A 1% winsorisation sets everything below the 1st percentile to the 1st percentile value and everything above the 99th to the 99th. The observation count is unchanged, so you have not thrown away sample, but the leverage of the extremes on a mean or regression is capped.
Example: daily returns with one -35% print from a single stock. That one day can swing an estimated beta materially. Winsorised at the 1st and 99th percentiles, it becomes perhaps -8% and the estimate stabilises.
The honest caveat: winsorising is appropriate for estimating a signal, and dangerous for estimating risk. Never winsorise the series you use to size positions or model tail-risk, because you would be capping exactly the events you need to survive.
Related: outlier, normalisation, data-quality-check, fat-tails