Prediction is central in human language processing, as the brain continuously predicts upcoming words using prior knowledge and context. Surprisal theory quantifies predictability using word surprisal. While previous studies link neural activity to surprisal during passive listening or reading, we investigate how surprisal is tracked in dynamic face-to-face conversations. Two key challenges arise: estimating surprisal as well as identifying predictions in EEG data in natural conversation. We address the first challenge by adapting a pre-trained large language model to a dataset of spontaneous conversation capturing features like hesitations and repetitions. We then relate the surprisal estimated by the adapted model to EEG data using temporal response functions. Our experimental results show neural tracking of surprisal at different time lags after word onset, supporting the surprisal theory in face-to-face conversation. To the best of our knowledge, we are the first to address the application of surprisal theory in such interactive settings.