Machine translation post-editing (MTPE) has become the default workflow for most SaaS and FinTech companies localizing into Japanese. DeepL, ChatGPT, and Google Translate produce output that reads like Japanese. That is precisely the problem.
When translation looks fluent, reviewers relax. Errors that would be obvious in broken machine translation slip past because the overall text feels natural. But in Japanese, the gap between "fluent" and "trustworthy" is wide, and the errors that close that gap are exactly the ones MTPE reviewers most often miss.
This article describes a QA checklist built from 15+ years of Japanese MTPE review work. It focuses specifically on the issues that are invisible to AI review tools and easy to miss under time pressure.
Why Japanese MTPE Needs a Different QA Approach
Most MTPE QA frameworks were developed for European language pairs where the structural distance from English is smaller. Japanese presents unique QA challenges:
- Register flexibility: Japanese has multiple politeness levels (です/ます vs plain form vs keigo) and machine translation frequently mixes them within a single document.
- Subject dropping: Japanese commonly omits subjects; MT often over-inserts explicit subjects that read as redundant or robotic.
- Particle accuracy: Japanese particles (は、が、を、に、で) are grammatically critical but can be technically correct yet contextually wrong in ways that create subtle trust issues.
- Industry terminology: In FinTech, SaaS, and legal contexts, there are often two or three Japanese options for a single English term — and the wrong one, while understandable, signals poor domain expertise.
The MTPE Japanese QA Checklist
Layer 1: Register Consistency
Verify a single register is used throughout
Check that the entire document uses either です/ます (polite) or plain form consistently. MT frequently shifts register mid-paragraph, especially when source sentences vary in formality.
Check keigo usage in customer-facing UI
Notification messages, error screens, and help center text require respectful language (丁寧語). MT often generates plain or casual forms in these contexts.
Layer 2: Terminology Accuracy in Domain Contexts
Payment and billing terminology
Verify 決済/支払い distinction (see: "決済 vs 支払い" article), and check that 請求、課金、引き落とし are used in contextually correct ways. MT defaults to the most common translation, which is often wrong for B2B SaaS billing contexts.
Feature and UI terminology consistency
Build a termbase before MTPE begins. MT will translate "dashboard" as ダッシュボード in one sentence and 管理画面 in another. Both are correct; inconsistency is the problem.
Legal and compliance language
Privacy policy, terms of service, and billing disclosures require legally precise Japanese. MT output in these sections should be treated as a first draft requiring full human review, not light post-editing.
Layer 3: Redundancy and Over-Literalness
Check "以上/以下" expressions
MT frequently doubles these — "53以上以上" or "53+以上" — because it translates the English "+" and then adds the Japanese "以上" independently.
Check repeated subject insertion
Japanese naturally drops subjects in connected sentences. MT often re-inserts the subject in every sentence, creating an awkward, robotic tone that signals non-native production.
Layer 4: CTA and Conversion Copy
Critical QA zone: CTA buttons, signup prompts, and pricing CTAs receive the most user attention and generate the highest stakes errors. These require human review regardless of MT quality elsewhere on the page.
CTA register and action clarity
MT often produces CTAs that are grammatically correct but too passive. "無料で始められます" (You can start for free) is weaker than "無料で始める" (Start for free). In our QA engagements, the direct imperative form consistently converts better in Japanese B2B contexts.
Error messages and empty states
MT error messages are often too blunt or too vague. "エラーが発生しました" (An error occurred) provides no recovery path. Check that error messages include what went wrong and what to do next, in natural Japanese.
Layer 5: Visual and Format QA
Punctuation normalization
Japanese uses 「」 for quotation, 。for periods, and 、for commas, not their ASCII equivalents. MT output often mixes ASCII and fullwidth punctuation, which reads as unpolished.
Numeric and unit formatting
Japanese numeric conventions differ: large numbers use 万 (10,000) units, not commas every 3 digits in some contexts. Dates are written 年月日 order. MT frequently leaves Western formatting that looks out of place.
Text expansion and UI truncation
Japanese text is typically 1.1–1.3× the character length of English. Button labels, tab headings, and navigation items localized by MT often exceed their UI containers. Review in context, not just as text.
The QA Gap That Matters Most
AI-assisted review tools like built-in MT quality estimators can catch many of the above issues — but they struggle with the most commercially important category: trust register.
Trust register is whether the Japanese sounds like it was written by someone who understands the buyer's industry, level of formality expectations, and professional context. A payment platform's Japanese should sound like a bank. A developer tool's Japanese should sound like an experienced engineer wrote it. An HR SaaS product's Japanese should match the formal HR vocabulary Japanese professionals use.
No AI tool currently measures this. It requires a native Japanese reviewer with domain experience — which is exactly what structured MTPE QA provides.
When MTPE QA Is Not Enough
Light MTPE with QA review is appropriate for: help center articles, FAQs, blog content, and internal documentation where conversion stakes are lower.
For these content types, full human translation (or at minimum, heavy MTPE) is recommended regardless of MT fluency:
- Pricing pages and billing copy
- Legal and compliance documents
- Onboarding flows and signup screens
- Error messages in payment or authentication flows
- Sales and enterprise proposal materials
If your current workflow applies light MTPE to these content types, you are likely leaving conversion on the table. The trust gaps are invisible in your analytics but visible to every Japanese enterprise buyer who reads your product.