Great product translation starts with prepared source copy

Source copy carries the burden of translation quality

Product copy is often written quickly in one language and then handed to translation as if meaning were self-evident. Most translation problems begin earlier, in the source. A sentence with two possible readings produces two possible translations. A label that omits what the action applies to forces the translator to guess. A button that says Continue on one screen and Next on another creates inconsistency that no review can fully repair. The first step in a translation workflow is a writing habit rather than a tool: decide what each piece of copy means, say it plainly, and say it the same way twice.

Short complete sentences travel well across languages, and concrete nouns carry more meaning than abstractions. Keeping one term for one idea is a small effort that saves many hours of revision later. It also helps your own team, because reviewers can tell a wording problem from a translation problem instead of debating both at once.

Agree on a handful of source conventions before the first request goes out:

• Sentence case for interface labels and headings.
• One voice and one tense for product messages.
• A practical length limit for buttons, tabs, and navigation.
• Placeholders that name the data they stand for, so no language has to restructure the sentence around them.

These rules are not bureaucracy. They reduce ambiguity for people and for any automated step you add later, because well-formed source is easier to segment, match against previous work, and check for length.

A glossary keeps product meaning consistent across markets

A glossary is simply a list of your product's important words with an agreed translation and a note on when to use it. It covers feature names, button verbs, error categories, units, and the terms that appear in support articles. Without one, the same English word can come back as three different words in a single language, and readers start to wonder whether the product has three different features.

Build the glossary from real copy rather than from theory. Collect terms from the screens customers see most, mark which ones are fixed brand names that should not be translated, and record a short reason for each decision. A reason matters, because the person who maintains the glossary in a year will not remember the discussion behind it.

Glossaries work best when they are boring and boringly maintained. Give one person ownership, review new terms as they appear, and let translators flag a term instead of silently improvising. Treat disagreement as information: if two competent translators cannot agree, the source term is probably unclear.

Style notes belong beside the glossary. They describe register, formality, punctuation habits, and how to handle humor, gendered language, or polite forms in each market. Languages differ in how much formality they demand, and a literal translation of a friendly English tone can sound careless in one language or overly familiar in another.

Translators need context beyond isolated strings

Modern tools let you send strings one at a time, and that efficiency has a cost: translators see a fragment and not the screen it lives on. The word Delete in a menu means something different from Delete in a confirmation dialog, and only context resolves the difference.

Give every string three things: where it appears, what the user is trying to do, and any constraint such as length limits. A short note saying this appears beneath a price and must fit one line on mobile is worth more than a paragraph of generic instructions.

Screenshots help, but plain descriptions age better and travel further across teams.

• Name the surface: onboarding, checkout, billing, settings, or error states.
• State the action the user takes immediately before and after the string.
• Flag strings that are reused in several places.
• Explain any wordplay, idiom, or cultural reference that cannot survive a literal transfer.

Context also shapes word order and grammar. Many languages need more room than English, some need less, and several require agreements that English leaves invisible. A translator who knows the surrounding sentence can keep the grammar coherent, which is why reassembling fragments blindly produces copy that is correct word by word and wrong as a whole.

Review passes catch errors before customers do

Translation quality is not a single moment of judgment but a sequence of checks, each with a different job. Separate them so that no reviewer is asked to fix meaning, spelling, and layout at the same time.

• A linguistic review by a native speaker who checks meaning, grammar, and tone against the source.
• A functional review inside the real product, where someone clicks through the flow and reads the copy in order.
• A final check for consistency, punctuation, capitalization, placeholders, and truncated text on the narrowest supported screen.

Keep the reviewer independent of the translator when the stakes are high, and keep the feedback loop short. A comment that says this sounds unnatural in the billing screen is useful; a comment that says rewrite everything is not.

Measure something. Track first-pass acceptance, the number of strings returned with questions, and the defects found after release. Over a few releases these numbers reveal where the process breaks: if questions cluster around one feature, the source copy there needs work; if defects cluster in one language, the review step there needs attention.

Error reports should name the string, the language, the expected meaning, and the observed text. That turns a vague complaint into a change the next release can absorb.

Automation supports the process without owning it

Machine translation and large language models can draft quickly, and they are genuinely useful for bulk content, internal notes, and first passes on low-risk strings. They are less reliable where wording carries brand, legal, or emotional weight, and they fail most quietly when the source itself is ambiguous.

A sensible setup gives the machine the same information you would give a human: glossary terms, style notes, length limits, and the context described above. Ask for consistent terminology rather than creative phrasing, and keep a person in the loop for anything a customer reads at a decision point.

Teams that automate often route requests through an API relay, which forwards each request to the official model named in the request rather than substituting another one. That detail matters for translation work, because quality comparisons depend on knowing exactly which model produced a draft. This article is published by BestRelay, a quality-first relay covering Claude, GPT, Gemini, DeepSeek, Kimi, GLM, Qwen, Grok and more, which works with Claude Code, Codex, Gemini CLI and any OpenAI SDK. Billing is pay-as-you-go from the token counts the provider reports, with cache reads and writes priced separately, and members earn a referral commission when people they invite top up. For a quick first pass on low-risk strings, Gemini 2.5 Flash is worth evaluating alongside your existing review steps.

Whatever tool you choose, keep the workflow human-readable. Someone should be able to see which strings were machine-drafted, which were edited, who approved them, and what changed between versions. When a customer reports a wording problem, that history is the difference between a five-minute fix and an afternoon of archaeology.

Start small: pick one product surface, write its source copy carefully, build a glossary of twenty terms, and run the review passes on it. Create a BestRelay account when you are ready to try the drafting step.