What this playbook is for
A multilingual content website fails when translation becomes a rescue job. The real challenge is not generating another language version. It is keeping one source of truth, one publishing system, and one quality bar across every market.
A multilingual site scales when localization becomes a workflow, not a weekly act of copy-paste survival.
Quick take
| Layer | Best tools right now | What they are actually good at | What to avoid |
|---|---|---|---|
| Source language system | Notion, structured content fields | Keeping title, summary, slug, audience, status, and update history in one canonical record | Letting each language version become its own independent page logic |
| Translation | ChatGPT, Claude, DeepL | Draft localization, preserving structure, adapting tone, and speeding first-pass translation | Assuming translation and localization are the same job |
| Localization operations | Lokalise, glossary rules, review queues | Managing terminology, status, and continuous localization when volume grows | Using a full localization platform before there is enough content flow to justify it |
| Website publishing | Astro i18n, Astro content collections, Vercel | Clean multilingual routing, predictable content structure, and deployment discipline | Shipping language versions through ad hoc folder sprawl |
Choose the right translation model first
| Approach | Best when | Main strength | Main tradeoff |
|---|---|---|---|
| LLM-first translation | You need nuance, restructuring, or editorial adaptation | Better at preserving intent and adjusting phrasing by audience | Needs stronger review for consistency |
| DeepL-first translation | You need fast, consistent, high-volume translation with glossary control | Very efficient for repeatable language work and document translation | May still need editorial adaptation for flagship copy |
| Lokalise-style workflow | You have enough content volume to justify structured localization ops | Status tracking, glossary, and continuous localization discipline | Too heavy if the site is still small and manually manageable |
The operating model
Translation gets easier when the source is structured
Every source-language page should have stable fields before translation starts:
- title
- summary
- slug
- audience
- status
- last updated date
- notes for translators or reviewers
This is the difference between a site that scales and a site that drifts.
Review by business importance
Pages that need the most human review first:
- homepage
- pricing or conversion pages
- flagship guides and reviews
- pages with claims, metrics, or legal sensitivity
Pages that can usually move faster:
- short updates
- lower-stakes support content
- archive items with stable structure
This one decision saves a huge amount of wasted review time.
The website layer should help operations, not fight them
Astro's i18n routing is useful when you want clean locale-aware URLs, fallback handling, and predictable route generation. Astro content collections are useful when you want structured content with repeatable fields. Vercel is useful when you want a lightweight deployment path for content-heavy sites.
The key is simple: content structure and route structure should agree with each other.
A practical multilingual publishing workflow
- Publish or update the source-language page.
- Generate the first-pass translation.
- Apply glossary and terminology fixes.
- Review only the pages that have high trust or high conversion impact.
- Push all updated locales in one deployment batch.
- Log which localized pages are out of date, missing, or underperforming.
What to standardize first
| Priority | Asset | Why it matters |
|---|---|---|
| 1 | Source schema | Structured source content makes every later step easier |
| 2 | Locale routing rules | Readers and search engines need predictable paths |
| 3 | Glossary | Core terms should not mutate across languages |
| 4 | Review priority system | Not every page deserves the same level of human attention |
| 5 | Sync cadence | Content drift is easier to prevent than to clean up later |
A simple stack by stage
| If you are at... | Suggested stack | Main result |
|---|---|---|
| Early-stage multilingual site | Notion + ChatGPT or Claude + Astro + Vercel | Fast localization with light human review |
| Content-heavy multilingual operation | Notion + DeepL + glossary workflow + Astro or your CMS + Vercel | More consistent translation and less manual cleanup |
| Higher-volume localization system | Structured source content + Lokalise + website publishing pipeline | Better status control and continuous localization |
Common mistakes
- Treating translation as a one-off task instead of an ongoing system.
- Giving every page the same review depth.
- Letting each locale invent its own slug or structure.
- Translating without a glossary once brand language starts mattering.
- Publishing updates one locale at a time and creating silent drift.
Checklist
Best source material from the old site and current project
Operator note
A multilingual content website starts feeling mature when translation quality is no longer the bottleneck and update discipline no longer depends on memory.