Latency first, with the first probe thrown away. A warm-up request pays TCP and TLS setup, and counting it would inflate the idle baseline — which would quietly flatter the bufferbloat grade by making the added latency look smaller. Up to ten probes then follow, with an eight-second cap so a slow link cannot stretch the phase indefinitely — the run detail reports how many actually landed. Ping is the best (minimum) round trip of those, because the minimum is the closest thing to the path’s floor; jitter is the mean absolute difference between consecutive probes, which is the same definition the PulseLAN app uses. Fewer than two probes cannot produce a jitter figure, so in that case it reads “—” rather than a confident 0.

Then throughput, over four concurrent transfers each way by default (one in Single mode). A fixed 1.5-second warm-up is discarded, then bytes observed during the next ten seconds are counted. Waiting for final acknowledgements cannot reduce the finished rate, and callbacks after the deadline cannot add bytes. Both this page and app internet profile 2 retain the existing 4% overhead compensation. Downloads use adaptive chunks; browser uploads use bounded XMLHttpRequest POSTs with progress events, while the native app streams its POST bodies. Four browser transfers can share one HTTP/2 or HTTP/3 connection; they are not necessarily four independent network connections. Upload progress describes the browser network stack, not proof of receipt at the server.

Latency is sampled throughout, every 300 ms. Each phase reports the median of its samples, not the mean, so one scheduler hiccup cannot decide your grade. The added-latency figure is the worse of the two loaded medians minus the idle median.

The ping shown is an HTTP round trip, and it will read higher than ping in a terminal. A probe is a one-byte request, so the time we measure includes Cloudflare running the code that answers it — roughly 20 ms of their own processing on the runs we tested — plus a few milliseconds of browser and JavaScript scheduling on top. That is not an error, it is what the number is: the time for a real request to be answered. The run detail below the results reports Cloudflare’s own view of the connection’s minimum round trip alongside ours, precisely so the gap is visible rather than hidden. Note that an inflated idle baseline makes the added-latency figure smaller, so it errs toward a kinder grade, never a harsher one.

Nothing here is ever filled in with a zero. A phase that fails shows an error and no number; a metric with too few samples shows “—”. Values under 10 ms are shown with decimals, because rounding a fast link to “0 ms” reads as a broken tool rather than a quick connection.

A browser test measures the browser’s path and transport. Browser connection sharing, progress-event batching, CPU load, Wi-Fi conditions and request overhead can change the result in either direction. Safari may also use iCloud Private Relay while the native app’s HTTPS traffic takes a direct route. This page cannot detect Private Relay reliably and does not bypass it. A low browser result alone does not establish a fault in your internet connection.

Keep the test in the foreground. Browsers can throttle timers and network delivery in a background tab. This test stops when the tab becomes hidden, clears the incomplete readings and asks you to restart. This avoids presenting a throttled run as a valid measurement.

It measures your internet path, not your LAN. Everything on this page is about the route from this device to a Cloudflare edge. It says nothing about how fast this device can talk to your NAS, your desktop, or the laptop in the next room — that is a different link entirely, and measuring it is what the PulseLAN app does. If a file copy at home is slow, this page cannot find it.

We do not name your ISP or your location, because we cannot read them. Cloudflare does not expose that information to a page in your browser, and we are not going to guess at it. The edge code shown in the run detail is a Cloudflare data-centre identifier and nothing more — it is not a city. There are no comparisons to other users on this page either: we have no data on other users, so any such figure would be invented.

One run is one moment. Whatever else was using your connection is part of what you just measured. Run it twice before you conclude anything, and three times before you phone anyone about it.