Large groundwater level (GWL) data sets are often patchy with hydrographs containing continuous gaps and irregular measurement frequencies. However, most statistical time series analyses require regular observations, thus hydrographs with larger gaps are routinely excluded from further analysis despite the loss of coverage and representativity of an initially large data set. Missing values can be filled in with different imputation methods, yet the challenge is to assess the imputation performance of automated methods. Assessment of such methods tends to be carried out on randomly introduced missing values. However, large GWL data sets are commonly dominated by more complex patterns of missing values with longer contiguous gaps. This study presents a new artificial gap introduction approach (TGP- typical gap patterns) that improves our understanding of automated imputation performance by mimicking typical gap patterns found in regional scale groundwater hydrographs. Imputation performance of machine learning algorithm missForest and imputePCA is then compared with commonly applied linear interpolation to prepare a gapless daily GWL data set for the Baltic states (Estonia, Latvia, Lithuania). We observed that imputation performance varies among different gap patterns, and performance for all imputation algorithms declined when infilling previously unseen extremes and hydrographs influenced by groundwater abstraction. Further, missForest algorithm substantially outperformed other methods when infilling contiguous gaps (up to 2.5 years), while linear interpolation performs similarly for short random gaps. The value of the developed gap introduction approach lies in aiding the selection of imputation methods, a task not limited to groundwater level time series alone. The results further provide insights into region-specific data peculiarities that can assist groundwater analysis and modelling.