Before making your articles multilingual with AI translation: checking steps from our four-language rollout

The short answer

If you publish AI-translated articles in several languages, do not stop at checking the translations once. It is practical to decide first on a page structure that keeps each translation tied to its original, and on a step in which a separate pair of eyes compares the translation with the original.

In September 2026, Rootpublish made its own journal available in four languages: Japanese, English, Simplified Chinese and Korean. The articles approved and published in Japanese served as the originals, and the translations were checked in a separate AI call from the one that translated them. The check found translations that widened the meaning and literal carry-overs from Japanese. It also produced findings that should not have been adopted.

This article sets out, within what the records state, the structure, the checking method, the mistranslations found and the progress toward being indexed by Google. It ends with checking steps for doing the same on your own site.

Note that the translations were checked by AI; no person or native speaker has reviewed them yet. Nothing here guarantees translation quality, search rankings or citations by AI.

Publish translations as separate pages tied to the original

In Rootpublish's journal, the article approved and published in Japanese is the original. The English, Simplified Chinese and Korean translations are each published as separate pages.

Each page lists the four languages and x-default (English) as alternates of one another. Simplified Chinese is marked as zh-Hans, as Google's documentation describes. From any language version, it is clear that the other versions are alternates of the same content.

The other decision was what happens when the original changes. When the original changes, the publishing process stops until the translations have been checked again. If only the Japanese is corrected and an old translation stays up, pages in different languages end up saying different things. Deciding first which language is the original also makes this stop condition easier to define.

Translations were checked in a separate call from the translation

The translations were checked with a Claude Opus call separate from the one that translated them. It compared the original with the translation and reported omissions, additions, changes in meaning, terminology and unnatural phrasing.

The number of findings was as follows:

• Chinese and Korean translations of two articles: 11
• English, Chinese and Korean translations of three articles (nine pairs): 25
• Translations of one other article into three languages: 10

Claude decided whether to adopt each finding by comparing it with the original, and only adopted findings led to corrections. In other words, translating, checking and deciding were all done by AI, and no person or native speaker has reviewed the result yet. Note that these counts are numbers of findings, not numbers of mistranslations, and they do not measure translation quality.

What was found: overstatements and literal translations

Broadly, the mistranslations found either changed or widened the meaning, or carried Japanese over literally.

An example of a change in meaning was a translation of 「変更を復元」 ("restore the change", meaning revert it) that read as re-applying the change. Examples of widened scope included translating 「多く」 ("many") as 「大部分」 or "most", and a translation of 「97%以上だったモデルが」 ("models that scored 97% or more") that read as "all such models". In each case, a limitation or condition in the original was lost in translation. In a company's verification results or product descriptions, overstating quantity or scope like this becomes a claim that differs from the facts.

Literal translations from Japanese included the Chinese 「课题」 and 「实绩」 and the Korean 「독자 블록」. In another case, the review-only summary and the WordPress excerpt were translated with the same word, so they could no longer be told apart. Product-specific terms mix easily unless their translations are decided in advance.

An existing English translation also overstated 「多く」 as "most", and it was corrected at the same time. Not only newly translated languages but also translations already published need to be reviewed.

Some findings were not adopted

Applying every AI finding was not the right answer.

Because this check compared text with the tags removed, it repeatedly reported, wrongly, that the numbers of numbered lists and the URLs of links were missing. They are not missing on the displayed pages, so these findings were not adopted. This shows that the format used for comparison can itself produce false findings.

For Korean, one finding suggested changing "publish" from 「공개」 to 「발행」. However, the wording has to stay consistent with articles already published, so rather than changing it right away, we decided to wait for a native speaker's judgment. Consistent terminology is hard to decide article by article, and this was a decision to hold off rather than conclude on the strength of an AI check alone.

Progress toward being indexed by Google

On September 25, 2026, we submitted the sitemaps to Search Console and also sent the URLs to IndexNow.

Right after submission, six pages were indexed, including the home page, the WordPress edition's page and article 5. About 11 hours later on the same day, the number had grown to 11, including the Japanese and English versions of article 8, two Chinese articles and the Korean journal index. There are no numbers yet for impressions or clicks in search.

This is one day on one site; it does not show how quickly pages get indexed in general or what they achieve in search. Google explains that the fundamentals of SEO still apply to AI Overviews and AI Mode, and states that meeting the requirements for appearing in these features does not guarantee crawling, indexing or display. Rootpublish does not guarantee search rankings or citations by AI. Its search improvement feature is designed so that after one improvement is approved and published, the same page is observed for seven days and judged on measured results.

Checking steps for making your own articles multilingual

From here on is a suggested procedure based on the records above and the approach of the Rootpublish editorial guide. It is not a measured result showing improved search rankings.

1. Decide which language is the original. Publish translations as separate pages and list the languages as alternates of one another.
2. Decide what happens when the original changes. Stop the publishing process until the translations have been checked again, or at least hold back updates from going live until then.
3. Compare the original and the translation in a call separate from the translation. Look for omissions, additions, changes in meaning, terminology and unnatural phrasing. Pay particular attention to words for quantity, scope and conditions, and to product-specific terms. Review existing translations from the same angles.
4. Decide whether to adopt each finding by comparing it with the original, and record the reason. Some findings are better not decided on the spot, such as false findings caused by the comparison format or terminology that must stay consistent with published articles. Send the latter to a native speaker.
5. Before publication, a person with publishing rights checks the text and sources and then decides whether to publish.
6. After publication, record when the sitemaps were submitted, which pages were indexed and how impressions and clicks change, and judge on measured results.

Google asks for a focus on accuracy, quality and relevance when content is prepared with generative AI, and says that mass generation that adds no value for users may violate its spam policies. Rather than making more translations a goal in itself, it is better to start with the articles that readers of each language need.

In Rootpublish's own records, review by a person or native speaker is not finished yet. Because an AI check can only find so much, we recommend building into your schedule the point at which a person reviews the translations.