Translation is the last decision, not the first
Companies approach a multilingual website by asking which languages to translate into. The questions that decide whether it works come earlier: which markets you are selling to, how the URLs will be built, what happens to a visitor who lands on the wrong version, and who maintains the second language when the first one changes. Translate first and you get a site that reads well in four languages and cannot be found in three of them.
Language or market, decide before you build
Serving Spanish is not the same as serving Spain. If your product, delivery, terms or commercial approach differ by country, you are building market versions and the language follows from that. If the offer is identical everywhere and only the words change, you are building language versions. This decision sets your URL structure, your hreflang codes and how many versions you will be maintaining in two years. Companies that never made it explicitly end up with one Spanish page for Spain, Mexico and everyone else, and it converts in none of them.
Every version needs its own URL
The most damaging mistake is switching language without changing the address. A switcher that swaps content behind the same URL, or sets a cookie, gives search engines one page to index and buyers no link they can send to a colleague. Pick one structure and keep it: subdirectories under one domain, which is the simplest to run and inherits the authority of the main site, or country domains, which cost more and signal a local presence. Choose once. Changing later means redirecting every page you have.
hreflang is where most sites quietly break
These tags tell search engines which version belongs to which audience, and they only work when every version points at every other version, including itself. Miss one reference and the whole set gets ignored. The other common failure is a canonical tag on all versions pointing at the English page, which instructs search engines to disregard everything you translated. Use real region codes rather than invented ones. Then verify it, because this is invisible from the front of the site and the symptom is that the wrong version ranks in the wrong country.
Do not redirect people by location
Automatically sending visitors to a version based on their IP address feels helpful and causes problems. It traps travelling users, it overrides people who deliberately chose a language, and it stops crawlers seeing any version other than the one their servers appear to be in. Detect location if you want, then offer a suggestion the visitor can dismiss. Keep the choice visible in the header, label the languages in their own names rather than with flags, and let people stay where they landed.
Machine translation and where it stops working
Machine translation is good enough for support articles, documentation and pages nobody makes a decision on. It is not good enough for the pages that sell. Your homepage, your service descriptions, and anything carrying a claim or a legal term need a native speaker who understands the business, because the sentences that convert depend on connotation and the ones that expose you depend on precision. A thin translation of your key pages tells buyers in that market they were an afterthought.
Design has to survive a longer language
German runs long, Arabic runs right to left, Turkish stacks suffixes onto words that were already sized to fit. Buttons that sit comfortably in English break in translation, navigation wraps onto two lines, and headlines that were balanced turn ragged. Build the layout to flex before the translation arrives, and test the longest language rather than the shortest. Include the parts everyone forgets: form validation messages, error pages, automated emails, and the document you send after a form is submitted.
Launch narrow and prove it
A second language doubles the maintenance forever. Before committing to four, launch one market properly: the pages that sell, translated by a person, with correct structure and its own path to conversion. Then look at whether anything came of it. Most companies find that one well-executed market outperforms three half-built ones, and that the version they never update starts working against them within a year.
Written by
Idennex
Strategy-first agency, Istanbul