EEG phase is increasingly used in cognitive neuroscience, brain-computer interfaces (BCIs) and closed-loop stimulation devices. However, it is unknown how accurate EEG phase prediction is across cognitive states. We determined the EEG phase prediction accuracy of parieto-occipital alpha waves across rest and task states in 484 participants over 11 public datasets. We were able to track EEG phase accurately across various cognitive conditions and datasets, especially during periods of high instantaneous alpha power and signal-to-noise ratio (SNR). Although resting states generally have higher accuracies than task states, absolute accuracy differences were small, with most of these differences attributable to EEG power and SNR. These results suggest that experiments and technologies using EEG phase should focus more on minimizing external noise and waiting for periods of high power, rather than inducing a particular cognitive state. Significance Statement EEG phase is a neural signal related to many moment-to-moment behaviors and has consequently been used to inform brain-computer interfaces and closed-loop stimulation devices. However, prior research and demonstrations have forced the user to be in a single cognitive state, such as rest, making it unclear how EEG phase can apply to the varied contexts that real individuals are placed under. The current study showed that EEG phase can be consistently well predicted across different cognitive contexts after accounting for EEG power and signal-to-noise ratio. These findings represent an important next step for both understanding the cognitive and neurobiological correlates of EEG phase and optimizing EEG-based devices to administer more effective interventions.