Insights of YouTube View Check System

Abstract

YouTube is the most popular website for videos nowadays. Its business model is based on advertisement: ads are introduced in the videos and uploaders are paid based on the number of views of their videos. For this reason, it is imporant for YouTube to be able to check that views are real, and discard those that are faked. The main purpose of this paper is to understand the mech- anisms implemented by YouTube to detect and discard faked views. To this aim, we have developed a robot that performs faked views; by building on a thorough study of the patterns and the most common traffic sources of videos, and emulating this patterns, our robot shows a very similar behavior to real users. From the results obtained from our experiments, we observe that: (i) we are able to overcome the 301 limit, which corresponds to the first thorough analysis YouTube performs to detect faked views, (ii) we are capable of pushing the total number of views to an unlimited large number, but at a limited growth rate, (iii) YouTube appears to be insensitive to many behaviors that show that views are faked, except for the IP address used by the client, and (iv) regardless of how sophisticated the robot is, YouTube only counts a fraction of the views realized (around one half). In order to compare the behavior observed by our robot to other behaviours, we have considered (i) an experiment with real people watching the video, and (ii) services offered in the Internet that sell views for YouTube videos. We have observed from the first experiment that even with real views, YouTube only seems to be counting about half of them as real views. Furthermore, the second experiment confirms that even some basic tools are able to increase the number of views (e.g., in some cases we found that the average view duration was 0 but still for those cases YouTube counted a large number of views). The experiments conducted seem to show that YouTube implements a very basic scheme to detect fake views that is easily tricked by simple tools, and that YouTube discards a subsantial number of the views performed even for real views. Such results seem to point to the need for revisting the algorithms implemented by YouTube to detect fake views.