The most awkward multilingual-site failures are rarely whole pages left untranslated. They appear just as someone is about to click: a Chinese validation message on an English form, a different price, date, or link in the other language. Manual browsing catches some of these, but omissions become inevitable as content and releases grow.
AI can help with localization QA, but “it looks translated” is not a useful standard. A dependable review first exposes verifiable differences in titles, buttons, form errors, SEO metadata, links, numbers, dates, and product terms. Only then should AI assess meaning and language fit. This is a lightweight workflow for Chinese-English sites using ChatGPT or Claude.
Start with URL pairs, not a request to read the whole site
Build a URL-pair sheet. Each row contains a Chinese URL, its English counterpart, page type, and owner: tool detail, category list, article, checkout, help page, or email template. Do not assume every path can be paired by replacing a locale prefix. Some tools are Chinese-only and some articles have no English edition; label these as intentionally unavailable rather than treating them as defects.
For each pair, capture four things: title and description, visible text, link targets, and structured data or important numbers. Record the date and whether the capture is production or staging, otherwise a later difference may be impossible to attribute to translation or a new feature.
Handle missing translation and wrong translation separately
Localization QA often mixes two distinct jobs. A missing translation is source-language text that should have been localized. A wrong translation already has target-language text, but its meaning, tone, or context is wrong. They need different owners and priorities.
- Missing text: navigation, buttons, empty states, validation messages, modals, email subjects, alt text, cookie notices, and footers.
- Factual mismatch: price, quota, date, feature availability, platform support, version, or legal link.
- Term inconsistency: one feature is “workspace” on one page and “space” on another; the same action becomes Create, Build, and Generate.
- Tone or order: literal phrasing, overly long buttons, double negatives, mixed punctuation, or inconsistent title case.
Do not ask AI only whether text sounds natural. Require evidence: URL, location, source, current translation, issue type, severity, and proposed revision. “Could be improved” without a location rarely enters a real fix queue.
Compare the following Chinese and English page text. Report only verifiable issues: source-language strings left untranslated, opposite meaning or factual mismatch, inconsistent terminology, or links and button instructions likely to block a task. For every item give the URL, area, source text, current translation, issue type, risk level, and recommendation. Do not count product names, code, URLs, versions, or explicitly non-translated terms as omissions. Use “needs human confirmation” when text alone is insufficient.
Use rules to narrow the work, then use AI for semantic judgement
AI does not need to do everything. Simple checks can find high-value anomalies: long runs of Chinese characters on an English page, an entire untranslated English paragraph on a Chinese page, a title and h1 that do not match, a link that points to the other locale, or changed currency, dates, percentages, and versions. Rules provide recall; AI explains context and helps sort the results.
This is particularly effective after each release. Check changed URLs and adjacent pages instead of rescanning all historical content. Keep approved product names, brand terms, and code fragments in an allow-list so valid strings such as “API Key” do not create repeat alerts.
A glossary needs context, not just preferred words
A practical glossary contains a source term, preferred target term, allowed variants, discouraged variants, context, and an example sentence. The value is not its size but its context. “Workspace” may be a Chinese product term for everyday users while remaining in English in developer documentation; “Plan” can mean a subscription tier or a project plan. Without the context, AI can only substitute mechanically.
Add one glossary rule when a real error is found. Do not batch-replace a word across the whole site after one style discussion. Validate the phrasing in a few high-traffic pages first, then extend the rule.
Review samples by severity
Not every imperfect sentence should block a release. A practical severity system is:
- P0: incorrect payment, permission, deletion, privacy, security, or legal information; fix or take down immediately.
- P1: interferes with registration, purchase, download, a core task, or search indexing; fix in the current release window.
- P2: terminology or meaning problems on core product pages; schedule in the next content-fix batch.
- P3: style, punctuation, or minor grammar that does not impede understanding; collect and resolve together.
AI can assign an initial label, but people must review P0 and P1. For compliance, pricing, and availability, the model is never the source of truth.
Include SEO and non-visible text
Invisible content often reveals localization gaps first: title, meta description, Open Graph, canonical, hreflang, structured data, sitemaps, and image alt text. Chinese and English pages each need a fitting title and description, not a Chinese title dropped into English. If a tool supports only Chinese, the English directory should hide it or fill its position correctly; it should not produce an empty English detail page.
Also test whether a locale switch reaches equivalent content rather than always returning to the home page. When no translation exists, a clear alternative route is more respectful than a forced redirect.
A 15-minute pre-release checklist
- Are all changed URLs in the pairing sheet?
- Do text sources include buttons, errors, empty states, and modals?
- Do prices, dates, versions, and links match the source page?
- Are remaining source-language strings either fixed or allow-listed?
- Are high-frequency glossary terms consistent on key pages?
- Do locale switches, canonical tags, and hreflang target the correct pages?
- Have P0/P1 items been confirmed and regression-tested by an owner?
AI localization QA does not replace translators. It turns issues scattered across browsing, search, and screenshots into tasks with evidence, location, and priority. Let rules find suspicious items, use AI to explain the context, and let people who understand the product verify facts. That is how a bilingual experience stays dependable without an oversized maintenance burden.