Cross-border selling is one of the most reliable growth levers an online retailer has: the catalogue already exists, and each new market is incremental demand. But the moment you serve more than one language, your search box becomes the part of the store most likely to quietly fail. A shopper in Munich, Lyon, and Warsaw will describe the same product in three different ways — and if your search only understands one of them, you lose the others without ever knowing.
Why multilingual search is harder than it looks
Translating your interface is the easy part. The genuine difficulty is that search has to bridge the gap between how shoppers phrase queries and how your catalogue is written — and that gap widens with every language.
A few specific problems recur:
- Catalogue and query languages diverge. Your product data might be in English while shoppers search in their own language, so nothing matches.
- Translation is lossy. A literal product-feed translation often produces phrasing no native speaker would type.
- Morphology varies wildly. German compounds words (“Regenjacke”), Slavic languages inflect heavily, and Romance languages carry gender and accents. Naive matching breaks on all of this.
- Mixed-language behaviour is normal. Shoppers routinely search brand names and technical terms in English even when browsing in their own language.
The result is a high zero-result rate in non-primary markets, which looks like weak demand when it’s really a search failure. This is the same dynamic we describe in fixing zero-result searches, amplified across languages.
Decide on a translation and indexing strategy
Before tuning anything, you need a clear model for how language flows through search. Three approaches dominate.
1. Per-language indexes
Maintain a separate search index for each market, each with translated product data and language-specific processing (stemming, stop words, synonyms). This gives the best relevance but the most operational overhead, and it’s the right choice for your priority markets.
2. Translate the query, not the catalogue
Keep one index and translate incoming queries into the catalogue language. Cheaper to run, but quality depends heavily on translation accuracy for short, ambiguous queries — exactly the kind machine translation handles worst.
3. Cross-lingual semantic search
Match meaning across languages directly, so a French query can retrieve products described in English without an explicit translation step. This is where embeddings change the economics — a single multilingual vector space lets one index serve many languages. Our primer on vector search explained covers how this works, and semantic search for eCommerce explains why meaning-based matching outperforms keywords here especially.
For most retailers the pragmatic answer is a hybrid: per-language processing for your top two or three markets, and cross-lingual semantic matching as the safety net for everything else.
Language-specific processing that actually matters
Generic search engines apply English assumptions by default. To serve other markets you need:
- Correct stemming and lemmatisation per language, so inflected forms collapse to the same root
- Compound splitting for German and Dutch, so “Gartenmöbel” matches “Möbel” and “Garten”
- Accent and diacritic folding, so “cafe” finds “café” and vice versa
- Per-language synonyms, because a synonym list never translates one-to-one
- Local stop words, which differ in every language
Skipping these is the most common reason a search that works beautifully in English collapses in German or Polish.
Handling the messy realities of cross-border shoppers
Real queries are rarely clean. Build for the edge cases:
- Mixed-language queries. Allow brand names and model numbers to match regardless of the interface language.
- Typos in non-Latin and accented scripts. Typo tolerance must be language-aware, not just edit-distance on ASCII.
- Regional vocabulary. “Trainers” versus “sneakers”, “jumper” versus “sweater” — and the equivalent splits within German-speaking or Spanish-speaking regions.
- Currency, units, and sizing. A search experience that returns the right product but shows the wrong size system or currency still feels broken. Localise the result page, not just the matching.
Don’t stop at search — localise the journey
Search is the entry point, but conversion depends on the whole localised experience. Recommendations, merchandising, and ranking should reflect local demand, not your home market’s bestsellers. A product that leads the page in Estonia may be irrelevant in Spain. Apply the same merchandising controls per market so local teams can boost what sells locally.
Practical localisation checklist:
- Currency, taxes, and shipping shown in local terms
- Local sizing and unit conversions
- Translated, native-quality product titles and key attributes
- Market-appropriate ranking and recommendations
- Local payment methods surfaced early
Measuring search across markets
Aggregate metrics hide market-level failure. A healthy overall search conversion rate can mask a market where search barely works. Break every metric down by language and market:
- Zero-result rate per market — your fastest signal of a broken language
- Search conversion rate per market, compared against your primary market as a baseline
- Search-to-click, to catch cases where results return but relevance is poor
- Top zero-result queries per language, which feed both synonym lists and catalogue gaps
When a secondary market shows a zero-result rate several times higher than your home market, that’s not low demand — it’s a fixable defect.
Common pitfalls
- Trusting raw machine-translated product feeds as if they were native copy. They mistranslate attributes and produce phrasing no one searches for.
- One synonym list for all languages. Synonyms are deeply language-specific.
- Ignoring compound and inflected forms, which silently kills recall in German, Dutch, and Slavic markets.
- Measuring search globally, so a broken market never surfaces in the numbers.
- Localising the UI but not the ranking, leaving every market with the home market’s bestsellers up top.
Where to begin
Pick your single largest non-primary market and pull its search logs. Look at the zero-result queries first — they’ll tell you immediately whether the problem is translation, morphology, or genuine catalogue gaps. Fix language-specific processing and synonyms for that market, measure the change in zero-result and conversion rates, then repeat for the next market. Cross-lingual semantic search is worth adding once you’re serving more languages than you can hand-tune.
Done well, multilingual search turns “we technically ship there” into a market that actually performs. If you’re expanding across borders and want search that works in every language you sell in, our AI search and recommendations work is built for exactly this. Get in touch and we’ll review your per-market search data together.