Cell-type deconvolution methods aim to infer cell-type heterogeneity and the cell abundances from bulk transcriptomic data. The large number of currently developed methods (>50) and the vast inconsistency in their results in many cases, create an urgent need for guidance on method selection. Although previously suggested benchmarks have paved the way to better understand the performance of deconvolution methods, the proposed tests focus on simulated data and have only been applied to a handful of datasets. Besides, there is pressing interest to achieve the decomposition of database-level transcriptomic data of different tissues, conditions and species, scenarios in which deconvolution methods have not been tested thoroughly. Here, we propose a large-scale, multi-level assessment of 29 available deconvolution methods, leveraging single-cell RNA-sequencing (scRNA-seq) data from different organs and tissues. We extend previous benchmarks, suggesting a comprehensive simulation framework to evaluate deconvolution across a wide range of scenarios and we provide guidelines on the preprocessing of input matrices. We show that single-cell regression-based deconvolution methods such as DWLS are performing well but their performance is highly dependent on the reference selection and the tissue type. Our study also explores the vast impact of various batch effects in deconvolution including sample, study and the technology effect which have been overlooked. We validate deconvolution results on a golden standard bulk Polymorphonuclear Peripheral/Blood Mononuclear cell (PMN/PBMC) dataset with known cell-type proportions. Importantly, we suggest a novel methodology for consensus prediction of cell-type proportions for cases when ground truth is not available. The large-scale assessment of cell-type prediction methods is provided in a modularised pipeline for reproducibility (https://github.com/Functional-Genomics/CATD_snakemake). Last but not least, we envision that CATD (Critical Assessment of Transcriptomic Deconvolution) pipeline can be used for fast, simultaneous deconvolution of hundreds of real bulk samples with the use of different references. We also envision it to be used for speeding up the evaluation of newly published methods in the future.