Smarter models, better results: The evolving AI models behind ZMS ads
We’ve rebuilt the AI models behind our two main ad products. The result: more relevant ads for shoppers, higher returns for partners, and lower costs per click.
Display Ads and Sponsored Product Ads from Zalando Partner Marketing Services (ZMS) now run on advanced deep learning models that can weigh many signals at once. This update marks the most significant architecture change since launch: a new technology, and models that now also “see” product images.
What this means for partners
Brands want their ads to reach shoppers who are actually interested, and to get the most out of every euro they spend. In randomised tests comparing our previous and new models, the new ones have improved results across both products.
Since the switch, partners have been seeing:
In short, partners’ budgets work harder. Shoppers see ads that fit what they’re looking for, and partners pay less for every click.
These gains start with shoppers. Whether someone is searching for running shoes or browsing new-season coats, the ads they see now fit that moment more closely – in style and in what they're actually looking for. That makes ads feel more like a part of the shopping experience. The higher click-through rates suggest shoppers find them more useful too.
The new models are now live and apply automatically to all Display Ads and Sponsored Product Ads campaigns – no action needed from partners.
Two goals, two solutions
Every campaign has a different job to do. A brand launching its new season's collection wants shoppers to discover it. A partner with a new sneaker launch may want to drive conversion in a specific period. That’s why we offer different ad products, each running on its own model. Here’s how our two main ad products work.
Display Ads cover our mid- to upper-funnel formats. They bring shoppers into a partner’s brand world. They show collections, curated boards, or awareness formats on the homepage, in the catalogue and in search results. The goal is engagement and awareness, so the model predicts how likely a shopper is to click through to a collection.
Sponsored Product Ads are our lower-funnel product. They help partners sell specific items by showing a single product in the catalogue and on product pages. The goal is a sale, so the model predicts two things: how likely a shopper is to click, and how likely they are to buy afterwards.
Why transformer models make the difference
Transformers are a type of AI model first developed to understand language. They’re the technology behind today’s large language models. Their key strength is called attention. Instead of reading each signal on its own, the model looks at all of them at once and learns which ones matter most in relation to each other.
Think of a stylist putting together a look. They don’t judge a jacket in isolation – they consider the shoes, the occasion and the person wearing it. Attention works in a similar way. The model weighs a product’s image, a shopper’s recent activity and the page they’re on together. That’s how it picks up the subtle style preferences earlier models missed.
It’s a real step forward. For Display Ads, we used to work in two stages: one system picked suitable collections based on simple similarity scores, and a second ranked them. Now a single transformer model does both. For Sponsored Product Ads, we moved step by step from a basic statistical model to a neural network and then to a transformer. It now also uses near real-time shopping data and timing signals, such as the hour and day of the week. Weighing all signals together, the new models match partners’ ads to interested shoppers even more precisely.
We'll keep developing these models to make every ad more relevant – so shoppers discover content that inspires them, and partners get more from every euro they spend.