As two-position is the most widely used scheme of multi-position alignment, azimuth error analysis regarding the whole alignment procedure as an entity is explicitly discussed. A formula to calculate an equivalent north accelerometer bias drift rate is developed. The effecting extent of each inertial measurement error is also theoretically deduced and validated through simulation. It is pointed out that the main error sources causing heading-sensitive azimuth error are accelerometer triad non-orthogonality and lever arm error. In a Kalman filter alignment algorithm, affected by the equivalent accelerometer bias change rate, extra azimuth error emerges from the mistake estimation of fiber-optic gyroscope drift. A three-sequence scheme and a reciprocating slow-rotation scheme are proposed to achieve the most inertial measurement error self-compensation. Theoretical error comparison and a turntable four-orientation alignment test show the superiority of the reciprocating slow-rotation scheme over the other two schemes. The heading-sensitive azimuth alignment error is reduced from 0.2268° to better than 0.0015° through scheme modification.