Compare
Two sites drawing side by side, replayed from the filmstrips Lighthouse captured while measuring them. The clock is the load that was actually recorded, so a lane that is further along really did get there first.
This page animates two filmstrips against each other, which needs JavaScript. Each site's own filmstrip is on its site page, linked from the leaderboard.
Frames are thumbnails from the median Lighthouse run on the emulated mobile profile, timed from navigation start and resampled here onto one shared grid, since each run samples on its own schedule. The clock is that recorded load. Each strip is only eight samples of its own run, and Lighthouse sizes that run's timeline to the run, so a lane simply stops where its capture stops rather than holding its last frame across an axis built for a longer one.
These frames are barely throttled. The runs use Lighthouse's
default simulated throttling, which does not slow the connection or the
processor: it clears network throttling before the load and never applies
CPU throttling at all. Device emulation is real — a phone-sized
viewport and a mobile user agent, so this is genuinely the
mobile layout — but the page otherwise arrives
as fast as the measuring machine and its connection allow. The one
exception is requestIdleCallback, which Lighthouse shims to
hand back a quarter of its usual budget per callback, so a page that
schedules work that way does spread it thinner here than it would on your
machine.
Slow 4G and a 4× CPU slowdown are applied afterwards instead, by Lighthouse's Lantern model, as an estimate of what that load would have been over a slow connection. That is where TTFB, FCP and LCP come from, and it is why none of them are drawn on this axis: the frames are one load and those numbers describe a different, slower one that never actually ran. Both are correct about different things, and no offset reconciles them: a page can be fully painted in these frames well before the FCP it reports. So the finish line here is the observed one: when a lane's picture stops changing. The rings beside each lane are that site's scores, which are about the site rather than about this clock. This is a picture of two measurements, not a benchmark.
These frames are of a throttled load. The runs applied Lighthouse's Slow 4G profile and a 4× CPU slowdown to the browser itself — not to a model afterwards — so the page arrived over a deliberately slow connection on a deliberately slow processor. What you are watching is much closer to what someone on a phone would sit through than an unthrottled capture is.
Because the throttling was real, TTFB, FCP and LCP for these sites are read from this same trace rather than estimated from a faster one, so for once they describe the load in front of you. They are still not drawn on the axis: the finish line here is the observed one — when a lane's picture stops changing — which comes out of the frames themselves and needs no second source. Numbers measured this way carry more run-to-run variance than the modelled ones, so a close result is closer than it looks. The rings beside each lane are that site's scores, which are about the site rather than about this clock. This is a picture of two measurements, not a benchmark.