Big data coming from social media offer companies the potential to develop knowledge of the behaviour of their customers (Chua, 2011), especially within the TV industry (Proulx & Shepatin, 2012). Indeed, the rise of smartphones and tablets that allow discussing around TV shows on social networks as Twitter (Doughty, Lawson & Rowland, 2011), the so-called Social TV phenomenon (Harboe, 2009), generates a huge quantity of data useful to analyse viewers’ behaviour in order to better design TV contents and marketing activities. Notably, explaining the drivers of social media activity (Hill & Benton, 2012) is one of the major goals for TV companies (Proulx & Shepatin, 2012). This research aims at investigating the drivers of viewers’ interactions on Twitter, by considering different types of viewers. We collected approximately 550,000 tweets generated during a popular TV show, the contents shown on TV, the Twitter elements shown on TV screen and the viewership. Through the hierarchical linear regressions (Frazier et al., 2004), we investigated the minute-by-minute correlations between the viewers’ Twitter activity (dependent variable) and TV show’s contents and Twitter elements (independent variables). This extensive analysis was conducted on subsets of viewers classified by their loyalty to the TV show and their intensity of Twitter activity. Although the analysis of big data can improve the knowledge on viewers’ behaviour, their inherent heterogeneity can cause a lack of clarity in results. Results show that this lack of clarity may be caused by the fact that different types of viewers exhibit very different kinds of behaviour. Therefore, the study reveals some interesting insights on the phenomenon, since viewers are separated based on their loyalty and intensity. To our best knowledge, this research is the first demonstrating that the drivers of online engagement depend on the kind of viewers. In the context of Social TV, big data coming from Twitter can be useful to understand what makes people engaged with a TV show. Our findings demonstrate that ignoring differences between different types of viewers can lead to little knowledge from big data and, consequently, make ineffective business decisions.