International SEO Architecture: Getting Multilingual Sites Right
Most multilingual SEO failures are architectural, not linguistic. Here is how to structure, signal, and localise a global site correctly.
Expanding into new markets is usually framed as a translation project. It is really an information architecture project with a translation component. The sites that struggle almost never fail because the copy reads awkwardly — they fail because engines cannot tell which variant belongs to which audience, or because six locales quietly compete for the same queries.
Choose the URL structure first
This decision is expensive to reverse, so make it deliberately.
| Structure | Example | Strengths | Costs |
|---|---|---|---|
| Subdirectory | example.com/de/ | Consolidates authority, simplest to operate, fastest to launch | Weaker local signal than a country domain |
| Subdomain | de.example.com | Independent hosting and stacks, clean separation | Authority does not transfer as reliably |
| ccTLD | example.de | Strongest local trust and geo signal | Separate authority to build per domain, high operational overhead |
For most organisations, subdirectories win. Choose ccTLDs only when local trust materially affects conversion — regulated services, marketplaces, or markets where users actively check the domain.
Language versus market
Separate the two dimensions. Language targets how a person reads; market targets where they buy. German content for Germany, Austria, and Switzerland shares a language but differs in price, currency, tax, and often legal disclosure. Structure as /de-de/ and /de-ch/ when commercial detail diverges, and keep a single /de/ when it genuinely does not.
Implement hreflang precisely
Hreflang is where most technical audits find breakage. The rules are simple but unforgiving:
- Reciprocity. Every page in a cluster must reference every other, including itself.
- Absolute URLs. Relative paths are ignored.
- Canonical agreement. A page's self-canonical must match its hreflang self-reference, or the entire cluster is discarded.
- Valid codes. Language in ISO 639-1, region in ISO 3166-1 alpha-2.
en-ukis invalid;en-gbis correct. - x-default. Point it at the language selector or global fallback for unmatched users.
Choose one delivery method — HTML head tags, HTTP headers, or the XML sitemap — and use it consistently. Sitemap-based annotation scales best for large catalogues because it avoids bloating every page's head.
Localise deeper than words
Translation moves text; localisation moves intent. Keyword research must be redone per market, because literal translations of high-volume English terms are frequently not what local buyers search. Beyond copy, localise currency and tax display, payment methods, date and address formats, units, imagery, testimonials and case studies, and legal or compliance content.
On translation quality: machine translation with professional post-editing now performs on par with fully human translation for most commercial content at materially lower cost. Raw unreviewed output does not, and it is the main reason multilingual sections get treated as thin content.
Geo-detection without self-sabotage
Never automatically redirect based on IP address. It breaks crawling from any single-country crawl location, traps travellers in the wrong locale, and hides content behind logic engines cannot follow. Instead, keep every locale on a stable, crawlable URL and offer a dismissible suggestion banner that respects the user's choice via cookie or account preference.
The failure modes to audit for
- Cross-locale canonicals. Pointing
/fr/pages at their English equivalents deindexes the French site. - Partial coverage. Translating templates but leaving product descriptions, filters, and error states in English.
- Duplicate near-identical locales. Multiple English variants with no differentiation splitting authority.
- Untranslated URLs and metadata. Slugs, titles, and descriptions left in the source language.
- Crawl budget waste. Locale parameters generating combinatorial URL explosion.
- Local link poverty. No in-market references, so the locale never accrues independent authority.
A sane launch sequence
Do not launch six markets at once. Prove the pattern in one: run market research and keyword mapping, build one fully localised locale including commerce and support content, verify hreflang and indexation for four to six weeks, then template the process. Rolling out a broken pattern across ten locales multiplies the remediation cost by ten.
Frequently asked questions
Subdirectories or subdomains?
Subdirectories for most organisations, because they consolidate authority and are simplest to operate. Use subdomains or ccTLDs when separate infrastructure, distinct branding, or strong local trust signals are required.
Is hreflang required?
Not strictly, but without it engines frequently serve the wrong variant or treat translations as competing duplicates. For any site with more than two locales it is effectively mandatory.
Is machine translation acceptable?
With professional post-editing and locally researched keywords, yes. Published raw, it usually underperforms and can be treated as low-value content.
How long until a new locale ranks?
Expect three to six months for meaningful visibility on a subdirectory inheriting domain authority, and considerably longer for a fresh ccTLD building authority from zero.