Is machine translation good enough for China?

The vendor guides we read say raw machine translation isn’t enough for China. Each vendor also sells a related service, and none cites data. So it’s fair to ask what the evidence says without the sales pitch.

The short version: the evidence is thinner than the vendors suggest, and the case for review rests on quality, not on a penalty anyone has measured.

This guide is part of our series on translating and localizing a website for China.

What does the evidence say about machine translation into Chinese?

Raw output still contains serious errors, and the best automated correction cut them. But no study we read measures how many a human reviewer removes, or tests your exact case.

That case would be English into Simplified Chinese marketing pages, scored by human readers, with Chinese reported on its own.

Study	What it tested	Main finding	Limit
Welocalize, EAMT 2026 1	English into 10 languages, including 7,763 zh-CN segments out of 71,262	AI post-editing beat raw engines and direct LLM translation overall	Figures pooled across languages. The authors work for the firm that built the system
Zou et al., June 2025 2	English to Chinese, 3,498 sentences	DeepSeek-V3 0.7790, GPT-4o 0.7655 on an automatic score	Academic papers. No human scoring
Chen and Lin, July 2025 3	Chinese to English tourism texts, 7 raters	ChatGPT 3.87 to 4.07, DeepL 2.20, Google 1.90, out of 5	Reverse direction, GPT-3.5, no human baseline
Alconost, May 2026 4	Client projects, zh-CN	DeepSeek 72.2 (n=12), Gemini 71.9 (n=17) on a composite score	Small samples. The firm sells post-editing
Venkatesan, June 2025 5	Survey of postgraduate translation students	Vocabulary and style need human work most	Opinions, not error counts. We read the abstract

The Welocalize paper is the closest fit, because 26% of its segments are marketing text and professional translators reviewed a sample blind.¹ Here’s what they found per marketing segment.

1. Google Translate output + AI post-editing0.06
2. Production AI post-editing0.09
3. Google Translate alone0.09
4. Direct LLM translation0.11
5. DeepL alone0.12

Welocalize 2026, ten languages

That’s one serious error roughly every 8 to 11 segments for raw output.¹ Edit distance against human post-edits tells the same story: 0.20 to 0.22 for raw systems, 0.11 to 0.12 with AI post-editing.¹

The gain isn’t automatic, though. The production setup, which also reused older translation memory, stayed at 0.09 on marketing text, level with raw Google Translate.¹

The paper gives no error rate for Chinese alone. Its charts do show zh-CN’s share: 9% to 10% of each system’s errors, and 12% to 14% of its major and critical errors.¹

An even split across ten languages would be about 10%, so serious errors in Chinese don’t look lighter than average. These are shares across all content types, not rates.

Note what “AI post-editing” means here: an LLM correcting a draft, using style guides and approved examples.¹ It isn’t a human review, so these numbers don’t show what a human reviewer would catch. It also isn’t what you get by pasting text into a free tool.

Which engine should you use for Chinese?

No engine wins clearly, so don’t spend long choosing. The studies disagree, largely because they test different things.

• Welocalize: translators ranked Google Translate about level with direct LLM translation, and DeepL lowest.¹ Neither engine had a glossary in that test.¹
• Chen and Lin: ChatGPT finished far ahead, but on Chinese to English, with GPT-3.5.³
• Alconost: DeepSeek and Gemini were effectively tied for zh-CN on its composite score. Across all languages, no engine topped 67.8 out of 100 with its linguists.⁴

Setup may matter as much as the engine. Terminology made up 32.3% of Google Translate’s errors, with no glossary. It made up 19.9% for the LLM, which had style guides and example pairs.¹

They’re different systems, so that’s a hint, not proof.

If you use DeepL, select Simplified Chinese explicitly. Its documentation calls plain ZH an unspecified variant kept for compatibility, and recommends ZH-HANS or ZH-HANT.⁶

Will Baidu penalize machine-translated pages?

We found no Baidu rule that names machine translation, though one old notice comes close. Translation doesn’t appear in the chapter on cheating and penalties in Baidu’s 2014 SEO guide.⁷

That proves little, because the chapter gives three examples of cheating and doesn’t claim to be a full list. Our searches for a current rule on translated or AI content found nothing either, which doesn’t prove none exists. The rest is inference, including ours.

Here’s what’s actually on record. Some of it is old, and nothing confirms it still applies as written.

Source	What it says	Why it’s relevant
Baidu notice, July 2012, reposted by OSChina 825	Low-quality sites have little original content, and much that’s scraped or machine-generated or modified	Closest wording to translation, but the excerpt we read doesn’t name it
Hurricane algorithm, 2017 to 2019 9	Targets scraping, mirror sites and duplicate content	Translating someone else’s content could look similar
Drizzle algorithm, reported 2019 10	B2B sites: scraped content and sentences that don’t read smoothly	Relevant if raw output reads stiffly, but this is one industry article, and it concerns B2B sites
Baidu quality guide, reported 2020 11	Judges the producer’s authority, page experience, depth and user reputation. Text should be clear, consistent and free of typos	None of this concerns how the text was made, but stiff or wrong output could fall short

Vendors go further than this record allows. Linguise says poor translation leads to lower rankings, and Chinafy says Baidu favors native-sounding content.¹²¹³ Neither gives data.

The likely mechanism is ordinary quality. A page that reads badly gives a Chinese reader a reason to leave, and it does little to answer a search. That’s a sensible assumption, not a measured effect.

What about Google and Bing?

Google’s spam policies don’t ban translation. They target scale without value. “Translating” appears once, as one way of turning scraped content into many pages with little value for users.¹⁴

The same policy covers using AI tools to generate many pages without adding value.¹⁴

Google told SEO consultant Glenn Gabe in 2025 that its policies “do not strictly define content that has been translated by AI as spam.”¹⁵

Even so, Gabe reports sites receiving manual actions or spam-update hits after scaling machine translation.¹⁵

On that evidence, the pattern to avoid is a plugin that creates thousands of unreviewed pages at once.

This matters if you target Chinese readers outside the mainland. StatCounter gave Google 89.59% of search in Hong Kong and 80.02% in Taiwan in October 2026.¹⁶¹⁷

We couldn’t read Bing’s guidelines ourselves. Search Engine Journal reported in February 2026 that they now name “Large-scale content generated without oversight, quality control, or editorial review.” Its report doesn’t mention translation.²⁶

Bing matters here: StatCounter gave it 21.69% of search in China in October 2026.²⁷ Check Bing’s own documentation before relying on this.

Free check

Is your website visible on Baidu?

Enter your website. We check Baidu, access from China and how your site is set up, then email you the report.

Your websiteCheck my site

Where does machine translation fall short on a Chinese site?

It translates the sentence in front of it. It doesn’t know your market, your legal exposure or what people type into Baidu. Those gaps matter more than grammar.

Weak spot	Why an engine misses it	What to check
Search terms	A correct word isn’t always the word people search	Main terms against Baidu keyword research
Regional vocabulary	Words differ by region: 鼠标 on the mainland, 滑鼠 in Taiwan, both “computer mouse” 21	Terms fit your target script and region
Your name and products	Without a glossary, terms can vary from page to page	One approved Chinese brand name and fixed product terms
Added meaning	Chen and Lin found ChatGPT could distort or embellish the source 3	Specs, guarantees and numbers, line by line
Absolute claims	China’s Advertising Law restricts terms such as “the best” 18	Superlatives, with a local adviser
Local conventions	NN/g: Chinese promotions usually show the share paid, not the discount 19	Offers, dates, units, calls to action
Punctuation	Mainland uses double quotes first, Taiwan single. Mainland marks sit in the corner of the character square, not centered 20	Quote order in the text. Mark position also depends on your font

The advertising point deserves care. China Briefing reports fines from RMB 200,000 to RMB 1 million for advertisers, with exceptions.¹⁸ How this applies to a foreign company’s website is a question for a local adviser.

NN/g’s B2B testing in China makes the broader point: “Merely translating into the appropriate language is not enough.”¹⁹

Readers notice when a site doesn’t match local conventions, even when the translation is correct.¹⁹ That research dates from 2017, so read it as direction, not measurement.

How much review does each page need?

Match the review to the cost of being wrong. This table is our judgment, not a standard, because no source we read sets these levels for Chinese.

Page type	Treatment	Why
Home, landing	Rewrite by a native writer	These pages have to persuade
Product, service, contact	Full post-editing	These pages carry your credibility
Titles, descriptions, headings	Written by a person from your keyword list	They decide what a searcher sees first
Legal, safety, regulated claims	Specialist translator and local adviser	Errors carry legal risk
Blog posts and guides	Full post-editing	They need to earn search visits
Support articles, documentation	Light post-editing	Clarity matters more than style
Internal use only	Raw output is fine	Nobody outside relies on it

Full post-editing is the human process covered by ISO 18587.²²

Linguise describes the light version as fixing significant grammar and word errors. In the full version, the translator revises everything to fit the style you want.²³

Budget is often the real question. Centus said in June 2025 that basic machine translation is often free or $0.01 to $0.05 per word, against $0.10 to $0.20 for human marketing work.²⁴

Those are one vendor’s figures. See what a China-ready website costs.

How to test machine translation on your own pages

A small test shows what no benchmark can: how your own pages fare. No study we read covers your product, so run your own before committing a whole site.

1. Pick five pages. Include your home page, one product page and one technical page.
2. Set up the engine. Give it a short glossary of brand and product terms.
3. Find a native reader. They need to know your field and not have produced the draft.
4. Have them mark errors. Use the labels from the Welocalize study: minor, major, critical.¹
5. Ask one direct question. “Would you contact this company after reading this page?”
6. Check search terms. Compare the headings and titles with your keyword list.
7. Decide by page type. Use the results to adjust the table above.

If you do have a reviewer, how to translate your website into Chinese covers the full workflow.

The bottom line

Use machine translation to draft, and don’t publish it unread on any page that represents your company.

The evidence says output still needs correcting. It doesn’t prove that Baidu punishes translation, and anyone who says so should show the data.

How Index in China can help

We translate your site into Simplified Chinese with a human reviewer, as part of getting it indexed on Chinese search engines. We don’t promise rankings.

Want to talk through which of your pages need that review? Book a 30-minute call.

Book a 30-minute call

Monthly report BaiduIndexed Bing ChinaIndexed SogouIndexed