The positioning accuracy of multi-joint robots is much lower than the repetition accuracy. Currently， kinematic calibration and spatial error compensation are the primary techniques adopted to improve robot positioning accuracy. In spatial error compensation， inverse distance weighting is typically used to predict the error of a positioning point. However， as a single weight， the inverse distance cannot accurately represent the influence of each grid point because the weight is too average， resulting in a poor compensation result. To improve the positioning accuracy of robots， a positioning prediction and compensation method， including distance and orientation， is proposed based on error similarity. First， a positioning error model is derived from the kinematic model using differential kinematic theory， and then， the relationship between joint angles and the positioning error is analyzed. The robot workspace is divided into a grid of several cubes， and the effect of the direction of the positioning point relative to the reference point on error similarity is investigated. The error transfer function is then derived using the cosine of the included angle and the distance as the error transfer factor. Second， a method for anisotropic similarity modeling and error compensation is proposed based on an error transfer function that uses the error of reference points to calculate the error of the positioning point. The proposed method is verified through experiments and compared with the traditional inverse distance weighting method. After compensation， the maximum and average positioning errors of the robot decrease significantly along all directions， as demonstrated by the experimental results. The maximum and average errors along x-， y-， and z-directions decrease from 1.03 mm and 0.30 mm to 0.11 mm and 0.04 mm， respectively. Furthermore， the proposed method provides greater compensation accuracy than the traditional inverse distance weighting method.