This paper proposes a novel methodology to identify the geographic market of local newspapers when information on their diffusion is not available or is not sufficiently granular. We illustrate the methodology using historical data from 154 newly digitized newspapers published in Italy between 1919 and 1922. Combining machine learning-augmented optical character recognition techniques, multi-way fixed-effect regressions, and GIS tools, our approach allows us to estimate markets based on news content. Text-based location of newspaper markets considerably improves over assuming that market boundaries coincide with administrative aggregations. We discuss how our technique strengthens the usage of newspapers as a granular and time-varying source of historical information and offers new avenues for identification strategies.