Purpose: The purpose of this paper is to propose and demonstrate how Tourism2vec, an adaptation of a natural language processing technique Word2vec, can serve as a tool to investigate tourism spatio-temporal behavior and quantifying tourism dynamics. Design/methodology/approach: Tourism2vec, the proposed destination-tourist embedding model that learns from tourist spatio-temporal behavior is introduced, assessed and applied. Mobile positioning data from international tourists visiting Tuscany are used to construct travel itineraries, which are subsequently analyzed by applying the proposed algorithm. Locations and tourist types are then clustered according to travel patterns. Findings: Municipalities that are similar in terms of their scores of their neural embeddings tend to have a greater number of attractions than those geographically close. Moreover, clusters of municipalities obtained from the K-means algorithm do not entirely align with the provincial administrative segmentation.