that keeps changing during translation is how three versions of a page end up live in different languages, and the freeze is the cheapest possible prevention. Two to three days.
Build the glossary and the do-not-translate list. Product names, feature names, brand vocabulary, and the terms that must be identical every time they appear. This is the first deliverable and it is also fed directly into the model prompt when a machine draft is involved, which measurably reduces drift. Without it, one feature acquires three different names across a site and both support and search break. Two to three days.
Draft and post-edit. Either a new translation or a post-edit of a machine draft, depending on what was agreed. The post-editing pass is where idiom is rebuilt, register is corrected, and anything that reads as a translation is rewritten until it reads as though it was written there. Every deviation from the source is annotated with what it originally said and why it changed. Scales with the word count and is scheduled with it — a page set runs in a few days, a full site and app in two to three weeks.
Run the locale pass separately. Dates, currency, units of measurement, address shape, phone number format, name order and number separators — checked as their own pass because they are a different kind of error from a language error and get missed when the two are done together. The date format alone is a live hazard — 03/04/2026 means the fourth of March in the US and the third of April in the UK, and one unconverted date is a wrong delivery promise on every product page. About a day per language.
Native review and sign-off. A native reader of that market goes through the finished copy for anything that reads wrong, sounds automated, or would cause a problem locally, and signs it off. Nothing is delivered without this step. Three to five days.
Deliver in your own format, with the memory. Back into the CMS, or as JSON, PO or XLIFF with keys and placeholders intact, plus the translation memory and glossary handed over so the next batch is cheaper and consistent with this one. One to two days.

Tools — DeepL, which is stronger than the general engines across more pairs; Claude or ChatGPT for transcreation and register work; Smartcat or MateCat for translation memory and glossary enforcement; and your own CMS or localisation files for delivery.

What moves the price

Work is priced per word at $0.04–$0.15/word, and the relationship is usually ongoing, because product copy, pricing pages and app strings keep changing after a market launches. Three things decide where a job falls inside that range.

The language pair. English into German or Japanese sits at the top of the range. Pairs with a deeper pool of qualified marketing writers sit lower. This is supply, not difficulty of the text.

Subject matter. General marketing copy is one rate. Medical, financial and legal content is another, because a mistranslation there is not an embarrassment but a liability, and the review is correspondingly heavier.

New translation or post-editing. Editing a machine draft is priced below translating from nothing, and which one you need is agreed before the quote rather than discovered halfway through. Repeated segments across a large site are discounted through translation memory.

What usually goes wrong

AI output ships without a native reader. This is the most expensive mistake in the whole service, and the reason is that it is invisible from the inside. The copy looks fine to everybody on your side, because nobody on your side can read it. Nothing fails, nothing errors, traffic keeps arriving, and the first indication of trouble is a customer writing in to say that the line above the checkout button reads as impolite. By then it has been live for months. If you cannot personally read the output and catch the error, it does not ship without someone who can.

The words are translated and the locale is not. A perfectly natural German paragraph sitting above an American date format, a price in the wrong currency symbol, weights in pounds for a metric market, or an address form that demands a state. Each is small and each tells the buyer this page was made for somebody else. They are missed because they are not language errors, which is exactly why they get their own pass.

No glossary is agreed, so terminology drifts. The same feature is rendered three different ways across the homepage, the pricing page and the app, because three different passes made three reasonable choices. The customer cannot tell whether these are the same thing, internal search stops matching, and support answers questions about a feature name that does not appear in the interface. The glossary is written before any body copy, and it is enforced by the translation memory rather than by memory.

See where the site stands before you brief anyone.

If the site is already live, start from what it scores rather than from a guess. The audit reports what it checked, what it found, and — the part most tools leave out — what it could not check, and why.

See the website audit Talk about this work

Questions people ask before they commit.

What is the difference between translation and transcreation?: Translation moves the words. Transcreation moves the effect, which sometimes means the sentence in the new language shares almost nothing with the original. Break a leg translated literally is an insult. A transcreator writes whatever that market says to wish someone luck. Product documentation wants translation. Headlines, taglines and ad copy want transcreation, and they are priced differently because they are different work.
We already ran it through DeepL. Can you just check it?: Yes, and that is now the most common version of this job. It is called post-editing and it is priced below new translation, which is part of why the per-word range is as wide as it is. What you are paying for is the class of error a machine cannot see — idiom, register, cultural fit, and legal wording — rather than the class it handles well, which is ordinary sentences.
Why is a native review non-negotiable?: Because the failure mode is silent. Bad marketing translation does not look broken to anyone who cannot read the language, so nothing alerts you. The first signal is a customer telling you a line at checkout reads as rude, by which point months of traffic have seen it. A native reader of that market signs off before anything goes live, and it is priced into the work rather than offered as an upgrade.
Do you handle app strings and website files, not just documents?: Yes. JSON, PO and XLIFF files are handled in their own format with the keys and placeholders left intact, which matters because a translated placeholder breaks the build rather than the sentence. Website copy goes back into the CMS in the structure it came from.
What about legal text and terms of service?: Regulated and legal wording is looked at separately from marketing copy and is not covered at the marketing rate. Ad rules differ sharply by market — what is an ordinary comparative claim in one country is prohibited in another — and that review needs someone reading for compliance, not for tone.
How is the per-word rate calculated?: On the source word count, agreed before work starts, after the source text is frozen. Language pair, subject difficulty and whether it is new translation or post-editing decide where in the range it falls. Repeated content across a site is discounted through translation memory, so paying twice for the same product description is avoided.

Often taken together.

Start with the brief, not the invoice.

Send what you are trying to achieve. You get a written scope, a price and a date — or an honest “this is not the right service for that”.

Talk about Translation & transcreation All Content, Social & Communications services