Multilingual Product Search for a Pan-European Retailer

One catalogue, many languages of query

A pan-European retailer gets searches in a dozen languages against a catalogue whose content may be in a few. "Robe rouge," "rotes Kleid," and "red dress" should all reach the same products. Multilingual product search is about making the language of the query independent of the language of the catalogue, cleanly, rather than translating queries and hoping.

The core change from a single-language search is the embedding model. A multilingual model places a phrase and its translations near each other in vector space, so a French query retrieves German-described products directly. Our semantic search project notes this explicitly as the point where the smaller English-focused model is outgrown and a larger multilingual one earns its cost.

Translation as a fallback, not the strategy

Machine-translating every query into the catalogue language and matching lexically works, but it stacks a translation error on top of a retrieval error and loses nuance. It is a reasonable fallback for a rare language with no embedding coverage, not the primary approach for markets that matter.

Per-language relevance tuning

Synonyms, stopwords, stemming, and common misspellings are language-specific. Relevance that is well tuned for English will underperform in Polish or Finnish until each language gets its own attention. The merchandising console has to support per-language rules, not one global set.

Locale-correct results

Currency, units, sizing conventions, and availability differ by market. A correct product match shown with the wrong currency or an unavailable-in-this-country item is still a poor result. Locale has to flow through ranking and presentation.

One index or many

A single multilingual index is simpler to operate and lets cross-language queries work naturally; per-market indexes give more control and isolation at the cost of duplication. The right choice depends on how different the assortments and rules are between markets.

Where the data limits it

Cross-language retrieval is strongest when product content is rich in at least one language and consistently structured. Sparse or inconsistent catalogue text hurts more here, because the model has less to align translations against.

A worked example

A retailer sells across France, Germany, and Poland with product content mostly in English. A shopper in Lyon searches "robe rouge été." A multilingual embedding model retrieves the English-described summer dresses directly, and the result page shows prices in euros with French sizing. A Polish shopper's query for the same category works the same way. Without the multilingual model, each query would be machine-translated first, stacking a translation error onto retrieval and losing the seasonal nuance in "été."

Rolling out per market

A practical sequence is: launch with the multilingual model and a shared index, then tune each market's synonyms, stopwords, and merchandising rules in priority order by revenue. Locale formatting — currency, units, sizing — comes with the market launch, not after. Treating every new language as a full sub-project is what makes multilingual search feel perpetually unfinished. A useful checkpoint before each market goes live is a small set of real queries in that language, reviewed by a native speaker, to catch the cases where the model retrieves a plausible-looking but wrong product because a term is a false friend across languages.

Where this stops being right

• A single-language market does not need any of this — a monolingual model is smaller, cheaper, and often more accurate for that language.
• Right-to-left and non-Latin scripts add rendering and tokenisation work beyond the embedding model choice.
• Legal and labelling requirements for product information vary by country and are not a search concern to paper over.

FAQ

Do I need a separate search per language? Usually not. A multilingual embedding model lets one index serve many query languages. Per-market indexes are a choice for isolation and control, not a requirement.

Is translating the query good enough? As a fallback for rare languages, yes. As the main approach it stacks translation errors on retrieval errors and loses nuance — a multilingual model matching directly is better.

What still has to be done per language? Synonyms, stopwords, stemming, misspellings, and merchandising rules are all language-specific and need tuning per market.

ISTRALLEN builds multilingual search with a multilingual embedding model and per-language relevance tuning — see AI for Retail.