In this work we dig into YouTube's policy to ensure the authenticity of engagement metrics: the platform decreases the views counts of videos when it deems some views to be produced not by humans but rather by automatic systems. Drawing on data that we collected during 18 months from 1300 French YouTube channels, we highlight the massive scale of YouTube views corrections, which affect the majority of the monitored channels and half the videos in our corpus. Given the extent of YouTube corrections, and bearing in mind the dynamics that regulate content diffusion on the platform, we investigate whether manipulating engagement metrics could produce some effect in boosting the popularity of videos. Our analysis shows that the number of illegitimate views removed from videos is not independent from their final popularity. Videos corrected later are on average more popular than those corrected earlier or the uncorrected ones and the median popularity of slowly corrected videos is almost 4 times bigger than the one of uncorrected videos. Our results bring forth the crucial role that time plays in the diffusion of online content: views that are corrected too late can interfere with the platform's trendiness dynamics and help a content reach a broader audience.