Abstract: This paper reviews various selection methods of real records for time-history analysis of structures in recent years. The selections of criteria and required numbers of records are summarized and classified. Specific earthquake scenarios, such as the magnitude, distance and site classification, are generally considered in the first step. Then, the target spectrum and other intensity measures (IMs) of the ground motions are taken into consideration. Presently, dozens of IMs have been developed, but there is no agreement on the relationship between IMs and structural responses, as well as the advantages of IMs in the selection of records. The method that selects and scales real ground motion records according to target spectra is the most popular. There are many kinds of target spectra, among which the design spectra in codes are the most popular because of the easy implementation in practice. The Conditional Mean Spectrum (CMS) is more accurate. How to introduce CMS into the Chinese code is worth of further studying. The selection of deviation parameters and scaling methods for estimating the mean responses of a structure should be taken in aspect of getting the optimum selection in real records that can match the spectral shape. With the development of the performance-based seismic design and the gradually perfection of Probabilistic Seismic Hazard Analysis (PSHA), estimations of structural responses have been developed from mean values to probability distributions. The selection criteria and required number of records for estimations of the probability distributions of structural responses need to be further studied.