Why your PageSpeed score changed when your site didn't.

What the number is

PageSpeed Insights loads your page once, on a simulated slow phone over a simulated slow connection, measures a handful of timings — how long until the largest thing on screen appears, how long the page is frozen while scripts run, how much the layout jumps around — and turns those timings into a score from 0 to 100.

The turning-into-a-score step is where most of the confusion lives. The score is not a percentage of anything. Each timing is placed on a curve that is steep in the middle and flat at the ends, the curves are weighted, and the weighted results are combined. In current versions the two heaviest weights are the time the page is blocked by scripts and the time to the largest visible element. So two pages with very different problems can land on the same score, and a small change in one timing can move the score a lot or barely at all, depending on where on the curve it sits.

Why it moves when nothing changed

Because it is one load, on one simulated device, from one test location, at one moment. Anything that varies between moments varies the score.

• The network between the test machine and your host. Not your connection — Google's, to your server, at that second.
• Your server's response time at that second. Shared hosting in particular has good and bad minutes.
• Third-party scripts. Chat widgets, analytics, booking embeds, ad tags. Each one is a request to someone else's server, and their speed is not yours to control. A page with six third-party scripts has six extra sources of variance.
• The test's own throttling. The simulation of a slow phone is itself an approximation, and it does not produce identical results twice.

The practical result is a band, not a point. A page that is "a 64" is really a page that scores somewhere between the high 50s and the low 70s depending on the run. That band is the honest measurement.

Lab versus field

The score described above is a lab measurement: a synthetic load under controlled conditions, useful for diagnosing what is slow and why. PageSpeed Insights also shows, when it has enough data, a field measurement: what real Chrome users experienced on your page over the last 28 days, aggregated. The field numbers move slowly, because they are an average over thousands of loads, and they reflect your actual visitors on their actual phones.

These two disagree constantly, and both are right. The lab run tells you what to fix. The field data tells you whether your visitors are actually suffering. And it is the field data — not the lab score — that Google's search ranking looks at when it assesses page experience. A lab score of 55 with passing field metrics is a page that is fine for the people who use it. A lab score of 85 with failing field metrics is a page that tests well and performs badly for real owners on real phones, and the second is the one that costs you.

The scale is not linear

Because of the curves, the distance between scores is not what it looks like. Getting a page from 30 to 60 usually means fixing one or two large, obvious things — an uncompressed hero image, a script that blocks everything. Getting from 60 to 90 means fixing many small things. Getting from 90 to 100 means fixing things a visitor could not perceive. Across the veterinary practice sites I measured in May 2026, mobile scores ran from 24 to 71 out of 100. The practice at 24 and the practice at 71 do not have problems of the same size; the first has a page that takes twenty-five seconds to show its largest element, and the second has a page that is merely slower than it should be. The score compresses that difference into 47 points. The seconds do not.

Which is the practical rule: read the seconds, not the score. Largest Contentful Paint in seconds is a real quantity a pet owner experiences. Google's threshold for it is 2.5 seconds. A page at 12.5 seconds — one emergency site I measured — is failing in a way a score cannot soften and a good run cannot hide.

What the same page looks like across five runs

I ran the test on this site's homepage five times in a row on 15 September 2026, changing nothing between runs.

Across the five runs the performance score ran from a minimum of 98 to a maximum of 99, median 98; Largest Contentful Paint ran from 0.99 to 1.17 seconds, median 1.05. On a page this fast the score barely moves, because the scoring curve is flat at the top; the load time still varied by nearly a fifth between runs. The same proportional variance on a twelve-second page is a spread measured in seconds.

How to read it without being fooled

• Run it three to five times and take the median. Never act on one run. Never let anyone else act on one run either.
• Test on mobile. That is where your visitors are and where the scores are lower.
• Compare seconds, not scores, and compare against Google's published thresholds, not against a competitor's number from a different day.
• Check the field data if it is shown. It is the number that matters for ranking and it is the one a single bad run cannot fake.
• Be suspicious of a screenshot of one run. Especially one in a sales email. The next section of this site's writing covers what those audits usually are.

Why an ads person is writing about this

Because the same timings that produce the score also feed Google's scoring of an ad's landing page, and that score sets what an advertiser pays per click. The page that loads in 12.5 seconds is not only ranking badly; every click it receives from an ad costs more than it should. Speed is one number that moves two channels, and it is the first thing I measure on any practice site, always more than once.