Original measurement, published whichever way it lands. Every number here we took ourselves, with the method written out so you can check it.
Variable fonts get recommended constantly and quantified rarely. We measured both options for the two typefaces on this site, at the exact weights in use.
Finding. For our two typefaces, variable files cut 156,616 bytes across 9 requests down to 80,640 bytes across 2. That is a 48 percent saving and 7 fewer round trips, with more weights available than the static set offered.
September 2026 · Read it
The median mobile page is 2,164 KB. We measured every byte of this one and it came to 94.4 KB across 7 requests, with an interactive scene, two self-hosted typefaces and no third-party requests at all. The surprise was which part costs the most.
Finding. 94.4 KB over the wire across 7 requests, against a median mobile page of 2,164 KB across 66. Typography accounts for 83 percent of it. The interactive scene accounts for 4.7 KB.
August 2026 · Read it
Three rules, and they are the reason this page fills slowly.
Not a summary of somebody else's study. If we did not run it, weigh it or count it, it belongs in the journal as an opinion, not here as a finding.
Every piece says what we measured, how, and on what date. If you repeat it and get a different answer, that is the point, and we will publish the correction.
Two of the results so far contradicted what we expected going in. One of them argues against a decision we had already made. Those are the useful ones.
Studies in progress, listed here before they have results so we cannot quietly drop the ones that go badly.
We have been recording prices at around nineteen Coachella Valley stations several times a day since June. The question is how far apart the cheapest and dearest stations really sit, whether the gap is stable, and whether it is worth a driver crossing town. Original local data nobody else is collecting.
Weight is not speed. The follow-up to our first study is measuring the same pages on a mid-range Android device over a real mobile connection, because that is the test that matters to a customer.
Search is increasingly answered rather than listed. We want to know what share of local business sites carry structured data, allow AI crawlers, and are legible enough to be quoted. A sample of real sites, measured the same way.
Every AI vendor claims hours saved. Almost none show the before. We run enough of our own automated systems to measure the honest version, including the tasks where automation cost more time than it saved.
Every method here is written so somebody else can run it. If your numbers disagree with ours, that is worth more to us than agreement, and we will publish the correction with your name on it if you want it there.