Language models struggle to search, not due to an architecture problem, but a data one! They rarely see how to search or backtrack. We show how LLMs can be taught to search by representing the process of search in language as a flattened string, a stream of search (SoS)!