AI Translation for Your Own Site: Bulk Multilingual Translation, Term Consistency, and Quality Checks
1. How AI translation differs from traditional machine translation
Traditional machine translation (MT) relies on a glossary for consistency: the same term gets the same translation on every page, at the cost of stiff sentences. Large models translate fluently, but each call is stateless — if you don’t tell it how to translate a term, it improvises.
This means that when you use a large model for a multilingual site, you can’t treat translation as a one-off black-box call — you have to put the glossary into the prompt. For how to build a glossary, see translation post-editing (MTPE), and after launch there are five more checks, see Multilingual Content Publishing Process.
| Dimension | Traditional MT + glossary | Direct large-model translation | The right combination |
|---|---|---|---|
| Fluency | Standard | Good | Good |
| Terminology consistency | High | Low | High |
| Number handling | Stable | Occasional drift | Caught by automated checks |
| Cultural Adaptation | Weak | Stronger | Backed up by human spot checks |
| Cost | Low | Billed per token | Spread thin by bulk runs |
2. Glossary first: start with 50 to 100 terms per language
术语库分三层：品牌与产品名（公司名、产品线、型号，绝不许翻）、行业术语（行业固定译法，如”compressor”固定译”压缩机”）、自定义口径（你自己定的说法，如”24小时响应”固定译法）。
For each language, start by compiling 50 to 100 entries into a bilingual table. When you translate, paste this table at the top of the prompt and write: “The following terms must be translated strictly according to the table and must not be replaced.” Get this one step right and term consistency jumps from a random 60% to over 95%.
举个例子：你的核心产品叫”环控冷水机”。术语表里这一项写死，中文”环控冷水机”、英文”precision chiller”、德文”Präzisionskühler”。
When the model hits this term it must output the specified translation and must not swap in “industrial water chiller” on its own. After each translation run, a script regex-matches this line and flags any variant in red. For how to lay out multilingual GEO, see Multilingual GEO.
| Glossary layers | Content | What to do |
|---|---|---|
| Brand and product names | Company names, product lines, model numbers | Leave as is, no translation |
| Industry terms | Compressors, heat exchangers, etc. | Fixed translation, no variants allowed |
| Custom glossary | Commitment lines such as “24-hour response” | Fixed translation, aligned sentence by sentence |
| Units and numbers | 380V, 68 dB | Leave as is, no conversion |
| Placeholders | {name}、{date} | Leave as is, no translation |
3. Batch translation pipeline: four steps from English to multiple languages
Four steps: prepare the English source files, run a first pass with the glossary attached, check automatically, then spot-check by hand. Split source files by page type: technical docs, marketing pages, FAQ templates, each with its own prompt. Technical docs need accuracy, marketing pages need flow, FAQs need a conversational tone.
Take a scenario: you have an English blog post to translate into German and Spanish. Paste the glossary into the prompt, add “keep the original table structure, don’t translate numbers or units,” and run both languages at once. When it’s done, scan with a script: did every term use the glossary translation, does every number match the source, and did any placeholder like {name} get mangled.
| Page type | Prompt focus | Sampling rate | Go-live requirement |
|---|---|---|---|
| Technical docs | Accurate, correct terms | 5% | Pass automated checks |
| Marketing homepage | Natural, with tone | 100% | Read through by a native speaker |
| FAQ templates | Conversational, short | 15% | Spot-check 15 entries against the source |
| Product spec pages | Zero number drift | 100% | Check every number one by one |
| Blog posts | Smooth and natural | 20% | Spot-check first and last paragraphs |
4. Three automated checks: terms, numbers, placeholders
The automated check does three things. First, term consistency: use regex to pull every industry term out of the translation and compare it against the glossary, flagging any variant not in the glossary. Second, zero number drift: extract every number from the source and the translation and compare them; one mismatch and it goes back.
Third, placeholder integrity: template variables like {name} and {date} must appear exactly as-is in the translation. All three are script jobs, under 10 seconds per article. Humans only review what the automated check flagged, plus a 10% to 20% read-through by page type.
| Check | How to check | If it fails |
|---|---|---|
| Term consistency | Extract terms with regex and compare | Flag it, fix it back to the glossary translation |
| Number consistency | Extract numbers from source and translation and compare | Send that section back for retranslation |
| Placeholder integrity | Template variables present as-is | Send that section back for retranslation |
| Length fit | German length overflow check | German often runs 30% longer, leave room |
| Encoding and characters | Scan for garbled text and special characters | Send back for retranslation |
5. Pitfalls: five types of content you shouldn’t machine translate
First, pricing and payment terms, where one wrong number means a contract dispute; second, certification names and regulatory clauses, since machine translation flattens legal text; third, brand names and slogans, which need localization, not translation; fourth, culturally sensitive metaphors and puns, which can offend local readers if translated literally.
The fifth category is form buttons and error messages, where a one- or two-word mistranslation can directly cost you conversion. Use human translation or native-speaker review for these; don’t hand everything to machine translation just to save money. For how to rank market priorities, see Market Priority for Going Global.
| Content type | Why machine translation won’t work | The right way |
|---|---|---|
| Pricing and payment terms | A wrong number means a wrong contract | Human, line by line |
| Certification and regulatory clauses | Legal text loses its meaning | Native-speaking lawyer / translation reviewer |
| Brand slogans | They need localized copywriting | Write two options and let a local pick |
| Cultural metaphors | A literal translation offends readers | Translate the meaning or cut it |
| Form button copy | A mistranslation hurts conversion | Build a termbase by hand |
FAQ
Why can’t you just throw the full English text into an LLM and translate it into multiple languages?
How many terms should the termbase start with?
How do you set the manual spot-check rate?
German translations always stretch the page out of shape. What do I do?
What content is worth paying a native reviewer for instead of using machine translation?
One person maintaining 3 language versions: how do you schedule it?
Final thoughts
AI translation has driven the cost of “getting it translated” to nearly zero, so the real cost has shifted to consistency and accuracy. Whoever builds a termbase and automated QA first turns their multilingual site from “usable” into “trustworthy” — and trustworthy is exactly what it takes to get cited by AI engines. This doesn’t require new tools; a glossary and a QA script are enough.
Three steps to get there: week one, put together a 50-term glossary; week two, get the automated QA script running; week three, take one existing blog post through the whole pipeline. Don’t translate your entire site right away — get one language and one page type working smoothly first, then expand.
There’s an order to expanding: start with FAQs and technical docs, since those two translate most reliably by machine; then blogs; and only last your marketing homepage and pricing terms, because those two need people most.GEO Practice In a multilingual setting, it comes down to using consistent facts and terminology so AI in every language can cite you.
One line worth remembering: the large model handles fluency, the termbase handles consistency, and people handle the fallback. Miss any one of the three, and your multilingual site is merely “translated” rather than “usable.”