Automating translation and updates on a bilingual website: translation keys, glossaries and 7 common hreflang mistakes (2026)

Run a website in Japanese and English long enough and the English side falls behind: Japanese pages edited weekly, English ones six months old, a price stale in one language, a new article with no English version. We hear this from businesses in Japan and Europe alike.

Slow translation is rarely the cause. Almost always, nobody knows which part of which page has changed: the work around translation (tracking changes, consistent terminology, paired pages, hreflang) is still done by hand.

This article sets out how to automate translation and updates on a Japanese–English site in three parts: managing translation pairs, dividing work between machine translation and human review, and implementing hreflang and avoiding the usual mistakes.

Main sources for this article

• Google Search Central on localized versions (updated 21 September 2026), multi-regional sites (10 December 2025) and spam policies (28 August 2026)
• W3C Internationalization on language tags and declarations; WCAG 2.2 criteria 3.1.1 and 3.1.2
• Google Cloud Translation and DeepL API pricing, limits and glossary documentation
• CSA Research, "Can't Read, Won't Buy – B2C" (2020, 8,709 consumers in 29 countries)
• ISO 18587 (post-editing of machine translation output) and OASIS XLIFF 2.1

1. What's changing: readers want their own language, and machine translation got cheap

Readers want their own language, and machine translation has become very cheap. The temptation to "just machine-translate everything" grows, while search engines get stricter about exactly that.

Four in ten readers won't buy from a site in another language

In 2020 the US research firm CSA Research surveyed 8,709 consumers in 29 countries and reported that 76% prefer products with information in their own language, and 40% will not buy from websites in other languages.

65% would rather have information in their language even if the quality is poor: readers want their own language first, a perfect translation second.

A 2011 European Commission survey (Flash Eurobarometer 313) likewise reported that around nine in ten internet users in the EU visit a website in their own language when they have the choice.

Machine translation costs USD 20 per million characters

Google Cloud Translation's standard neural machine translation (NMT) is free for the first 500,000 characters each month (a USD 10 monthly credit) and USD 20 per million characters beyond that. The large-language-model "Translation LLM" costs USD 10 per million characters each for input and output, which puts it on a par with standard NMT when input and output are about the same length (checked 25 October 2026).

DeepL API Free also translates up to 500,000 characters a month, and the paid plans are usage-based.

Search engines are stricter about "just translated"

Google's spam policies (updated 28 August 2026) list mass-producing pages through automated transformations such as translation, without adding value, as an example of "scaled content abuse".

Translating your own articles is not copying, but publishing unread machine output moves you towards that example. What has become cheap is the draft, not a translation fit to publish.

Google has also retired the International Targeting report in Search Console (hreflang itself is still supported). The table summarises these shifts.

All point the same way: you need systems for checking more than for translating.

2. Where bilingual maintenance time actually goes

Auditing operations work is covered in "Automating website operations starts with an audit"; here we take only the translation tasks. Time and errors concentrate in tasks before and after translation that rely on memory.

Five tasks around translation

Only the second row is translation itself. The other four are tasks a machine can almost entirely handle once the rules and data structure are decided.

"The English page is out of date" is a detection problem

The most common failure is not mistranslation but one language being out of date, such as an old price left on the English page. The problem is not speed but having no way of knowing a page went stale.

3. The big picture: five stages of a translation pipeline

Translation and updates are easier to reason about as one pipeline. Here is the flow, in five stages, for a site where Japanese is the source language. The key is to concentrate human involvement in one place.

The table lists the tools used at each stage and the records kept so that you can later explain who checked what, and when.

Skip stage 3 and you have "a site of machine translations". Skip stages 1 and 4 and people end up rereading entire pages every time.

4. Managing pairs (1): make the pair data with a translation_key

Keep the translation_key separate from the URL

The simplest way to express a pair is to write the same key (translation_key) on both pages. This column does exactly that: the Japanese and English front matter share one key, and hreflang is generated from it. If you match pairs by URL, a URL change breaks the pair, so the key never changes once set.

---
title: "Automating translation and updates on a bilingual website"
translation_key: bilingual-site-translation-automation
source_locale: ja                 # language of the source
source_hash: 3f9c2a…              # fingerprint of the source when translated
translation_status: reviewed      # machine / reviewed / outdated
reviewed_at: 2026-10-25
---

Detect stale translations with a source fingerprint

Record a hash (fingerprint) of the source at the time of translation on the translated page, and comparing it with the current source hash tells you whether the translation is stale.

A whole-page hash can't tell you where, so also hash each paragraph and keep those in translation memory. Only changed paragraphs then go to translation and review.

Decide publishing from the combination of states

Three states are enough: machine (unreviewed MT), reviewed (checked by a person) and outdated (source has changed). Tie them to publishing rules.

For typos, a few days' delay on the English side does little harm; for prices, dates, contact details or legal notices, hold both versions until the English one is ready.

5. Managing pairs (2): translation memory and glossaries

Translation memory: reuse approved translations sentence by sentence

Translation memory (TM) is a store of source–target pairs approved by a person, kept at sentence or paragraph level. When the same sentence appears again, the approved translation is reused; similar sentences are offered as suggestions.

Footers and notes under price tables recur all over a site. TM saves translating and checking them every time and stops approved wording silently changing.

Exchange formats include XLIFF 2.1, approved as an OASIS Standard in 2018, and TMX. A small site can start with a CSV kept in Git. What matters is not locking it inside one tool.

Glossaries: fix terms and pass them to the engine

A glossary is a dictionary that fixes how individual terms are translated. Both DeepL and Google Cloud Translation accept glossaries through their APIs and apply those terms first.

We checked these limits in each provider's documentation on 25 October 2026. Neither offers versioning, so it is safest to keep the glossary in your repository and push it to the API on every change.

What goes in a glossary, and what doesn't

An oversized glossary also fixes words that should vary with context. Limit it to terms where inconsistency causes harm.

A do-not-translate term stays as-is when source and target are identical. Style that a glossary can't capture (whether to write as "we", date and currency formats) goes into a one-page style guide kept beside the reviewer. With LLM-based translation you can also pass it as instructions.

6. Choosing a machine translation engine, and what it costs

For Japanese–English, the main options are DeepL and Google Cloud Translation. Judge them not only on how the output reads but on glossaries, data handling and how characters are counted. Quality varies by document type, so the reliable test is 20–30 paragraphs from your own pages.

Main options and prices

eTranslation, the European Commission's service, is free for EU-based public bodies and SMEs (including the self-employed). We could not confirm Japanese support, so it suits translation between European languages.

Adaptive translation and custom models cost 2.5 to 4 times standard NMT. For smaller sites, start with standard NMT or DeepL plus a glossary, and try higher tiers only where you remain unhappy.

Billing counts source characters

Billing is generally based on the number of source characters; DeepL says it counts Unicode code points in the source text.

In our own articles, one Japanese character corresponds to roughly two to three English characters, so translating from Japanese into English bills fewer characters than the reverse, which matters for free allowances.

Check how your data is handled

Drafts may contain unannounced prices or customer names. For its paid Pro service, DeepL says texts are stored only as long as processing requires and are not used for training; the free version may differ, so check the terms. Text that may contain personal data should only be machine-translated after checking the processor agreement under the GDPR (EU General Data Protection Regulation).

7. Splitting the work between machine translation and human review

The work a person does on machine output is called post-editing. The international standard ISO 18587 (2017) sets requirements for "full post-editing", in which a person fully revises the output; "light" post-editing appears only in an annex. Public pages are read as your company's own words, so review them to the full standard.

Vary the split by document type

Divide the work according to how much damage a mistranslation would do.

Google notes that translating only boilerplate such as menus makes for poor search results; untranslated posts belong to source-language pages.

Decide in advance what reviewers check

Fix the review items and show them on the review screen.

Run the checks software can handle before review and flag mismatches, so the reviewer can concentrate on meaning and style.

The reviewer must be able to judge whether the target language reads naturally; an external reviewer can be sent only the changed paragraphs. For how to record approvals, see also "Automated publishing depends on a quality gate".

8. Implementing hreflang: three methods and how to choose

hreflang tells search engines that versions of the same content exist in other languages. Once pairs are managed as data, it becomes a matter of generating it from that data in your templates or build. Written by hand, something always gets missed.

The three methods are equivalent: pick one

Google treats HTML link elements, HTTP headers and XML sitemaps as equivalent. It states plainly that using all three adds no search benefit and only makes maintenance harder.

For a Japanese–English site, output the following in each page's head. List every version including the page itself, and write fully qualified URLs starting with https.

<link rel="alternate" hreflang="ja" href="https://example.com/column/translation-automation" />
<link rel="alternate" hreflang="en" href="https://example.com/en/column/translation-automation" />
<link rel="alternate" hreflang="x-default" href="https://example.com/en/column/translation-automation" />

x-default is where everyone else goes

x-default is the page shown to readers whose language matches none of the listed versions. On a Japanese–English site, pointing it to the English version is the usual choice.

Keep language codes short

Language codes follow ISO 639-1 (ja, en); regions, if added, follow ISO 3166-1 Alpha 2 (JP, GB). W3C recommends keeping language tags as short as possible, so ja is enough for Japanese. Use en-GB and en-US only if the pages genuinely differ in spelling or prices.

The lang attribute does a different job

hreflang is often confused with the lang attribute on the html element. Google says it uses neither hreflang nor lang to detect a page's language; it works from the visible content.

You still need lang. web.dev explains that declaring it helps screen readers and translation services identify the language, and W3C says to always put it on the html element. In WCAG 2.2, language of page (3.1.1, Level A) and language of parts (3.1.2, Level AA) are success criteria.

Don't redirect automatically

Google advises against automatically redirecting users based on a guessed language and recommends links to the other versions so readers can choose. Googlebot usually crawls without sending an Accept-Language header, so some versions may become invisible to it.

9. Seven common hreflang mistakes

hreflang mistakes rarely show; you notice when the wrong language version appears in search. With no Search Console report any more, you need your own checks. We have grouped the mistakes Google lists and those we see most often into seven.

Mistakes 1 to 4 are requirements and errors spelled out in Google's documentation. On 5, Google's canonicalisation documentation asks hreflang sites to specify a canonical page in the same language. Point the English page's canonical at the Japanese one and the English page may be treated as a duplicate.

Mistake 6 isn't a stated requirement, but it is the one we see most: an unpublished English page still listed, or an old URL lingering. Emitting hreflang only for live pairs with a reviewed translation (section 4) prevents it by design.

Checks that catch the mistakes

All of these can be checked automatically. At the end of the build, verify the following for every page due to be published, and stop publishing if anything fails.

The language match check catches English templates copied from Japanese ones that still say lang="ja". You can also check that English titles, alt text and body copy contain no more than a small share of Japanese characters, and that English pages don't link to Japanese ones.

10. Cost and time ranges

These figures are DevNJ's working assumptions, not measured statistics. The model site has 40 pages in each language, four articles a month (about 6,000 Japanese characters each) and around ten edits to existing pages a month, totalling just over 30,000 source characters a month.

Monthly hours

Review before automation is often a skim, hence three hours; afterwards the full text is reviewed. Part of the time saved goes into checking, so quality improves.

Machine translation cost

At just over 30,000 characters a month, Google's standard NMT and DeepL API Free both stay within the 500,000-character allowance; even retranslating all 40 pages (about 200,000 characters) fits. Beyond that, with hundreds of pages or three languages, the cost is still small next to people's time.

Implementation stages and costs

Stage 1 alone fixes "we can't tell the English page is out of date", and external translation requests shrink to the changed paragraphs. A translation management system (TMS) subscription would be on top of this.

11. Common failures

Failures we see in practice, mostly from how operations are set up rather than technology.

The costliest is the last row. Readers can't tell it is wrong, so they go straight on to a contract.

12. Japan and Europe compared

The context differs somewhat between sites aimed at Japan and at Europe; these are differences in conditions, not a ranking. In Japan, the Japan Tourism Agency's 2014 guidelines on multilingual signage set rules for rendering proper nouns, such as writing 清水寺 as "Kiyomizu-dera Temple", with example term pairs, and they work well as a model for building a glossary.

For a site aimed at Europe that starts in Japanese and English, this design pays off when you add a third language such as Dutch or German. With a translation_key and a glossary in place, adding a language just means adding one more member to each pair.

13. Worked examples (three fictional companies)

Three fictional companies (not real businesses) show how the answer changes.

A had twice published outdated English specs, so it manages spec tables with TM and number matching. B only marks the English page outdated and sends an alert when a price changes; the owner translates. C passes a glossary and style guide to Translation LLM for drafts of its seasonal notices and reuses human-translated booking terms.

14. A six-week rollout

Each week has a clear end state, so progress survives a pause.

1. Week 1: inventory the pairs — map every Japanese and English page; find pages in one language only or paired only by URL
2. Week 2: translation_key and fingerprints — add keys and write source_hash and translation_status, so stale translations are listed automatically
3. Week 3: hreflang generation and checks — generate hreflang from pair data, add the section 9 checks to the build, remove auto-redirects
4. Week 4: glossary and style guide — 50–100 common terms and a one-page guide in the repository
5. Week 5: machine translation and review screen — translate only changed paragraphs and show source and target side by side
6. Week 6: number matching and review — add checks for leftover Japanese and internal links, and log a month of hours and misses

15. What not to do

Each looks easier now and costs trust or search visibility later.

16. Checklist

Use it at rollout and in a monthly review.

Managing pairs

• Every page has a translation_key, and Japanese–English pairs can be listed
• Translations record a fingerprint of the source at the time of translation (source_hash)
• Outdated translations are listed automatically
• Pages in the machine state are never published
• The glossary and style guide live in the repository with change history

Translation and review

• Prices, contracts and legal notices are translated or checked word by word by a person
• Reviewers can judge whether the target language reads naturally
• Numbers, terms and paragraph counts are compared automatically before review

hreflang and publishing

• hreflang is output using only one of the three methods
• Every page lists all versions, including itself
• Language codes are ja and en; regions, if any, use codes such as GB or JP
• URLs are fully qualified, start with https and all return 200
• Canonicals point to the page itself or a same-language page
• The html lang attribute matches the page's language
• There is no automatic redirect by browser language

17. Summary

• The usual bilingual failure is not mistranslation but one side going stale. Start by turning pairs and staleness into data with a translation_key and source fingerprints
• Keep per-paragraph hashes and translation memory so only changed paragraphs are translated and approved wording never changes silently
• Start the glossary with 50–100 terms in the repository; put style rules in a one-page guide
• Machine translation is cheap (standard NMT USD 20 per million characters, 500,000 a month free). The main cost is human review, and prices, contracts and legal notices stay with people
• Generate hreflang with one method and check reciprocity, self-reference, full URLs and same-language canonicals
• Avoid auto-redirects and side-by-side bilingual pages; use switcher links
• On the model site, 22 hours a month drops to around 11, with part of the saving spent on review

How to decide the overall scope of automation is covered in "Automating website operations starts with an audit", and automating updates, backups and other maintenance in "Automating website maintenance".

Sources

Search and multilingual guidance (Google)

• Tell Google about localized versions of your page — Google Search Central: the three hreflang methods, requirements and common mistakes (updated 21 September 2026)
• Managing multi-regional and multilingual sites — Google Search Central: avoiding automatic redirects, one language per page (updated 10 December 2025)
• Spam policies for Google web search — Google Search Central: mass production through automated transformations such as translation (updated 28 August 2026)
• How to specify a canonical URL — Google Search Central: same-language canonicals on hreflang sites (updated 10 July 2026)
• The International Targeting report is deprecated — Search Console Help: retirement of the International Targeting report

Language declaration and accessibility (W3C, web.dev)

• Choosing a Language Tag — W3C Internationalization: keep language tags as short as possible
• Declaring language in HTML — W3C Internationalization: why to put lang on the html element
• Understanding Success Criterion 3.1.1: Language of Page — W3C WAI: language of page (Level A)
• Understanding Success Criterion 3.1.2: Language of Parts — W3C WAI: language of parts (Level AA)
• Document structure — web.dev Learn HTML: lang helps screen readers, search engines and translation services
• Tags for Identifying Languages (RFC 5646) — IETF: the language tag standard (BCP 47)

Machine translation pricing and features

• Cloud Translation pricing — Google Cloud: rates per model and the 500,000-character monthly allowance
• Create and use glossaries — Google Cloud Translation: glossary formats and limits
• DeepL API Free — DeepL Help Center: 500,000 characters a month on the free plan
• Usage limits — DeepL API documentation: how usage is counted (source code points)
• Glossaries — DeepL API documentation: glossary formats and limits
• AI translation and language tools — European Commission: eTranslation eligibility and pricing

Translation standards

• ISO 18587:2017 Post-editing of machine translation output — ISO: requirements for full post-editing and post-editor competences
• XLIFF Version 2.1 — OASIS: exchange format for translation data (OASIS Standard, 13 February 2018)

Surveys and Japanese guidance

• Consumers Prefer their Own Language — CSA Research: 76% and 40% (July 2020)
• Flash Eurobarometer 313: User Language Preferences Online — GESIS Data Archive: record of the European Commission's 2011 survey
• Guidelines for improving multilingual support (Japanese) — Japan Tourism Agency: rules for rendering proper nouns (March 2014)

All sources were checked on 25 October 2026. Prices and specifications change, so confirm the latest details on each provider's official pages before implementing. Hour and cost figures are DevNJ's working assumptions, not measured statistics, and vary with site size and team set-up.