If you're evaluating or using translation platforms, you know you need your brand's terminology preserved. But does your current platform do it? The best translation platforms can absolutely support brand terminology: they should enforce your terms through glossaries, do-not-translate rules, translation memory, and style guides. It should also include automated QA checks and provide in-context human review that then catch the terms that slip through.
You've probably seen what happens without those controls. A product name gets translated in one market and left alone in another, or a tagline shifts meaning in German but stays intact in French. Imagine if Nike's iconic "Just Do It" tagline ended up as "Hazlo justamente", meaning “Do it fairly.”
It mistakes “just” for “fairly,” losing Nike’s intended message: stop hesitating and take action.
Localize's 2019 glossary article cites industry estimates that about 15% of localization spend goes to redoing translations, with inconsistent terminology as the leading cause.
The stakes go beyond redoing work. In CSA Research B2B findings from March 2022, 66% of 1,116 B2B tech buyers said they'd pay up to 30% more for a localized product. CSA notes this reflects stated willingness and is largely hypothetical, since localized products usually cost the same.
That stated value only holds if your localized product reads like your brand in every language. For marketing teams, that means the same product names and taglines on every regional page.
Our marketing localization guide covers the broader strategy. This article focuses on one idea: treat terminology controls and QA as a single system you evaluate together, because each depends on the other.
Glossaries, term bases, and do-not-translate rules
A glossary in a translation platform defines how and when to translate each term. It tells a translator, or an AI engine, that "workspace" has one approved rendering in Spanish and that your product name stays in English everywhere.
The strongest setups apply the same glossary to AI engines and human translators. If your machine translation ignores the glossary while your linguists follow it, you end up paying reviewers to fix errors the system could have prevented.
What belongs in a brand glossary
Start with the terms that carry your brand: product names, taglines, slogans, and trademarks. Then add acronyms, approved translations for confusing terms, and a do-not-translate (DNT) list, as outlined in our article on the benefits of a translation glossary.
Your glossary will outlive any single vendor, so ask how it moves between tools. ISO publishes the ISO 30042 TBX standard for exchanging terminology data, and it names translation as an intended application. A practical vendor question is whether the platform imports and exports term bases in TBX.
Do-not-translate rules for brand names
Some terms must stay identical in every language, such as your company name and product names. DNT rules make that intent explicit so nobody has to guess, whether the translator is a person or a model.
On the web, the W3C ITS 2.0 specification from W3C defines Translate and Terminology data categories for marking this kind of content. In HTML, the native translate attribute carries do-not-translate intent, but each tool decides whether to honor it. Check that any platform you evaluate respects that markup instead of assuming it does.
Platform-level rules give you more direct control. Localize's Same as Source glossary setting, released in April 2025, marks glossary terms to remain unchanged across all languages automatically.
Keeping glossaries current at scale
Glossary management is manual work, and it's easy to deprioritize when launches pile up. Meanwhile, your product team keeps shipping new feature names, and your glossary falls behind.
Localize's Agentic Glossary Terms, released in June 2026, scans your content and recommends high-impact terminology as glossary terms. You review the suggestions instead of hunting for terms yourself.
According to the Localize glossary documentation, the glossary can provide terms to human translators and supported AI engines. Glossaries can be project-specific or shared across projects, and bulk import and export help when you're migrating an existing term list.
Translation memory and style guides
Glossaries govern individual terms, while translation memory (TM) stores whole approved segments for reuse. When a new sentence matches a stored one exactly or closely, the platform offers the prior translation as an exact or fuzzy match. Our explainer on how translation memory works walks through the mechanics.
A style guide covers what neither tool can capture on its own: voice, formality, punctuation, and formatting conventions for each market. Localize's evaluation article points out that these different "learning" mechanisms are not interchangeable, so a vendor that has one hasn't automatically solved the others.
Localize Advanced Translation Memory surfaces similar prior translations and shows who translated them. That context helps reviewers judge whether a past translation still fits your current brand voice. It also tells you who to ask when a term's history is unclear.
The most revealing question is what happens to an approved edit. Ask vendors whether machine translation enters TM only after human approval. Then ask whether an update to a stored translation corrects existing published content, or only affects future matches.
Which localization providers offer quality assurance?
Most established localization platforms and agencies claim to offer quality assurance. The differences show up in how they define and route checks. They also differ in how they verify that a fix actually reached the published page.
That's why you should judge each provider by its QA mechanisms. The four areas below show what to look for and what to ask.
Automated QA checks
Automated checks scan translations for terminology errors, inconsistencies, broken placeholders, and formatting problems. They catch mechanical misses at a volume no reviewer could match.
To compare scores across vendors, you need a shared vocabulary. The MQM Council's MQM error typology defines seven high-level error dimensions, including Terminology and Style. MQM is a community framework, and it means glossary and style-guide violations become scoreable error classes.
The same W3C ITS 2.0 specification also defines a Localization Quality Issue data category for recording the type and severity of errors found during QA. Its issue types were derived from early MQM work.
With AI Translation Quality Scoring, Localize automatically scores every translation. With Translation QA enabled, AI translations wait in Needs Review and publish once they pass review.
Still, don't accept a score as a substitute for a standard your team defined. Ask vendors what the score measures and who sets the threshold, and read our guide to machine translation quality metrics before you compare numbers.
Human review and post-editing
High-impact content still needs human review. Pricing pages and campaign headlines carry too much risk to publish on an automated score alone.
ISO's ISO 18587 post-editing standard covers full, human post-editing of machine translation output and the competences post-editors need. It's a useful reference when a vendor describes its post-editing process. ISO lists the 2017 edition as under revision, so check which version a vendor cites.
You don't have to post-edit everything, though. Localize's MT metrics article explains that sampling and escalation rules usually outperform full post-editing, because reviewers spend time where errors are most likely.
Localize pairs AI translation with human review in one workflow. Our guide to AI translation explains how to balance speed and review depth.
In-context review
A term that looks right in a spreadsheet can break on the page. Reviewers need to see translations where users see them, including button labels and dynamic content that only appears after a click.
The Localize On-Page Editor lets reviewers select phrases on a live page and preview edits before publishing. Each edit moves through states such as Needs Review, Rejected, and Published, and roles separate who can approve from who can edit.
Review shouldn't stop at launch. The Localize Feedback Widget lets real users suggest phrase-level improvements to translations on the page. Your team then reviews those suggestions in the dashboard, so a term that doesn't land gets caught.
Certifications and standards
Certifications help you judge the human side of a vendor's QA. ISO's ISO 17100 translation requirements set requirements for the core processes and resources translation service providers need to deliver a quality service.
The standard governs translation service providers and their processes. A software platform on its own falls outside its scope, so ask which of a vendor's language partners hold ISO 17100 and request their certificates.
Keep the standard's limits in mind as you read vendor claims. ISO states that raw machine translation plus post-editing is outside the scope of ISO 17100:2015, and the 2015 edition is now under revision.
Evaluation checklist for brand terminology and QA
Use this table to compare vendors side by side. Each row ties a terminology control to the QA step that verifies it.
A checklist tells you what a vendor claims. A pilot shows you what actually happens, so run one before you sign.
Pick test content that stresses both terminology and QA. Each of these exposes a different weak point:
- A high-conversion page, such as pricing or a signup flow
- A dynamic UI state, like an error message or modal
- Content dense with brand terminology and product names
- Layout-constrained content, such as buttons and navigation labels
Then measure results you can compare across vendors. Use the same content and languages for every pilot:
- Reviewer time spent per page or per language
- Defects found during review and after publishing
- TM match rates (exact and fuzzy) alongside glossary coverage
Ask each vendor to report those coverage numbers for your pilot content. For a broader framework, see our article on evaluating AI translation platforms.
