Multilingual content with AI translation: what ranks and what gets filtered

TL;DR: Machine-translated content ranks fine when it adds real value to the target market and fails when it doesn’t; Google’s own spam policy names “translating” as a technique that becomes scaled content abuse only when a page provides “little to no value to users.” Raw output from Google Translate holds meaning correctly in roughly 82.5% of cases with a range of 55% to 94% depending on the language pair, so the quality bar shifts by market, not by whether AI touched the text.

---

Table of contents

1. The short answer
2. What Google’s policy actually says about translated content
3. Where machine translation is accurate and where it breaks
4. Quality thresholds by market
5. Raw MT vs MTPE vs full localization
6. Hreflang: the technical layer most sites get wrong
7. How AI search engines treat translated pages
8. The decision tree: should you translate this page
9. What actually gets filtered
10. Building a multilingual content workflow that survives review
11. Measuring performance per market
12. Frequently asked questions
13. Key takeaways

The short answer

Google does not have a translation penalty. It has a value penalty, and translation is one of the ways pages fail that test.

The Google Search spam policies define scaled content abuse as producing “many pages generated for the primary purpose of manipulating search rankings and not helping users,” and it lists the exact mechanism this article is about: “scraping feeds, search results, or other content to generate many pages (including through automated transformations like synonymizing, translating, or other obfuscation techniques), where little value is provided to users.” Read that sentence closely. Translation sits in the same clause as synonymizing and obfuscation, but only as one method among several for producing pages with “little value.” A translated page with real value from the same clause is not the target.

That single distinction decides whether your multilingual expansion works. A German page translated from a strong English original, checked for terminology, and matched to how German buyers actually search behaves like any other content asset. A German page auto-generated in bulk from a sitemap crawl, published untouched, and never checked against German search intent behaves like the scaled abuse pattern Google names directly.

What Google’s policy actually says about translated content

Three separate Google Search Central documents bear on this, and none of them treat “AI-generated” or “machine-translated” as automatically disqualifying.

The helpful content guidance asks whether “the use of automation, including AI-generation” is “self-evident to visitors through disclosures or in other ways,” and asks creators to explain “why automation or AI was seen as useful to produce content.” It does not ask sites to avoid automation. It asks them to be honest about using it and to still clear the same quality bar as manually written content: original value, clear expertise, and no “spelling or stylistic issues.”

The spam policies page frames the violation as using “generative AI tools or other similar tools to generate many pages without adding value for users,” which is a scale-and-value problem, not a tooling problem.

The multi-regional and multilingual site management guide treats multilingual sites as a normal, expected structure, and even tells you how Google determines a page’s language: “Google uses the visible content of your page to determine its language,” independent of any code-level language attribute you set. That single line matters for AI-translated pages specifically. If your translation output is grammatically broken enough that Google’s language detection misreads it, you have a quality problem regardless of what the HTML lang tag says.

Put together, the policy reads as: translate freely, disclose if it’s material to trust, and make sure the result reads like a page a native speaker in that market would want.

Where machine translation is accurate and where it breaks

Machine translation quality is not one number. It is a range that depends on language pair, domain, and how much post-editing happens after the raw output.

A study reported by localization platform Weglot found that Google Translate “preserved the overall meaning for 82.5% of the translations,” with accuracy spanning “55% to 94%” depending on the language pair. That range is the number to internalize: an 82.5% average hides a 39-point spread between the best and worst language pairs. The same article notes a much older domain-specific study that put translation accuracy for medical phrases at 57.7%, and casual English-to-other-language translation at 72%, both far below the general-purpose average. Neural machine translation, which most engines shifted to around 2016, cut translation errors by “more than 55%-85%” across major language pairs compared to the phrase-based systems it replaced, per the same source.

Two patterns hold across nearly every language pair study available:

• Accuracy drops with domain specificity. General marketing copy translates more cleanly than legal, medical, or highly technical content, because general vocabulary has more training data and fewer terms with multiple correct translations depending on context.
• Accuracy drops with linguistic distance from English. Romance and Germanic languages (Spanish, French, German, Portuguese, Italian, Dutch) consistently score higher in comparative MT studies than languages with different sentence structure, script, or honorific systems (Japanese, Korean, Arabic, Thai).

The same Weglot article cites a stat worth sitting with if you are scaling translation without a review step: “99% of translation work produced in the world is not done by professional human translators,” and “an average of only 30% of machine-translated content is edited” before publication. Most translated content on the internet today is unedited machine output. That is exactly the population Google’s scaled content abuse policy is written to catch when it adds no value, and exactly why the 70% of unedited MT output is where ranking problems concentrate.

Quality thresholds by market

The quality bar you need to clear is not fixed. It moves with three variables: how far the target language sits from your source language, how competitive the local search results already are, and how much the topic touches YMYL (Your Money or Your Life) territory.

High-tolerance markets for raw or lightly-edited MT: Spanish, French, German, Portuguese, Italian, and Dutch, for general commercial and informational content. These language pairs consistently post the highest accuracy scores in the studies above, the search results in these markets are used to seeing competent MT and human writing side by side, and buyer intent research ports over from English with modest adjustment. Straight product descriptions, help docs, and blog content in these languages can launch with raw MT plus a terminology pass and a native speaker spot-check, not a full rewrite.

Medium-tolerance markets requiring MTPE (machine translation post-editing): Japanese, Korean, Arabic, Russian, and Polish. Sentence structure, honorific systems, and right-to-left text handling introduce errors raw MT does not catch on its own. A human editor working from the machine draft, rather than translating from scratch, is the workable middle ground, faster than pure human translation and materially more accurate than unedited output.

Zero-tolerance markets regardless of language: any YMYL content, in any language. Medical, legal, financial, and safety-related content needs a qualified human reviewer before publication no matter how good the source-language content or the MT engine’s general accuracy score is, because a 57.7% meaning-preservation rate in a medical context (the domain-specific figure cited above) is not a rounding error, it is a compliance and trust failure.

The market-by-market rule of thumb: the closer your target language sits to your source language structurally, and the less the topic touches health, money, or legal risk, the more raw MT output you can tolerate before a human needs to look at it.

Approach	Cost per word	Turnaround	SEO risk	Best fit
Raw MT, unedited	✓ Lowest	✓ Fastest	✗ High for YMYL, medium for general	Internal docs, low-stakes UGC
MT + terminology/glossary pass	✓ Low	✓ Fast	✓ Low for close language pairs	General commercial content, Romance/Germanic markets
MTPE (human edits MT draft)	Medium	Medium	✓ Low across most markets	Most commercial and informational content
Full human localization	✗ Highest	✗ Slowest	✓ Lowest	YMYL, brand flagship pages, legal copy

Raw MT vs MTPE vs full localization

The distinction between translation and localization is not cosmetic. Semrush’s international SEO guide draws the line clearly: translation converts text while keeping meaning, localization adapts the message to cultural preferences and regional context, and the two produce different search performance because they solve different problems. A literal translation of “free shipping over $50” is useless in a market that shops in euros and expects free shipping thresholds tied to local price psychology, not a currency-converted number.

Full localization also means re-researching keywords per market rather than translating your English keyword list. “Lentes” and “gafas” both mean glasses in Spanish, but usage splits by country, exactly the kind of regional variation the Semrush guide flags as a reason generic translation underperforms even when it is grammatically perfect. If you skip this step, an otherwise flawless translation targets the wrong query in the target market’s actual search behavior, which shows up as low click-through rate even after the page ranks.

A workable rule: MTPE for anything you expect to compete for commercial queries, full localization for money pages and anything explicitly local (currency, units, legal terms, cultural references), and raw MT reserved for support content where speed matters more than polish and the downside of an awkward sentence is low.

Hreflang: the technical layer most sites get wrong

Translation quality is necessary but not sufficient. Ahrefs analyzed hreflang implementations across 374,756 domains and found 67% contained errors. That is not a rare misconfiguration, it is the majority default. The most common failures: missing self-reference annotations, invalid language or locale codes, references to non-canonical URLs, and missing x-default tags. Google’s own John Mueller called hreflang “one of the most complex aspects of SEO (if not the most complex one)… feels as easy as a meta-tag, but it gets really hard quickly” in a public statement cited in the same Ahrefs analysis.

The core mechanics, per Google’s own localized-versions documentation:

• Every language version must list itself and all other versions. Google calls this bidirectional, and it will disregard annotations that are not reciprocated.
• URLs must be fully-qualified, including the protocol; //example.com/de/ is invalid, https://example.com/de/ is not.
• Language codes follow ISO 639-1, with an optional ISO 3166-1 Alpha 2 region code, so de-CH is valid and a bare country code like EU or UK is not.
• x-default catches users whose language or region does not match any listed variant, and it is the recommended target for a language selector page.

For URL structure, Google’s multi-regional site guide ranks subdirectories (example.com/de/) as low-maintenance with a single host to manage, country-code domains (example.de) as the clearest geotargeting signal but the most expensive to run, and URL parameters (site.com?loc=de) as explicitly “not recommended” because they are hard to segment cleanly. The guide is equally direct about geotargeting: don’t redirect automatically based on perceived language or IP, because “IP location analysis is difficult and generally not reliable.” Offer a link to switch language versions instead and let the visitor choose.

None of this is specific to AI-translated content. It is the same technical layer every multilingual site needs, and it fails at the same 67% rate whether the underlying copy came from a translator or an engine. Get hreflang wrong and even a perfectly localized page can show the wrong language version in search results, which looks like a content problem but is a markup problem. See our international SEO and hreflang guide for the full implementation walkthrough.

How AI search engines treat translated pages

The newer question, separate from Google’s organic ranking behavior, is how AI answer engines treat translated content when deciding what to cite.

A Weglot study analyzing 1.3 million citations across Google AI Overviews and ChatGPT, built on 34,992 queries across 236 Spanish and Mexican websites, found translated sites received “24% more total citations per prompt” than untranslated sites. The effect was not symmetric: English citations rose 33% after translation and Spanish citations rose 16%, and visibility on secondary-language queries improved 327%. For Spanish-language sites specifically, the citation gap between Spanish-language and English-language queries in Google AI Overviews dropped from 431% for untranslated sites to just 22% for translated sites in the Spain sample, and 59% in the Mexico sample. ChatGPT showed close to no language bias either way, with translated Spanish sites picking up only 0.3% to 1.8% more citations in the non-native language, which the study’s authors read as evidence that AI systems weight the presence of multilingual content itself as a trust and coverage signal, not just a keyword-matching exercise.

The practical read: if your content strategy already includes tracking how AI search treats your existing content, adding translated versions of pages that already earn citations in your primary market is one of the more measurable ways to expand share of answer into new-language queries, and the Weglot data suggests the lift compounds rather than cannibalizes.

flowchart TD
    A[Page ranks well in source language] --> B{Search volume in target market?}
    B -- No meaningful volume --> Z[Skip translation]
    B -- Yes --> C{Is topic YMYL?}
    C -- Yes --> D[Full human localization + qualified reviewer]
    C -- No --> E{Language pair distance from source?}
    E -- Close, e.g. Spanish, German, French --> F[MT + terminology pass + native spot-check]
    E -- Distant, e.g. Japanese, Arabic, Korean --> G[MTPE: human edits MT draft]
    F --> H[Re-research target-market keywords, not just translated ones]
    G --> H
    D --> H
    H --> I[Publish with hreflang self + reciprocal tags]
    I --> J[Track citations and rankings separately per market]

The decision tree: should you translate this page

Not every page is worth translating, and the volume question comes before the quality question. A page with strong source-language rankings but no search volume in the target market wastes translation budget regardless of how good the output is. Run the check in this order:

1. Does the target market search for this at all? Pull local-language keyword volume before committing. Translating a page nobody searches for in that market produces a page with no ranking upside and full scaled-content-abuse exposure, since it adds no value by definition.
2. Is the topic YMYL? If yes, budget for full human localization and a qualified reviewer regardless of language pair distance. Skip MT entirely for the first draft if the domain has legal or medical liability.
3. How far is the language pair from your source language? Close pairs (Spanish, German, French, Portuguese, Italian, Dutch) tolerate MT plus a terminology pass. Distant pairs (Japanese, Korean, Arabic, Thai, Russian) need MTPE at minimum.
4. Are you translating keywords or researching them? Re-run keyword research in the target market. A translated keyword list carries over search intent from the wrong country.
5. Is hreflang wired correctly before launch? Publish with self-referencing and reciprocal annotations from day one. Retrofitting hreflang across dozens of translated pages after the fact is where most of the 67% failure rate documented by Ahrefs originates.

What actually gets filtered

Reading Google’s spam policy language closely, the pattern of translated content that gets suppressed or excluded shares a small set of traits, not a translation flag:

• Bulk, unedited output published at scale with no market-specific research behind it. This is the literal scaled content abuse definition: many pages, automated transformation, “little to no value.”
• Content that fails Google’s own language detection, meaning the translation is broken enough that “the visible content of your page,” which is how Google determines language per its multi-regional guide, reads as neither the source nor the target language cleanly.
• Duplicate structure across markets with no localization, where currency, units, legal terms, and cultural references stay in the source market’s form despite a language change, signaling the page was translated but not adapted.
• No disclosure where automation materially affects trust, which the helpful content guidance flags directly as a question worth answering for readers, particularly relevant on YMYL topics.

None of these traits require AI involvement to trigger a quality problem, and none of them are automatically triggered by AI involvement either. A human translator working fast and skipping market research produces the same failure pattern as unedited MT output. The filter is value, and translation method is incidental to it.

Building a multilingual content workflow that survives review

A workflow that survives Google’s scaled content abuse policy and produces content buyers in the target market actually want looks the same regardless of team size:

1. Source content selection. Start from pages that already rank or convert in the source language. Translating unproven content multiplies an unknown into more markets rather than validating anything.
2. Market-specific keyword research, not translated keyword lists. Search intent and query phrasing shift by country even within one language, as the Semrush “lentes” vs “gafas” example shows.
3. MT or MTPE per the language-pair and YMYL rules above, with a terminology glossary maintained per market so brand terms, product names, and category language stay consistent across every page in that language.
4. Word-count and structure matching to the source page’s proven format. If the English version that ranks is 2,200 words with a specific H2 structure, the translated version should hold that shape rather than drifting shorter through literal MT compression or longer through over-localization. This is the step where translation projects lose control fastest, because MT output length varies by language (German commonly runs 20-30% longer than English for the same meaning, for instance), and teams either pad or trim inconsistently across dozens of pages. Tools built for AI article generation with word-count control help here specifically, since you can regenerate a market-adapted draft against the same target length constraint the source page used, rather than manually trimming or padding machine output page by page.
5. Native speaker review before publish, scoped to the market-tolerance tier from the table above, not a blanket full-rewrite policy that burns budget on markets that don’t need it.
6. Hreflang wiring at launch, self-referencing and reciprocal, verified with a crawler before the pages go live rather than audited after the fact.
7. Per-market tracking, separate from your source-market dashboard, since blended reporting hides which languages are actually working.

Measuring performance per market

Aggregate multilingual traffic numbers hide the market that’s failing inside the market that’s winning. Separate reporting by hreflang-tagged URL segment or subdirectory at minimum, and track three signals per market independently: organic rankings for the market-specific keyword set (not the translated keyword list), AI citation rate in that language if the market has meaningful AI-search usage, and conversion rate against a market-adjusted baseline rather than your source market’s benchmark. A German page converting at half your English rate is not necessarily underperforming if German buyers convert at half the rate site-wide for reasons unrelated to content quality, currency friction being the common culprit.

Review each market’s search console data separately every quarter at minimum, and treat markets where indexation lags translation launches (pages published but not indexed after several weeks) as a signal worth investigating before scaling further translation in that language, since it usually points to a duplicate-content or hreflang misconfiguration rather than a content quality issue.

Frequently asked questions

Does Google penalize machine-translated content just for being machine-translated?

No. Google’s spam policy names translation as one method that can produce scaled content abuse, but only when the resulting pages provide “little to no value to users.” A well-translated, market-researched page faces no translation-specific penalty. The trigger is bulk publication of low-value pages, and translation is incidental to that pattern, not the cause of it.

Is raw, unedited Google Translate output automatically thin content?

Not automatically, but it’s higher risk. Raw MT preserves overall meaning in roughly 82.5% of cases on average, with a 55% to 94% range depending on language pair. For close language pairs and low-stakes content, raw MT with a terminology check can pass. For distant pairs or YMYL topics, unedited output is far more likely to read as thin or inaccurate to both readers and Google.

What is MTPE and when do I need it?

MTPE (machine translation post-editing) means a human editor revises a machine-generated draft rather than translating from scratch. It’s the workable middle ground for language pairs that sit structurally far from your source language, such as Japanese, Korean, or Arabic, where raw MT error rates run higher and a human pass catches errors MT alone misses.

Do AI-translated pages need hreflang tags?

Yes, exactly the same as human-translated pages. Hreflang tells Google which language and region each URL targets; it has nothing to do with how the content was produced. Skip it and you risk the wrong language version showing in search results even when the translation itself is accurate.

How do AI search engines like ChatGPT and Google AI Overviews treat translated content?

A Weglot study of 1.3 million citations found translated sites received 24% more citations per prompt than untranslated sites, with the citation gap between native and secondary-language queries in Google AI Overviews shrinking from 431% to as low as 22% after translation. ChatGPT showed close to no language bias either way, suggesting multilingual presence itself reads as a trust signal.

Which languages does machine translation handle best?

Romance and Germanic languages close to English structurally (Spanish, French, German, Portuguese, Italian, Dutch) consistently score highest in comparative MT accuracy studies. Languages with different sentence structure, script, or honorific systems (Japanese, Korean, Arabic, Thai) show larger error rates and need more human editing before publication.

Should medical or legal content ever be machine translated?

Use MT as a first draft only, never as the published version. A domain-specific study cited by Weglot found medical phrase translation accuracy at just 57.7%, far below the roughly 82.5% general-content average. Legal and medical topics carry liability and trust stakes that require a qualified human reviewer regardless of the source language pair’s general accuracy.

What’s the actual difference between translation and localization?

Translation converts text while preserving meaning. Localization adapts the message for cultural preferences, currency, units, legal requirements, and regional search behavior. A literal translation of a US price point or shipping threshold often makes no sense in a market with different currency and buying psychology, even when every word is grammatically correct.

Can duplicate content issues apply across different languages?

Not in the classic sense, since different languages are different content by definition. The real risk is structural duplication: translated pages that keep the source market’s currency, units, and cultural references unchanged, which reads to both users and Google as a page that was translated but never localized for the market it targets.

Subdirectories, subdomains, or country domains for translated content?

Google’s own guidance lists subdirectories (example.com/de/) as low-maintenance with a single host to manage, subdomains (de.example.com) as easy to set up though less visibly geotargeted, and country-code domains (example.de) as the clearest signal but the most expensive to acquire and maintain at scale. URL parameters are explicitly not recommended.

Does it matter whether I use Google Translate, DeepL, or another engine?

A Weglot and Nimdzi comparison across Google Translate, DeepL, Amazon Translate, and Microsoft Translator found no single tool outperformed the others universally; results vary by language pair and content category. Test your specific language pairs and domain rather than assuming one engine wins across the board.

How do I know if my translated pages are actually indexed?

Check Search Console per hreflang-tagged URL segment, not blended with your source-market property view. Pages published but not indexed after several weeks after launch usually signal a hreflang misconfiguration or a duplicate-content read, not a content quality problem on its own.

What word count should a translated page target?

Match the source page’s proven structure and length rather than letting raw MT output length drift. Some languages run meaningfully longer than English for equivalent meaning (German commonly 20-30% longer), so translated pages need a length target set deliberately, not inherited passively from whatever the MT engine outputs.

Do I need to re-research keywords for each target market, or can I translate my keyword list?

Re-research per market. Search terms for the same product vary by country even within a single language, the “lentes” versus “gafas” example for glasses in Mexican versus Spanish Spanish being the commonly cited case. A translated keyword list often targets the wrong phrasing for that market’s actual search behavior.

How often should translated content be refreshed?

At the same cadence as your source-language content, tied to performance data rather than a fixed calendar. A translated page that isn’t ranking or converting after a reasonable indexing window needs the same content refresh evaluation as any underperforming source-language page, not a lower bar because it’s translated.

What is x-default hreflang for?

x-default is a reserved hreflang value that catches visitors whose language or region doesn’t match any of your listed page variants. Google recommends it for language selector pages so unmatched visitors land somewhere sensible rather than on a guessed default that might be wrong for them.

Can machine translation at scale trigger a scaled content abuse penalty?

Yes, specifically when it produces many pages with little value, per Google’s own spam policy wording, which names translation directly as one obfuscation technique that qualifies. The trigger is volume plus low value together, not translation alone; a handful of well-researched translated pages does not resemble the pattern the policy targets.

How much does professional human localization cost compared to machine translation?

Costs vary widely by language pair, domain, and vendor, and no single verified industry-wide rate holds across markets, so get current quotes for your specific language pairs rather than relying on a general figure. What’s consistent across sources is the ordering: raw MT is cheapest and fastest, MTPE sits in the middle, and full human localization costs the most and takes the longest, with SEO risk running in the opposite direction.

Should customer reviews and user-generated content be translated?

Generally leave UGC in its original language with an optional MT toggle for readers, rather than replacing it with translated text as the primary display. Reviews carry authenticity signals tied to the original language and author; replacing them wholesale risks looking edited or fabricated, which cuts against the trust signals E-E-A-T guidance is built around.

What metrics actually show a translated market is working?

Track three per market independently: organic rankings against market-specific (not translated) keywords, AI citation rate in that language where relevant, and conversion rate against a market-adjusted baseline instead of your source market’s numbers. Blended, aggregate multilingual reporting hides which languages are actually converting and which are just adding page count.

Key takeaways

• Google’s spam policy names translation as one path to scaled content abuse, triggered by volume and low value together, not by AI involvement or translation itself.
• Raw MT preserves meaning in roughly 82.5% of cases on average, ranging 55% to 94% by language pair, so the safe-to-publish bar shifts by market rather than staying fixed.
• Close language pairs (Spanish, German, French, Portuguese, Italian, Dutch) tolerate MT plus a terminology pass; distant pairs (Japanese, Korean, Arabic, Thai) need MTPE; YMYL content needs full human review regardless of language.
• Hreflang fails in 67% of implementations studied, independent of translation quality, and is the most common reason a well-translated page still shows the wrong language version in search.
• Translated content measurably improves AI citation rates in Google AI Overviews and ChatGPT, with the largest gains in the secondary language’s own queries.
• Translate proven content, re-research keywords per market rather than translating the list, and report each market separately instead of blending multilingual traffic into one number.

Building the workflow above by hand across a dozen markets is where most multilingual programs stall on execution, not strategy. Vrid.ai handles the AI article generation piece with word-count control so market-adapted drafts hold the same structure and length as the source page that proved out, alongside keyword research and multi-channel publishing, if you want to test the process on a market or two before scaling it further.