Twenty-Two of Twenty-Nine

Across 1,091 dealer sites measured once between 2026-01-17 and 2026-05-12 and again between 2026-07-02 and 2026-08-02, the typical site’s mobile LCP changed by 71 milliseconds, with a 95% confidence interval of minus 1 to plus 152 milliseconds; 568 sites got slower, 514 got faster, and 9 did not change; the share meeting Google’s 2.5 second mobile LCP threshold in a standardized lab test moved from 2.66% to 2.93%, while 22 of the 29 sites meeting it at the first measurement did not meet it at the second, a 76% turnover in the passing set, and 25 different sites crossed in; and 751 of the 1,091 sites measured above 5 seconds at both measurements.

The two halves of that sentence disagree with each other, and the disagreement is the reason I wrote this down.

The share is the number anyone would quote. It moved by a quarter of a point between the two measurements, which reads as a layer sitting still. Underneath it, the set of sites producing that share was mostly replaced. Twenty-nine sites met the threshold at the first measurement. At the second, twenty-two of them did not, and twenty-five sites that had not met it did.

I had been treating a pass rate as a description of which sites are fast. It is not. It is a headcount taken on one day, and this is the first time I have had the same sites measured twice closely enough to see the headcount hold while the heads changed.

That has a direct consequence for my own published work. I publish pass rates. Fewer than 4% meet Google’s 2.5 second mobile LCP threshold in a standardized lab test, and I have put that figure on the front of this site. It is still correct. What this panel adds is that on the one population where I have been able to check it, the share held while most of the sites underneath it were replaced. I cannot tell you the cohort figure behaves the same way, because I have not measured the cohort twice. I can tell you that meeting this threshold once is not evidence a site will meet it again.

I cannot tell you what produced the turnover. Nothing in a PageSpeed run reaches inside a site to find out, and this panel does not separate what the sites did from what the instrument does when it measures the same site twice, which is a comparison the paper reports and does not resolve. It says nothing about any individual site either, because this measurement is noisy per site and precise only across hundreds of them.

Underneath both of those facts the layer keeps its shape. Those 751 sites sat above 5 seconds at both measurements, and the distance between the 8.7s median this cohort measures and a 2.5 second threshold is far wider than anything a repeat audit moves. What a single run cannot pin down is the individual grade at the line. That is why this research is about the layer and never about one site.

What I will do differently is stop reading a single pass as durable, including my own.