What Is p75? Why Percentiles Beat Averages
The 75th percentile explained: why Core Web Vitals use p75, what averages hide, and how to read percentile metrics correctly.
p75 — the 75th percentile — is the value that 75% of measurements fall under. When Google says your LCP is 2.8 s at p75, it means three-quarters of real visits painted faster and one quarter slower. Core Web Vitals are judged exclusively at p75, and the choice of that number over the average is the most consequential statistics decision in web performance.
Why averages fail for performance
Load times are not bell-curved; they are right-skewed with a long, ugly tail — most visits fast, a meaningful minority terrible (old phones, hotel Wi-Fi, cold caches). The average splits the difference into a number no actual user experienced: a site where 80% load in 1 s and 20% in 8 s 'averages' 2.4 s — sounds fine, while a fifth of your visitors suffer. The conversion damage lives precisely in that hidden fifth.
Why 75 specifically
It is a deliberate compromise: p50 (median) ignores the struggling half entirely; p95+ chases unfixable outliers (someone is always on a train in a tunnel). p75 says: make it good for most people including the moderately unlucky, without being held hostage by the pathological tail.
Reading percentiles in practice
- Track p75 as the primary, glance at p95 for the tail. Clycyo's load-time reporting exposes both (avg and p75 appear on the public demo) — a widening avg/p75 gap means your tail is growing even while the typical visit holds.
- Segment before percentile. A global p75 blends desktop fiber with mobile cellular; the actionable numbers are p75-per-page-per-device-class (the monitoring routine).
- Sample size humility: p75 of 40 visits is noise wearing a decimal point. Wait for volume before declaring regressions on low-traffic pages.
The habit to build: any time someone quotes an average for a skewed metric — load time, INP, session length, revenue per user — ask for the percentiles. The average is where the truth goes to hide.