It uses one number per day and ignores the range the market travelled between closes. A day that opened flat, ran 3% up, and closed unchanged registers as zero movement, which is obviously wrong as a description of the risk taken.
Its virtues are that it needs only a close series, it handles gaps correctly by construction, and everyone computes it the same way, so comparisons are meaningful. It is the default for good reason.
Its cost is efficiency: to reach a given precision it needs several times more data than range-based estimators like parkinson-volatility. On a 20-day window the noise in the estimate is substantial, which is one reason volatility signals look unstable.
Related: realised-volatility, parkinson-volatility, garman-klass-volatility, yang-zhang-volatility