p75 hides a bimodal distribution — check the histogram

p75 hides a bimodal distribution — check the histogram
A single p75 LCP of 2.4s can mean two different sites.
— Site A: tight distribution, most users at 2.2-2.6s — one fix lifts everyone
— Site B: 70% at 1.4s, 25% at 4.5s — a bimodal split (cached vs uncached, or fast vs slow geo)
— Same p75, completely different fix: Site B needs the slow cohort isolated, not an average speedup
— CrUX gives you the 'good/needs-improvement/poor' buckets — read the bucket %s, not just p75
Takeaway: a 25% 'poor' bucket at the same p75 is the real target.