Neural network-based sequential recommendation models, such as SASRec and GRU4Rec, struggle with highly repetitive recommendations due to a problem known as the SoftMax bottleneck: it is hard to model a multi-modal probability distribution, which is common in repetitive recommendations, using a single-vector sequence representation. As a solution, a recently proposed state-of-the-art approach called CPR uses separate sub-models to compute scores for repeated and non-repeated interactions. However, CPR does not explicitly consider the counts of past repetitions, one of the most important signals for predicting future interactions, and counting repetitions in the models’ inputs is known to be a hard problem for neural networks. Building upon CPR ideas, we propose CountNet, which models repetition frequencies using Laplace and Dirichlet models; two simple yet efficient statistical approaches frequently used to describe probability distributions in highly-repetitive data. As a result, we improve the effectiveness of sequential recommendation models in repeated recommendation scenarios. For instance, on the Gowalla dataset, we improve Recall@10 of SASRec by 23% and by 5.96% when compared to SASRec enhanced by CPR. CountNet can be easily applied on top of many existing sequential recommendation models, and we show that these improvements generalize through extensive experiments with three different backbone models and three real-world sequential datasets with repeated interactions.