Search Engine Journal

How Recommender Systems Like Google Discover May Work via @Sejournal, @Martinibuster

Google Discover operates as a large‑scale recommender system built on the Two‑Tower architecture originally devised for YouTube. The model creates separate embeddings for users—based on watch history, search tokens, location and demographics—and for content items, matching them via similarity scores to retrieve personalized feeds. To address the freshness challenge, the system treats newly uploaded content as zero‑age during inference, reducing bias toward older videos. Research shows this deep neural approach outperforms traditional matrix‑factorization and tree‑based methods, driving higher watch time despite noisy click signals.