Why AI Translation Gets Your Brand Name Wrong (And How to Fix It)

You ran your product through AI translation. It came back fast, looked mostly fine, and then a customer in Germany emailed asking why your product was called something completely different from what it said on your website. This is the most common — and most avoidable — failure mode in AI translation. And it's not a model problem. It's a glossary problem.

What Actually Goes Wrong

AI translation models are probabilistic. Given the same input, they don't always produce the same output. More importantly, they have no idea what your brand name means, what your product is called, or which technical terms you've specifically decided to use (or not use). So when your model encounters a brand name in a German context, it might translate it, leave it as-is, or do something completely different on the next run. There's no guarantee.

The same problem hits product names (translated into the local language, destroying the brand), deprecated features (old names translated consistently even after you retired them), industry terms (translated differently across segments), and forbidden phrases (legal terms your team explicitly decided not to use).

Probabilistic vs Deterministic: Why It Matters

Traditional human translators with a style guide are more consistent than AI — not because they're smarter, but because they read the style guide. They know which terms to leave untranslated, which to always translate the same way, which to avoid. AI translation models don't get a style guide by default. They get a prompt and a string. They do their best with available context, and their best is 'statistically likely to be correct' — not 'definitely matches your brand standards.' This is the gap: probabilistic translation (AI's default) vs. deterministic enforcement (what enterprise content requires).

The Fix: Translation Glossaries

A translation glossary is a list of terms that tells the AI exactly what to do with specific words and phrases. Not 'suggest' — enforce. When the AI runs translation with an active glossary, it's no longer guessing what to do with these terms. The output is deterministic — the same every time, across every segment, every run.

How to Build Your First Glossary in 30 Minutes

You don't need to start comprehensive. Start with the terms that matter most and expand from there.

Step 1: List your non-translatable terms (10 minutes)

These are brand names, product names, feature names, and legal identifiers that should never be translated. Go through your product and list every proper noun: company name, product names, feature names, integration names (Slack, GitHub, etc.), and any acronym that's part of your brand identity.

Step 2: List your forbidden terms (5 minutes)

These are words or phrases you've explicitly decided not to use — deprecated names, competitor references, legally sensitive terms. Check with marketing and legal for their no-go list.

Step 3: List your consistency terms (15 minutes)

These are terms that can be translated but need to translate the same way every time. Look at your existing translated content and find where the same English term has been translated differently in different places. Pick the right one, add it to the glossary. You now have the 80% that covers most quality issues.

Glossary vs Translation Memory: What's the Difference?

These two are often confused. They solve different problems. Translation memory (TM) stores previously approved translations and reuses them — it's reactive, learning from past work. Glossary defines rules for specific terms upfront, regardless of whether those strings have been translated before — it's proactive, enforcing standards before the translation runs. You need both: TM saves time and money on repeated strings, while glossary enforces brand standards on every string, new or repeated.

When Glossaries Aren't Enough

• **Brand voice.** A glossary can tell the model what words to use. It can't tell the model to be funny, or warm, or direct. Key marketing content still needs human review by someone who speaks the target language natively.
• **Humor and wordplay.** Puns don't translate. Idioms don't translate. Cultural references don't translate. No glossary entry fixes this — it requires a human who can find the equivalent that works in the target culture.
• **Context-dependent terms.** Some words have multiple correct translations depending on context. A single glossary rule for 'account' in a banking context vs. a social media context will get one of them wrong.

FAQ

What happens if a term isn't in my glossary?

The model translates it as best it can based on context. This is fine for most general vocabulary. The risk is with specialized or brand-specific terms — which is why building the glossary around those terms specifically is worth the investment.

How many terms should a glossary have?

Start with 20–50 terms covering brand names, product names, and your most-used technical terms. Glossaries with 400+ entries can cause quality issues by overriding too much of the model's natural judgment. Keep it focused.

Does every language need its own glossary?

Yes. A brand name might be treated the same in every language (don't translate), but a term like 'translation memory' has a different preferred translation in German, Japanese, French, and Spanish. Language-specific glossaries give you the control you need.

Can I use a glossary with any AI translation tool?

Not all tools support glossary enforcement. Some only support glossary suggestions (shown to human reviewers), not automatic enforcement. When evaluating tools, specifically ask whether glossaries are enforced at the model level or only surfaced as suggestions.