# Machine-translation QA flag (September 11, 2025 QRG) Google's January 23, 2025 Quality Rater Guidelines update added explicit language about **machine-translated content with no human review** as a form of scaled content abuse (§4.6.5). This is also the **enforceable webmaster-facing spam policy** (updated 2026-05-15), which names "automated transformations like synonymizing, **translating**, or other obfuscation techniques" as scaled content abuse: https://developers.google.com/search/docs/essentials/spam-policies > "Using automated tools (generative AI or otherwise) as a low-effort > way to produce many pages that add little-to-no value." Machine-translated content is fine **when reviewed by a human speaker and corrected**. Untranslated MT — or "lightly post-edited" output that still contains hallucinated terms, wrong gender/number agreement, or untranslated proper nouns — is treated as scaled content abuse. ## Signals the seo-hreflang audit should surface | Signal | Severity | Notes | |---|---|---| | Multiple `hreflang` alternates point to URLs whose content is identical except for header chrome | Critical | Indicates the body wasn't translated; just template wrapped. | | `lang="xx"` attribute on `` doesn't match the body language | High | Translation pipeline output without a final QA step. | | Auto-translated `` longer than 160 chars (untrimmed) | Medium | The translator overran the snippet limit — no human reviewer caught it. | | `lang` attribute is `auto` or missing entirely | Medium | Pages that don't declare their language confuse hreflang + AI crawlers. | | Untranslated proper nouns or product names sprinkled in body | Low (heuristic) | Common MT failure mode; hard to detect automatically. | | Schema.org `inLanguage` field absent or wrong | Medium | Multi-language audits should cross-check `inLanguage` vs. body. | ## What the audit should NOT flag - Sites that have a few MT pages clearly labelled as MT (a human-translation-fallback pattern). Google's QRG explicitly permits MT *when honestly labelled and clearly scoped*. - Machine-translated UI strings — those are i18n, not "content". - Content with `lang="auto"` if the audit can't fetch a fallback signal (be conservative; don't claim what we can't verify). ## Cross-skill delegation - For per-page hreflang validation, stay inside `seo-hreflang`. - For broader scaled-content scoring (entropy of translated pages, AI-pattern detection in body), defer to `seo-content` via `"${CLAUDE_PLUGIN_ROOT}/scripts/claude-seo" run content_quality.py`. - For "is this translated by Google's own auto-translate widget" detection, look for the `.goog-te-banner-frame` iframe. There is **no per-mechanism exemption**: in June 2025 Google removed the robots.txt-blocking guidance for auto-translated pages, and auto/AI- translated content is now judged **case-by-case on user value**. Use page-level `noindex` for low-value translated pages, not sitewide blocks; the widget also produces poor passage-citability. ## Primary sources - Current QRG (September 11, 2025; 182 pages; nothing newer through 2026-07-09): https://services.google.com/fh/files/misc/hsw-sqrg.pdf - John Mueller on MT (multiple SOTR episodes, 2024-2025): MT is OK when reviewed by a human; bulk MT without review is abuse. Last verified: 2026-07-09.