Why Multilingual Product Search Still Fails—and What AI Can Do Differently

Online marketplaces may display their interfaces in many languages, but that does not necessarily mean their product search truly understands those languages.

A shopper searching in English will often receive better results than someone describing the same product in Spanish, French, or Arabic. This happens even when the marketplace officially supports all of these languages.

The problem is deeper than translation

Marketplace listings are rarely written in clean, natural language. Product titles may contain:

• automatic or inaccurate translations;

• repeated keywords;

• technical supplier terminology;

• different names for the same product;

• spelling and grammatical errors.

Because of this, simply translating a user’s query into English may not be enough.

For example, a shopper might naturally describe what they need as:

“A small beige crossbody bag with an adjustable wide strap for everyday use.”

The relevant listing might instead contain a title such as:

“Women Fashion Shoulder Messenger Bag Casual Wide Belt Handbag.”

A traditional keyword-based system may fail to understand that these descriptions refer to a similar product.

Searching by intent instead of exact words

While developing OneFindMe, a multilingual product-search project for AliExpress, I began exploring whether AI could interpret the shopper’s actual intent before searching the marketplace.

Instead of matching only individual keywords, a search system could extract attributes such as:

• product type;

• colour and material;

• intended use;

• preferred style;

• size requirements;

• budget;

• delivery preferences.

The system could then generate several relevant search interpretations rather than relying on a single literal translation.

This approach may also help when users do not know the correct name of a product. They can simply describe what they saw, how they intend to use it, or upload a photograph.

Combining text and image search

Natural-language and image search solve different parts of the same problem.

An image can reveal shape, colour, style, and visual similarity. A written description can provide information that is not visible, such as budget, size, material preferences, or intended use.

Combining both inputs may produce better results than relying on either method alone.

For example, a user could upload a photograph of a bag and add:

“I want something similar, but in black, under $30, with delivery to Spain.”

The system would need to understand the image, the requested changes, the price constraint, and the shopper’s location.

The technical questions

Building this type of search raises several interesting questions:

1. Should the query first be translated into the marketplace’s dominant language?

2. Is multilingual semantic search more reliable than generating several keyword variations?

3. How should embeddings handle badly translated and keyword-heavy product titles?

4. How can relevance be evaluated when marketplace data changes constantly?

5. Should the system explain why each result matches the request?

I believe the biggest opportunity is not simply translating existing search. It is creating a layer that understands how people naturally describe what they want and then connects that intent with the unusual language used in marketplace listings.

I would be interested to hear from developers working with multilingual embeddings, semantic search, computer vision, or marketplace APIs.

How would you approach this problem?