What I learned building a 12-language AI product search for AliExpress

A developer built OneFindMe, a free multilingual search layer for AliExpress that resolves shopping intent to a canonical product query before translating, rather than translating raw phrases. The system trims descriptive queries to a product noun plus one or two qualifiers, ranks candidates by real order volume with a light penalty for unreviewed listings, and caches query resolution to cut latency. It supports 12 languages and applies locale-specific filtering for shipping thresholds and regional defaults.

AliExpress has one of the largest product catalogs on the planet, but its search assumes you already know the exact English keyword. If a shopper can only describe what they want — "that little stone thing for face massage" — the results collapse into noise. I spent a while building OneFindMe https://onefindme.com/en/ , a free multilingual search layer that sits in front of AliExpress, and these were the lessons that actually moved the needle. Naive translation is a trap. "לק ג'ל" Hebrew machine-translates to "gel polish", which on AliExpress surfaces floor lacquer as often as nail products. The fix was to resolve the shopping intent to a canonical product query first, then translate that — not translate the raw phrase. A small curated keyword map beat the general model for the high-traffic terms, because the model kept inventing plausible-but-wrong category ids. Long, descriptive queries rank worse than short ones on the marketplace API. The product noun has to lead. "comfortable running shoes for wide feet women" performs far worse than "wide running shoes". So the pipeline trims to the product noun + one or two qualifiers before it ever hits the API. The cheapest listing is almost never the best answer. Sorting candidates by real order volume, then lightly penalising listings with no reviews, produced results people actually clicked. Price is a filter, not a ranking signal. Product availability changes hourly, but the translation of "wireless earbuds" into a good query does not. Caching the query resolution and letting unknown terms cache their translation on first use, so the second shopper gets an instant answer cut latency dramatically without serving stale stock. The same query needs different handling per country: shipping thresholds, which categories pay, even modesty defaults in some regions where the marketplace's own "relevance" surfaces things a shopper did not ask for. Filtering the junk while never filtering a legitimate intent turned out to be the hardest, most locale-specific part. If you want to see the result, OneFindMe https://onefindme.com/en/ is live and free — type what you want in any of 12 languages and it does the translating, trimming, and ranking described above. Happy to answer questions about any of these in the comments.