Captain - ML Listing

In Allegro we want to simplify shopping and selling to be the most loved shopping destination in Europe. With ML Listing we bring 14 mln shoppers and many thousands of merchants together to maximize their mutual benefits - making it easier to find the best products and increasing merchants' sales. As part of this project, our aim was to improve the relevance of ads and search results for users by pushing irrelevant products further down in the search results. In the spirit of compressed commerce, this will make it easier for users to find what they want and for merchants to achieve better advertising results.

Allegro creates a platform that connects customers who want to buy specific products with merchants who want to sell them. To do this, we want to provide our users with a toolset that is effective and beneficial for all parties involved. Listings are an important tool that allows consumers to browse search results and merchants to advertise and reach a wider audience. Because these two goals are not perfectly aligned, mixing ads and organic search results can lead to listings of varying relevance. Our dynamic ad positioning aims to reorder listings to improve the contextual relevance of ads by positioning them more seamlessly among other relevant results. Better sequencing improves the browsing experience for consumers and increases the chances of merchants' promoted offers reaching interested parties. This sets us apart from platforms where ad positioning is static and arbitrary. Compared to competitors who also implement the dynamic approach, our solution has significant differences at the technological level. These differences range from the use of a different machine learning model architecture, to the development of a custom result scoring function, to a different dynamic adaptation algorithm. All this has led to improved results compared to benchmarks found in the literature.

To create ML ads in search results, we first had to integrate the search and ad engines. This required significant architectural and logical changes to Allegro's systems. A unified search engine for both ads and organic results was needed to create a single system for ranking and scoring offers. As a result of these changes, we increased the Return on Advertising Spend (ROAS) for merchants by 27% by providing better-matched sponsored offers to search queries.

Dynamic ad placement has recently been enabled globally on search listings. This means that Allegro users (customers and merchants) do not have to take any action to use the product. This makes it difficult to gather individual testimonials or user feedback. We can measure aggregate KPIs that show the value delivered to both customer and merchant objectives. From a customer perspective, we use the NDCG metric to measure the quality of the order of listed offers; the higher it is, the better the search experience. Our tests showed that the use of dynamic ordering resulted in an increase of around 1%, which is a significant gain for this type of metric. In the case of merchants, the benefits are even more visible, as we have seen a 33% increase in the number of views, an 18% increase in the click-through rate for desktop and around 27% for mobile web traffic, and a 22% increase in the raw number of clicks (and 28% for mobile web). All in all, the statistics show that both parties benefit from using our solution. This is all the more pleasing as, according to the Minds&Roses 2023 report, Allegro - not Google - is now the most popular search engine for product searches in Poland.

www.allegro.pl