We provide a method to track active prevalence of COVID-19 in real time, correcting for time-varying sample selection in symptom-based testing data and incomplete tracking of recovered cases and fatalities. Our method only requires publicly available data on positive testing rates in combination with one parameter, which we estimate based on a representative randomized field study of nearly 10,000 individuals in Utah. The method correctly predicts prevalence in two state-wide, representative randomized testing studies. Applying our method to all 50 states we show that true prevalence is 2–3 times higher than publicly reported.