YouTube videos of dog bites present an unexplored opportunity to observe dog bites directly. We recorded the context of bites, bite severity, victim and dog characteristics for 143 videos and for 56 videos we coded human and dog behaviour before the bite. Perceived bite severity was derived from visual aspects of the bite. Associations between bite severity and victim, dog and context characteristics were analysed using a Bayesian hierarchical regression model. Human and dog behaviour before the bite were summarised with descriptive statistics. No significant differences in bite severity were observed between contexts. Only age of the victim was predictive of bite severity: adults were bitten more severely than infants and infants more severely than children. Non-neutral codes describing dog body posture and some displacement and appeasement behaviours increased approximately 20 seconds before the bite and humans made more tactile contacts with dogs 21 seconds before the bite. This analysis can help to improve understanding of context in which bites occur and improve bite prevention by highlighting observable human and dog behaviours occurring before the bite.