All posts
Liferay · 12 min read · Sep 20, 2026

Localization in a Multi-Region Portal Is Not a Translation Problem

A portal serving several languages and regions

A request to make the portal available in five languages sounds like a translation exercise with a clear finish line. The interface strings are extracted, sent out, returned, and loaded, and the portal is demonstrably multilingual. Then the questions start arriving, and almost none of them are about language.

Who decides what appears in each region. Whether a French user in Belgium should see the same page as a French user in France. What happens when a page exists in English but not in Polish. Whether the Italian legal notice may be shown in English while translation is pending. These are governance and data questions, and they are the actual project.

Language and region are two axes

The single most useful distinction to draw early is that language is what someone reads and region is which market's rules, content and offers apply to them. They are correlated and independent.

UserLanguageRegionWhat they should see
Employee in BrusselsFrenchBelgiumFrench text, Belgian policies
Employee in ParisFrenchFranceFrench text, French policies
Expatriate in BerlinEnglishGermanyEnglish text, German policies
Visitor from outside coverageEnglishNoneEnglish text, global content only
The same language with different regional content is the case that breaks naive designs.

A design that treats the locale as the only dimension cannot express the first two rows, and the discovery usually happens after the content model is built. Deciding whether your sites are organised by region, by language, or by both is a structural decision, not a configuration one.

Three kinds of content, governed differently

  • Global content, authored once centrally and translated - brand, group policy, product information. One owner, many languages.
  • Regional content, authored per region within a shared structure - pricing, availability, local contacts. Many owners, one template.
  • Local-only content, existing in one region and nowhere else - a national campaign, a country-specific obligation. One owner, no translation.

Most disputes in these projects are really disagreements about which bucket a piece of content belongs to. Central teams want more global content for consistency; regional teams want more local content for relevance. That negotiation is better held explicitly at design time than rediscovered as a workflow problem later.

Fallback is a policy, not a default

When content is missing in a requested language, something has to be shown. The platform will do something sensible by default, and sensible is not always acceptable.

ContentFalling back isWhy
Marketing copyFineSlightly worse experience, no consequence
Product descriptionsUsually fineBetter than an empty page
Legal terms and noticesNot acceptableMay be unenforceable or non-compliant in that market
Pricing and availabilityDangerousAnother region's figures are simply wrong
Safety or regulatory instructionsNot acceptableThe reason the requirement exists
Some content must be hidden rather than shown in the wrong language.

The design consequence is that fallback has to be expressible per content type, with a defined behaviour when fallback is forbidden - hide the section, show a notice, or block the page. Choosing silently at the platform level is how a portal ends up serving one market's terms to another's customers.

Everything else that carries language

Interface strings are the part with a process attached. The gaps are always in the places where text is generated rather than authored.

  1. Notification and email templates, which are frequently written once in the project language and never revisited.
  2. Workflow task and stage names, which appear directly in the user's task list.
  3. Picklist and reference data labels, which the interface renders verbatim.
  4. Validation and error messages, including those produced by server-side rules.
  5. Generated documents and letters, which often matter most and are localised last.
  6. Anything assembled by concatenating fragments, which cannot be translated correctly in any language with different word order.

The last one is a genuine engineering constraint rather than an oversight. Building a sentence from pieces produces text that cannot be translated properly, because grammar is not a function of concatenation. Full sentences with named placeholders are the only reliable unit.

Locale-aware data, not just locale-aware text

Beyond language, a locale changes how data is formatted, ordered and validated - and these are where subtle, hard-to-reproduce bugs live.

  • Dates: ordering, separators, and whether the day or month comes first, which produces wrong data rather than ugly data.
  • Numbers: the decimal separator differs, so parsing a typed figure without a locale is a correctness problem.
  • Currency: symbol, position, and the fact that a figure without an explicit currency is meaningless across regions.
  • Sorting: alphabetical order is language-specific, so a list sorted by byte value is wrong in most languages.
  • Addresses, postcodes and phone formats, which cannot share one validation rule.
  • Names: assumptions about given and family names hold in fewer places than most forms assume.

The number separator case is the one that reaches production most often, because it is invisible in testing done in one locale and produces a value that is wrong by a factor rather than rejected.

Search has to be language-aware too

Indexing several languages into one field with one analyser produces results that are plausible and poor - stemming rules differ per language, so a German query will not match what a German user expects. Language-specific fields or indexes are the standard answer, and the cost is that changing this later requires a full reindex, which makes it an early decision in the search design rather than a tuning exercise.

Translations go stale silently

The failure mode of a translated portal is not missing text; it is text that was translated correctly a year ago and no longer matches the source. Nothing breaks, nothing is reported, and two markets are quietly describing different things.

Preventing it needs a relationship between the source and its translations, so a change to the source marks the translations as needing review. Without that link there is no way to answer which translated pages are out of date, and the answer is invariably more than anyone expects.

The hardest differences are not linguistic

The requirements that break a multi-region design are usually the ones where regions differ in substance rather than wording: one market requires an extra approval, another requires a field that elsewhere is forbidden, a third has a different retention period. None of this is localisation, and trying to solve it with translation produces a system where a single workflow and data model is bent out of shape by regional exceptions.

The workable approach is to treat regional variation as a first-class part of the model - configuration per region for required fields, approval chains and retention - rather than as special cases in code. It costs more initially and is the difference between adding the sixth market in a sprint and in a quarter.

Conclusion

Translation is the part of localisation with an obvious process and the smallest share of the risk. Separate language from region, decide explicitly which content is global, regional or local, make fallback a per-content-type policy with a defined behaviour when it is forbidden, localise the generated text as well as the authored text, treat formatting and sorting as correctness concerns, and keep translations linked to their sources so staleness is visible. Then handle the regional differences that have nothing to do with language, because that is where the real work has been waiting.

Contact

Let’s connect

Have a Liferay or Backend project you’d like to talk through? Fill in the form and I’ll get back to you.

Book Calendar