The Kroger Onsite Ad Technology team engaged 84.51° Data Science and Research Labs teams to create Dynamic Positioner, a solution with a proprietary two-phased approach.

In the offline first phase, a model uses a combination of historical insights from website interactions to estimate how likely an ad is to be clicked. To ensure the model can handle uncommon search terms, historical searches are clustered based on similar clickthrough insights for the most frequently clicked positions in the search results. This approach helps to make the model more robust and improve its performance in handling a wider range of search queries.

The online phase incorporates real-time optimization and searches the set of positions of the sponsored ads returned from the auction. It uses the modeled relevancy and position effects, and the clustering from the offline phase, to search for the optimal position of the sponsored ads, all in under 20 milliseconds so that customers experience no latency effect.

FULL ARTICLE