To simulate the nonlinearity of isolation layer and isolated structure simultaneously in seismic testing, this paper proposes a two-stage real-time hybrid testing method for isolated structures. In the first stage, the isolation bearing was tested separately, and the neural network was trained by the test data for fitting the dynamic characteristics of the isolation bearing. In the second stage, the trained neural network was used to build the numerical substructure model of isolation layer, which was then combined with the isolated structure to realize real-time hybrid testing (RTHT), so as to complete the dynamic performance test of the overall isolated structures. In numerical simulation, the peak error of the results between the proposed method and overall model simulation was within 3%, while in experimental verification, the peak error of the results between the proposed method and shaking table testing of overall structures was within 6%. Numerical simulation and experimental comparison show that the proposed two-stage real-time hybrid testing has a good accuracy. It avoids the influence of insufficient mass of isolation bearing caused by reduced scale in shaking table testing, and ensures the authenticity of its dynamic performance. Besides, it can solve the errors caused by inaccurate modeling of numerical substructure when using RTHT to test the isolated structures with strong nonlinearity in each part. The two-stage real-time hybrid testing provides a new method for seismic testing of isolated structures.