Most localization projects start with a spreadsheet and a translation vendor. Six weeks later the German storefront is live, the copy reads well, and the first order fails because the customer's postcode belongs before the street name and the address form will not accept it.
That is the usual shape of the failure. Translation is the visible part of entering a new market. The parts that break are money, addresses, tax, and URLs, and all four get decided by the data model months before anyone writes a headline.
Multi-market commerce is an engineering problem with a marketing surface. What follows covers the decisions that carry weight: how to store and display money, how to make content translatable without duplicating every template, how to survive the address and tax rules of the next country, and how to stop search engines from treating regional storefronts as duplicates of each other.
Model the market before the language
Language and market are separate axes, and conflating them is the first mistake. German is spoken in Germany, Austria, and Switzerland. Those three countries have different tax rules, different shipping realities, different payment preferences, and in the case of Switzerland, a different currency. A site that ships one "German" version serves a shopper in Vienna a price with German VAT at 19 percent baked in, when Austrian VAT is 20.
The fix is a market record that everything else points at. Each market carries a currency, a tax regime, a price list, a payment method set, a shipping zone, a default locale, and the legal text it requires. The storefront, the checkout, the invoice template, the email footer, and the sitemap all read from that record.
Once the market exists as a first-class row, adding Austria stops being a content project and becomes a configuration task with a tax table attached.
Store money as integers, price it per market
A price is not a float. Storing 19.99 as a double works until totals disagree by a cent, at which point finance notices and the ledger stops reconciling.
The safe representation is a value object: an integer amount in minor units plus a currency code, with the exponent read from ISO 4217 instead of assumed. This is where the usual * 100 shortcut fails. JPY, KRW, VND, CLP, and ISK have no minor unit, so multiplying by 100 turns a 1,000 yen item into a 100,000 yen charge. KWD, BHD, OMR, JOD, and TND have three decimal places, so the same shortcut undercharges by a factor of ten. Payment processors expect integer minor units and will reject or misread the request either way.
Keep money in one currency per value and make cross-currency addition impossible at the type level. Conversion happens at two boundaries only, display and settlement, and in both cases the rate used belongs on the order so the number can still be explained a year later.
A converted price is not a price
Finance teams do not want 49.00 EUR converted live into whatever the pound did this morning. They want a price list per market: 49.00 EUR for the eurozone, 44.00 GBP for the UK, 7,900 JPY for Japan. That list carries a margin floor, a rounding rule per currency, and a review cadence.
Rounding rules are not cosmetic. Swiss retail prices round to 0.05. Hungarian prices commonly end in 90 or round to 0.50. A conversion that produces 43.87 GBP reads like a math problem rather than a price. Store the rounded market price, not the conversion output.
Live conversion also produces a visible inconsistency, because the product page and the cart are separate requests. A shopper who sees 52.10 at 09:58 and 52.14 at 09:59 has a reason to leave, and support gets the ticket.
Tax inclusion changes the display, not just the arithmetic
In most EU consumer markets the displayed price must include VAT. In the US, prices are quoted before tax and the tax is calculated at checkout from the delivery address. Those are opposite display rules pointed at the same catalog.
A single price column cannot serve both. The model needs the net amount, the tax rate that applies to a given market, and a display rule that decides whether tax sits inside the price or is added at the end. The invoice needs all three as separate lines, because a business customer in Germany will ask for the net amount and the VAT rate apart.
Content needs keys and locales, not extra columns
The shortcut every team tries first is a column per language: name_fr, name_de, description_es. It holds for about six fields. Then the seventh language arrives, the template fills with null checks, the admin form grows wider than the screen, and the export feeding a shopping feed has a different shape for every country.
A translation store holds up better. Product text lives in rows keyed by entity, field, and locale, with a fallback chain that resolves fr-CA to fr to en. The fallback has to be visible. An editor should see at a glance that the Spanish short description is currently serving English text, rather than learning it from a customer complaint.
Do not build sentences by concatenation
String concatenation survives English and fails everywhere else. "You have 3 items" becomes "Du hast " plus the number plus " Artikel", and then breaks in Polish, which has four plural categories, and Arabic, which has six. Word order changes, gendered articles change, and the number needs locale formatting: 1,234.56 in en-US is 1.234,56 in de-DE.
ICU MessageFormat is the mature answer, applied on the server so the same message renders in the storefront, the confirmation email, and the invoice. Dates need the same treatment. 03/04/2026 is March 4th to an American and April 3rd to almost everyone else, which is the kind of ambiguity that ends up printed on a delivery estimate.
Addresses, phones, and the checkout edge cases
Every country has its own address syntax, and a form built for the US and Canada starts failing the moment it crosses an ocean.
- Japan orders the address largest to smallest: postal code, prefecture, city, then block. A form that puts street before city produces shipping labels that come back.
- The Netherlands places the postcode before the house number, and the house number often carries an addition like 12A.
- Ireland has no postal codes across most of the country, so a required postcode field blocks an entire market.
- The UK postcode is alphanumeric with a mandatory internal space, and a regex that strips spaces breaks label printing.
- Brazil wants a CPF on the invoice, and the tax ID appears on the printed document.
The approach that holds is to store the address as structured fields plus a country code, and let per-country metadata drive the form: which fields exist, the order they appear in, whether each is required, the validation pattern, and the label. Google's libaddressinput data and the CLDR address metadata both exist for this and stay better maintained than anything written by hand. Keep the raw user input next to the normalized version so support can resubmit a corrected address without retyping it, and be careful with normalization that silently drops an apartment number.
Phone numbers follow the same rule. Store E.164, parse and format with libphonenumber, and never assume ten digits.
Payment methods differ by market, and each one carries its own redirect flow and failure modes. iDEAL in the Netherlands, Bancontact in Belgium, BLIK in Poland, PIX in Brazil, Konbini and PayPay in Japan, Klarna and SEPA direct debit across much of Europe. A checkout with card and PayPal only is a checkout that loses a large share of European orders, and each added method is an integration with its own sandbox, settlement timing, and refund semantics.
URLs and hreflang, generated rather than maintained
Subdirectories such as /de/ and /fr-ca/ are the right default. Country-code domains are separate sites with separate authority, which means every new market starts its search visibility from zero, and subdomains sit in between. Pick one structure before the second market ships, because migrating later means redirect maps across every locale at once.
hreflang does not translate content and does not rank it. It routes a ranking that has already been earned to the right regional version. Four mistakes break most implementations, and all four fail silently:
- Annotations are not reciprocal. If the German page points at the English one and the English page does not point back, the annotation is ignored.
- Canonicals conflict with the annotations. A French page that canonicalises to its English equivalent has its hreflang ignored, so every regional page needs a self-referencing canonical.
- Codes are wrong. en-UK does not exist, the correct form is en-GB. Script variants belong as zh-Hans and zh-Hant rather than zh-CN, and pt-BR is not pt-PT.
- Two methods run at once. Head tags plus an XML sitemap describing the same set of pages is a conflict, not redundancy.
Generate the annotations from the table that already knows which pages are translations of each other. On a catalog with ten markets and 30,000 products, hand-maintained hreflang is roughly 330,000 link elements per crawl and a broken cluster the first time someone renames a slug.
x-default covers the visitor whose language you do not serve at all. Point it at a language selector or your primary market. It is a fallback, not a statement of which version is best, and pointing it at the biggest market is a common way to make the tag do nothing useful.
Crawl budget is the other half of the problem. Ten markets times 30,000 products is 300,000 URLs that mostly differ by a translated product name. Either keep the thin ones out of the index or accept the crawl cost, but decide deliberately, and make sure each locale's hub links to the pages that matter.
The recurring cost nobody budgets for
Translation is the cheap line item. The expensive part is that localization turns one release into many.
Every new product attribute becomes a translation task that never ends. A sprint that adds three fields is three fields across every locale, permanently. Legal text drifts by jurisdiction: the German Impressum, EU withdrawal rights, cookie consent rules that differ between member states, and the European Accessibility Act, which has applied to a broad set of consumer products and services since June 2025. Support has to exist in the local language and the local time zone. Return policies differ, and the EU's fourteen-day right of withdrawal is not universal.
None of that is avoidable. All of it is cheaper to estimate when the market record, the translation store, and the price lists already exist, because each new requirement lands in a table instead of a template.
A build order that survives the third market
The sequence matters more than the tooling.
- Create the market record with currency, tax regime, price list, payment methods, and legal text assignment.
- Replace float prices with a money value object that reads ISO 4217 exponents and refuses cross-currency arithmetic.
- Move translatable strings into a translation store with an explicit fallback chain and coverage reporting.
- Drive address and phone forms from country metadata instead of per-country templates.
- Generate URLs, canonicals, and hreflang from the market and locale table, using one method only.
- Add markets one at a time, with the tax and payment integrations for each tested end to end before launch.
Two markets can be shipped out of order with enough heroics. The third one exposes whatever was skipped, usually in the checkout, usually in the middle of peak season.
What to measure
Instrumentation is how a localization effort stops being a feeling and becomes a number.
- Conversion rate by market, split by payment method, because a broken local method shows up as a drop concentrated in one country.
- Checkout error rate by country, which surfaces address validation failures before customers give up.
- Organic impressions and clicks per locale in Search Console, filtered by country, to confirm hreflang is routing traffic rather than consolidating it.
- Share of orders needing an address correction after purchase, the clearest signal that the form is fighting the local format.
- Translation coverage per locale, plus the age of the oldest untranslated string in a live template.
- Contribution margin per market after FX, payment fees, and tax handling.
The first market is a project. The second is the real test of the architecture. If the third one takes a quarter of engineering time, the translation vendor was never the bottleneck.