Detection of online groomers has been investigated over the last 15 years. Most of the early research used a forensic approach where a full online conversation was classified as benign or grooming. In the last 8 years there has also been focus on early detection as a warning tool to prevent abuse of children. In this presentation we will go into a bit of the history of grooming/groomer detection from the early dictionary-based features combined with machine learning classifiers to the current use of LLMs. We will show how limited information would still reveal a grooming conversation, but most importantly, we will show that we can use clustering algorithms in combination with ego-graph features to provide sufficient information to find groomers in a huge unlabeled dataset.