This paper analyzes how the audience for internet news varies with the source of the referral, comparing the audience when users navigate directly to news outlets to that when users are referred by social media. On average, the audience referred by social media reads different types of articles, for example, articles with more emotional content and articles that show an individuals perspective. We find that much of the difference in audience can be explained by the fact that different types of users use social media as a news source with different propensities, and further different topics are consumed on social media. We also study the problem from the user perspective showing that a given user is more likely to consume news that aligns with their political preferences (as revealed through their past browsing) through social media. That is, social media seems to exacerbate polarization in information consumption. The paper develops a novel methodology for categorizing news at a large scale, combining text mining and crowd sourcing techniques. With these methods we are able to provide evidence about a variety of qualitative features of news consumption at a large scale, including subtle factors such as whether an article is of a type would typically be shared to impress one’s social media friends.