This presentation has a twofold objective. Substantively, we revisit a longstanding question using computational social science tools: Do people learn about politics through news exposure? To overcome the limitations of traditional self-reported exposure measures, we rely on exogenous (i.e., observed) measures of exposure. Methodologically, we validate several exogenous exposure measures that vary in granularity. Using machine learning techniques, we develop exposure metrics ranging from domain-level visits to specific content interactions. We assess these measures through multiple approaches, including their ability to predict political knowledge on an unexpected news event— the 2022 Russian invasion of Ukraine. Our findings highlight the importance of granularity: only visits and time spent on Ukraine-related articles significantly predict event knowledge, whereas broader measures—such as domain-level visits—show no effect when controlling for self-reported exposure and other key predictors.