Google Already Translates Your Site Into 20 Languages. Chinese and Japanese Aren't Among Them.
In short
- Google already serves English pages machine-translated into about twenty languages, shown in Search Console’s Translated results report. Chinese and Japanese aren’t among them.
- Google assumes a native page exists for any Chinese or Japanese query, so for a niche topic a translation is the only way those readers find the page.
- So every post here is in Simplified Chinese, Traditional Chinese and Japanese, and nothing else. On Mistboard, Chinese pages became six of the top ten organic pages.
- No person reviewed the translations. The model sees only prose segments, the checks compare meaning rather than tokens, and the whole run cost about $25.
Every post on this site is now in Simplified Chinese, Traditional Chinese, and Japanese, and in no other language. The list was going to be a hundred. It’s three because of one Search Console report.
Google translates your English pages for 20 languages
Open Search Console, Performance, and filter Search Appearance to Translated results. That report is Google serving your English pages machine-translated, on its own translate.goog proxy, to searchers in about twenty languages: French, German, Spanish, Portuguese, Arabic, Turkish, Persian, Indonesian, Vietnamese, Thai, Korean, and the major Indian languages from Hindi and Bengali down to Kannada. When a Hindi query is best answered by your English page, Google translates it and shows the result. You get the impression and the click either way.
| Languages | What your own translation does | |
|---|---|---|
| Google proxies | fr, de, es, pt, ar, tr, fa, id, vi, th, ko, hi, bn, ta, te, mr, gu, kn, ml, ur | Moves an existing click onto your domain |
| Google doesn’t | zh, ja | Opens a door that was closed |
A page of your own in those languages gains little
The proxy fires only where no good native page exists, which is exactly where your English page would have won anyway. A Portuguese page of your own there moves the click from Google’s copy onto your domain: your analytics, your embeds, your brand. Worth something, but the same reader. Where native Portuguese pages do exist, the proxy never fires and your own page has to beat them on their home turf, which a 38-page site rarely does. Redundant where you’d win, outmatched where you wouldn’t.
The report also tells you, for free, whether the door you’re about to build is needed. For this site, before any translation existed, the Translated results row was zero. Nobody in the proxied languages was finding these pages at all, so a Vietnamese version would have had nothing to move.
Chinese and Japanese are the exception
They aren’t on the list because the pattern in the list is big searcher populations with thin native webs (Hindi is about 0.1% of web content), and Japanese alone is about 5% of the web. Google assumes a native page exists for any Japanese query. That’s true of the language and false of a topic. There is no Japanese page about a tablebase for dōbutsu shōgi. A Taiwanese reader searching 象棋殘局 will never be shown my English endgame table, not because Google can’t translate it, but because Google decided it didn’t need to. For those two languages, a translation is the only door.
Mistboard already showed it
mistboard.com went into both Chinese scripts earlier this year with no native review. By September, six of its top ten organic pages were the Chinese versions and Taiwan was the top country by clicks. Those readers could not have found the English.
Scaling raw translation gets penalized
The one large site that scaled machine translation without review got hit. Reddit’s ?tl= pages, millions of URLs across 31 hreflang variants, grew for about fifteen months and then lost their translated visibility in the May and June 2026 core and spam updates, with AI Overview and ChatGPT citations falling alongside. Thirty unreviewed languages on a domain with 38 indexed pages is that pattern at a size where nobody would notice the recovery either.
How the three were produced
No reviewer read any of it. What made that defensible: the model never sees the markup, and the checks compare meaning rather than tokens.
Each post is split into prose segments by byte offset (paragraphs, headings, list items, table cells, image alts, captions). Code, HTML, URLs, and table pipes never leave the file. The model gets a JSON array of {id, text} and returns translations by id; the file is reassembled by splicing at the original offsets, so reassembly with no translation is the identity by construction.
My first fidelity checks then rejected correct translations 30 times in a 56-request batch before they caught a wrong one:
| English | Translation | What I had checked | Fix |
|---|---|---|---|
| 6.9 million | 690万 | the digits 6.9 survive |
compare values: 6.9 × 10⁶ = 690 × 10⁴ |
| $40-70k | 4 万至 7 万 | 40 and 70 survive |
a range shares its suffix |
| 202 million | 2億200万 | one scale character | sum compound CJK numerals |
| 90-95% | 90% 到 95% | 90 with no percent sign |
percent and plain compare equal |
| Tailwind CSS, Elo, Las Vegas | unchanged | output ≠ source | only lowercase common words count as untranslated |
The two real errors in that batch were caught by checks that don’t compare tokens: a Korean word in the middle of a Japanese sentence, and a bold link returned with its opening bracket gone. Script and structure checks hold; token matching mostly rejects correct output. After that, a second pass with the same model as judge, given the source, the translation, and the term table, took the open error count across 84 pages from 65 to zero in three rounds, about two dollars a round. The whole thing cost about $25.
The rule
Translate every post into the languages the search engine can’t substitute for, and none of the ones it can. For a site about xiangqi, banqi, jungle chess, and shogi, that rule names the readers who care most.
Whether anyone was behind the door is a Search Console question: impressions per page per country, by language prefix, eight weeks for impressions and sixteen for clicks, on a site doing 55 clicks a quarter. Not the coverage report, since hreflang alternates show as not-indexed by design (Illyes, August 2026). If Taiwan and Japan don’t appear at all, the theory that readers were finding this in the wrong language was wrong. The segmenter, checks, and judge prompt are in the site repo under _i18n/.