75th percentile (p75)

Line up every recorded visit from fastest to slowest and walk three quarters of the way along the line. The value you are standing on is p75: three out of four visits were at least that fast, one out of four was worse. Core Web Vitals are assessed at p75 over a rolling 28 days, split by phone and desktop, which is why your dashboard's cheerful average has nothing to do with whether you pass.

Percentiles beat the mean because latency distributions have a long ugly tail: one person on a tunnel connection can pull an average up by a second while nothing about the site changed. The median (p50) has the opposite problem, hiding the tail completely. Google landed on p75 as the compromise: reachable enough that a well-built site can pass, strict enough that you cannot ignore your slow quarter.

Gotcha: percentiles do not add up. The p75 of TTFB plus the p75 of resource load time is not the p75 of LCP, because different sessions are slow for different reasons. Same trap across segments: fixing your fast desktop pages moves nothing, because the p75 lives inside your worst traffic. And do not over-read a pass: a quarter of visits being slower than your p75 does not mean a quarter of people had a bad time. If your p75 sits comfortably inside the good threshold, plenty of that slower quarter is still inside it too. What the grade does hide is the shape of the tail, so look at p95 before declaring victory.