Multilingual Content Translation for Your Own Site: MTPE Workflow, Glossary, QA, and Publishing
1. Why raw machine translation can’t go live
Machine translation is already very good on general text, scoring above 80 BLEU. But content on a B2B company site is not general text. Product specs, regulatory names, MOQ, lead times, certification numbers — the MT model has neither your glossary nor your industry context. One wrong term can lead an overseas buyer to misread your specs.
Counterintuitively, raw MT that reads smoothly is more dangerous than MT that reads awkwardly. Buyers notice awkward text is a translation and read it with suspicion; fluent text with wrong terms gets taken at face value, and wrong specs go straight into purchasing decisions. This is the same call as Where AI Collaboration Fits in Content Production: machines handle volume, people handle the critical facts.
| Content type | Risk of raw MT | Recommended fix |
|---|---|---|
| Product spec sheets | Units, precision, and tolerances mistranslated | Strict glossary rules plus native-speaker proofreading |
| FAQ | Wrong tone, questions and answers don’t match | MT plus native-speaker polish |
| Long blog posts | Style drifts, long sentences fall apart | MT plus paragraph-by-paragraph editing |
| Pricing and lead times | Numbers shifted, currency symbols dropped | Protect placeholders; don’t machine-translate |
| Privacy policies and compliance text | Ambiguous legal wording | Handled separately by a native legal translator |
2. The five stages of the MTPE workflow
Run the five MTPE stages in order: preprocessing, MT engine translation, post-editing, QA, publishing. Preprocessing is the stage people skip most often, yet it determines how much rework the next four stages need. Wrap prices, lead times, MOQ, and model numbers in placeholders — for example, replace “MOQ: 500 pcs” with “[MOQ]” before sending it to MT — so the engine can’t split numbers from their units.
Post-editing is not rewriting. The industry splits PE into light PE (fixing errors and awkward phrasing, roughly 20% to 30% edit rate) and full PE (rewriting for native expression, 50%+ edit rate). Light PE is enough for product pages on a B2B company site; brand stories and About pages should get full PE. See the detailed scoring table at Post-Editing (MTPE): Glossary, Spot-Check Rates, and Scoring Table.
| Stage | Action | Quantitative thresholds | Common rework points |
|---|---|---|---|
| Pre-processing | Segment sentences, tag placeholders, protect code snippets | 100% placeholder coverage | Missed price placeholders cause mistranslated numbers |
| MT engine | Choose DeepL, Google, or a custom model | Use one engine per language | Mixing two engines in one document |
| Post-editing | Light edits by a native translator | Edit rate 20% to 30% | Too many edits turn into a rewrite |
| QA | Terminology, formatting, numbers, placeholders | Terminology consistency ≥98% | Missed check leads to lost placeholders |
| Launch | CMS backfill, hreflang setup | Spot-check another 20% before publishing | Missing hreflang for a language version |
3. How to build a terminology base, and which 200 terms to start with
Bigger isn’t better when it comes to a terminology base. To start, 200 to 300 core terms are enough to cover 80% of translation errors. Pick terms from four categories: brand and product names (fixed translations required), technical parameter units (mm, inch, tolerance), regulatory and certification names (the source wording for CE, UL, REACH), and high-frequency industry verbs (quote, lead time, MOQ, sample). With 50 to 80 terms per category, you can finish in three months.
Example: an industrial valve exporter built its first terminology base with 260 terms. “ball valve” is fixed as “Kugelhahn” in German customer manuals, but machine translation defaults to “Kugelventil,” which refers to a different type of valve in that market. Without this entry, German buyers would open the page and think they were looking at the wrong model. Write the terminology base as machine-readable CSV or TBX so you can feed it to CAT tools and MT engines as a hard constraint. For the localization boundaries of charts and images on multilingual sites, see Charts and tables on multilingual sites.
| Term category | Starting count | Example terms | Maintenance frequency |
|---|---|---|---|
| Brand and product names | 50 to 80 | Company name, product lines, model series | Update when new products launch |
| Technical parameter units | 60 to 80 | mm, bar, tolerance, IP rating | Review quarterly |
| Regulatory and certification names | 40 to 60 | Source wording for CE, UL, ISO | Update when regulations change |
| Transactional terms | 50 to 70 | MOQ, lead time, quote, FOB | Review quarterly |
4. The three QA gates and sampling rates
QA isn’t reading through once to catch typos. It’s three structured gates. Gate one: terminology consistency. Use a CAT tool to check whether the terms in the translation match the terminology base, targeting ≥98%. Gate two: placeholder integrity. Check that [MOQ], [price], and [model] in the source each have a match in the translation; any missing placeholder means the batch goes back. Gate three: numbers and units. Pull every number from the translation and compare it against the source, and verify unit conversions separately.
Sampling rates adjust dynamically by batch. The first language version goes live with 100% full review. After three consecutive batches with zero terminology errors, drop to 50%, and after two more stable batches, drop to 20%. If sampling turns up a terminology error, send the whole batch back and return the rate to the previous level. For the post-launch publishing workflow, see Multilingual content publishing workflow: 5 gates after translation is complete.
| QA gate | Check tool | Pass threshold | On failure |
|---|---|---|---|
| Terminology consistency | CAT tool terminology comparison | ≥98% match rate | Send the whole batch back to post-editing |
| Placeholder integrity | Script-based regex extraction and comparison | 0 missing, 0 misaligned | Locate sentences and retranslate |
| Numbers and units | Script-based number extraction plus manual spot checks | 0 mistranslations | Check the parameter table sentence by sentence |
| Formatting and layout | Browser walkthrough | No broken layout, no truncation | Adjust fonts and length to fit |
5. Rollout pace and language priority
Don’t launch 10 language versions at once. For a B2B standalone site, roll out languages in batches by market priority: English first (if you haven’t yet), then German, Spanish and French, the three European languages with the most search volume, followed by Japanese, Korean and Arabic for the Middle East. After each batch goes live, let it run a full 4 weeks and check inquiry volume and bounce rate before deciding on the next batch. For how to set market priority, see Market Priority for Going Global.
Example: a solar mounting company launches English and German first. In the third week after the German version goes live, inquiries from Germany are 35% higher than the English version over the same period, but the bounce rate is also 8 percentage points higher. On review, the German product pages are 40% longer than the English ones, and the tables get cut off on mobile. After switching the tables to a responsive layout, the bounce rate falls back to normal. Problems like this only surface after a real launch; QA during translation won’t catch them.
| Batch | Language version | Observation period | Criteria to start the next batch |
|---|---|---|---|
| First batch | English | 4 weeks | Cost per inquiry below the English baseline |
| Second batch | German, Spanish | 4 weeks | Terminology error rate<2%，跳出率不高于英语版本+5pp |
| Third batch | French, Japanese | 4-6 weeks | Inquiry share meets expectations for that language market |
| Later batches | Arabic, Portuguese, etc. | Decide by ROI | Launch only after the previous batch turns ROI-positive |
FAQ
How much more does MTPE cost than raw machine translation, and is it worth it?
How big does a terminology base need to be, and how do you start one?
Should you check 100% of the translation, and how do you set the spot-check ratio?
Can numbers like price, lead time and MOQ be machine translated directly?
How many language versions should we launch at once?
How do we choose between DeepL, Google Translate, and our own model?
Final thoughts
The core tension in multilingual translation: machine translation is already fast and cheap enough, but what B2B buyers want isn’t “roughly understanding it” — it’s “specs I can trust.” The MTPE process separates “fast” from “accurate”: machines handle volume, people guard the critical facts. Get these three things right — the terminology base, placeholders, and three quality gates — and the translation quality of your multilingual site won’t run into major problems.
Three steps to launch: week one, build a 200-term terminology base; week two, pick one product page and run it through the full MTPE process end to end; week three, publish the first batch of translations with a 20% spot-check rate. Don’t try to launch every language version at once — moving in small steps beats rolling everything out in one go.
One line worth remembering: publishing raw machine translation on a multilingual site looks like it saves on translation costs, but in reality it hands the authority to explain your product specs to an intern who knows nothing about the industry.