Scholars debate whether and how campaigns influence political behavior and electoral outcomes. No consistent theoretical framework, however, defines, measures, and analyzes election-related content from within the media's coverage, particularly in emerging democracies. We apply machine learning techniques on texts from nearly 100,000 news articles during South Africa's 2014 election, and use a theoretically-informed classification of election coverage to demonstrate how the conceptual scope of elections shapes voters' campaign information environment. Our results produce distinct representations of political actors and institutions during elections: a narrow classification provides heuristics cuing race, party, and incumbent performance; a broad definition reflects policy and service concerns parties debated. Topic models and word vectors show that campaign content clusters with parties and their associations with government performance and policies, but candidates vary in how much distinct coverage they obtain on valence issues. We provide methods and evidence to replicably study electoral news coverage across developing countries.