Rather than sampling blindly, it fits a surrogate model, usually a Gaussian process, to the trials so far and balances exploring uncertain regions against exploiting promising ones. On expensive objectives such as a slow event-driven backtest it can reach a good region in 50 trials where random search needs 500.
That efficiency is a double-edged tool in trading. Reaching the maximum of a noisy surface faster mostly means finding noise faster. It is most defensible for tuning machine-learning hyperparameters inside a properly purged validation scheme, and least defensible for picking a trading rule's thresholds.
If you use it, still report the trial count and still examine the neighbourhood of the chosen point.
Related: random-search, grid-search, overfitting