Starting around 2014, Prime Video uses a type of machine learning model called an autoencoder to predict which shows and movies each viewer will want to watch next, based on the order and timing of their past viewing. This approach outperformed the previous recommendation method by roughly two to one in internal testing.

The autoencoder is a neural network architecture that compresses viewing history into a compact representation and then reconstructs it to predict future preferences. The team framed recommendation as a prediction problem — anticipating what a viewer will want to watch over the coming one to two weeks — using chronologically ordered viewing data as training input. Earlier attempts failed by recommending safe, broadly popular titles rather than personally relevant ones; the temporal prediction framing solved this. The breakthrough was described internally as a "once-in-a-decade leap." This built on Amazon's earlier collaborative filtering algorithm, published in 2003, which won an IEEE journal "test of time" award in 2017.