Cross-border commerce has one very easy trap: translate the product titles and detail pages, publish, and wait for orders. Traffic arrives, but add-to-cart rate sits well below the home market. The problem is rarely translation quality. Translation solved comprehension; it did not solve willingness to buy.
Every funnel stage carries a language cost
Pull the funnel apart: search impression, click through, browse, add to cart, enter address, choose payment, complete. Language causes drop-off at every stage, but for entirely different reasons at each one.
Search-stage loss happens because your keywords do not match how people actually phrase things locally -- a German shopper searches for a term you would never reach by direct translation. Product-page loss happens because specifications are unreadable: sizes, materials, and voltages were never converted to local conventions. Checkout loss is the most hidden of all -- field order, postal code format, and payment method naming that do not match local habit will lose the sale outright.
Product titles are rewritten, not translated
A title has to do two jobs at once: match search queries and earn the click. Direct translation almost always fails both. Marketplace rules differ too -- some cap character counts, some require the brand first, some penalize keyword stuffing.
A workable process is to take a machine translation as the semantic baseline, correct the core terms against local keyword data, then restructure to the marketplace rules. Push the result into a glossary: once brand terms, category terms, and key attributes are fixed, new listings inherit them instead of being re-decided each time. That is exactly what the glossary feature is for.
Separate what must be exact from what may flex
Detail pages should not be handled with one standard. Specifications, ingredient lists, safety warnings, and return policies must be literally accurate -- one wrong figure can be a compliance issue. Selling points, use cases, and brand narrative should be free to be rewritten, and often should differ by market entirely.
Technically you can split this with style rules: strict output for specification blocks, freer tone for marketing blocks. Style rules support per-content-type configuration, which avoids applying one setting to an entire catalog.
Reviews and Q and A need handling too
User-generated content carries real weight in commerce conversion, and it is the part localization workflows most often skip. A page of foreign-language reviews erodes trust, but review volume also feeds ranking, so hiding them costs more.
The safer pattern is to keep the original, offer a translation alongside it, and label the translation clearly as machine-generated. You keep the volume without misleading anyone through an imperfect rendering. At scale, run this asynchronously through the translation API -- it is the only version of this that is economically viable.
Prove the return with A/B tests
Localization budget is hard to defend because the return is usually asserted rather than measured. Make it measurable: same market, same set of products, one group on the original translation and one on the improved localization, compared on add-to-cart and conversion.
In practice the biggest wins are rarely in product copy. They are in the parts that look least important -- shipping and duty explanations, expected delivery windows, and the returns process. On a domestic store these are fine print. Cross-border, they decide whether someone is willing to order at all. Test those first; the return is usually far better than rewriting descriptions.
The goal was never translation complete. It is getting each market close to home-market conversion. Quantify each funnel stage, find the widest gap, and invest there -- it beats translating everything once by a wide margin. For more on operating at scale, see scaling websites internationally, or review the enterprise solutions.