Pundits and scholars often claim that congressional primary elections favor extremist candidates, but the mechanisms by which primary voters might learn about candidate platforms are not well understood. In this paper, we collect a new dataset of roughly 16,000 local newspaper articles matched to candidates in contested U.S. House primary races from 1998 to 2012. Using supervised machine learning, we classify these articles into political topics. On average, we find little coverage of candidate platforms. However, we also find that the advantage of extremist candidates in House primaries—measured using the campaign contributions they receive—is concentrated in elections with low levels of newspaper coverage. Where newspaper coverage is higher, there is more coverage of candidate platforms, and extremist candidates do worse. The results suggest that the advantage of more extreme candidates in contested House primaries may be the result of information failures and not just the preferences of primary electorates, and that extremist candidates may do increasingly well as local newspaper coverage continues to decline.