Zalando’s latest search experiments are less about chat for its own sake and more about collapsing the distance between intent and outfit. Users describe occasions, constraints, and moods; the system has to translate language into SKUs without breaking brand tone or stock reality.
The hard part is not the model. It is the catalog: size curves, regional availability, sustainability attributes, and the editorial tags that make fashion search feel curated instead of mechanical.
Teams close to the work say the win condition is confidence — answers that cite why a look works, not only what to buy. That requires eval harnesses built with stylists, not only engineers.
For the industry, Zalando is a signal: conversational commerce in fashion will be won on data quality and taste boundaries, not on bigger prompts alone.
The concrete signal is not abstract. Zalando moved its AI assistant across its 25 markets in local languages, and OpenAI later described the assistant as a product-discovery layer built with Zalando’s own models and GPT-4o mini. That matters because it makes conversational search less like a lab feature and more like a customer habit.
Imagine the request: “I need something for a city wedding in Amsterdam, not too formal, good if it rains, and I hate tight sleeves.” A normal filter tree does not know how to hold all of that at once. It can handle size, color, price, brand. It struggles with mood, weather, insecurity, and the small social codes that make fashion shopping human.
For the team behind the assistant, the work becomes very specific. Someone has to map “not too formal” to silhouettes, fabrics, heel height, colors, and brands. Someone has to decide whether the assistant should suggest a raincoat when the customer asked for a dress. Someone has to make sure products are actually in stock in the right market. Someone has to check whether the explanation sounds helpful or strangely robotic.
This is where personal taste enters the system. A stylist may know that a satin slip dress and a sharp blazer can solve the wedding request. A data team can help the assistant find those products. A merchandiser can say which items should not be pushed because returns are high. Customer service can say which promises usually create disappointment.
The lesson is simple: conversational search is not a chat window added to a catalog. It is the catalog being asked to speak. If the catalog is poor, the answer will be poor. If the organization has no shared language for fit, occasion, mood, and availability, the assistant will expose that weakness quickly.
The next advantage will belong to teams that treat search as editorial work as much as technical work. The user is not asking for a query result. She is asking to be understood.
