Pinterest built a specialized AI model called PinSage that figures out which Pins are related to each other by analyzing the structure of how billions of users have organized content into boards. If many different users have saved both a recipe for sourdough bread and a specific mixing bowl to their cooking boards, PinSage learns those items are related — even if they look nothing alike visually. PinSage was first described in 2018 and its underlying approach still powers Pinterest's related content recommendations, search, and shopping features today.

PinSage is a graph convolutional network — a type of machine learning model designed to find patterns in graph-structured data (where items are connected to each other, like Pins saved to Boards). It was trained on a graph containing 3 billion nodes (2 billion Pins and 1 billion Boards) and 18 billion edges (the save connections between them). The model generates a numerical embedding for each Pin that captures both its visual and contextual meaning. These embeddings serve as foundational inputs for virtually every other machine learning system at Pinterest — including search retrieval, shopping recommendations, ad targeting, and home feed ranking. Successive systems built on this foundation include PinnerSAGE (for user-level preferences), SearchSAGE (for search queries), and OmniSearchSage (for unified multi-modal search).