Questions
How is download speed measured?
Your browser reads data from the server over several parallel streams (twelve over QUIC, six TCP connections over TCP) for a fixed time and counts every byte as it arrives. The first moments of a transfer are slower while the connection finds the path's rate, so the result is the rate after that ramp: from the first quarter-second window that reaches 90% of the median of the windows after it (never later than half the transfer). The average including the ramp and the best single second are shown beside it.
How is upload speed measured?
By the server. Every upload stream or connection carries the same random test id, and the server counts what each one delivers, as it arrives, on one meter in 50 ms buckets. Over QUIC the browser uploads on several WebTransport sessions at once (a browser keeps only about 128 KB in flight per session, so one session cannot fill a long path), as it uses six connections over TCP. The page reads that meter during the test for the live figure and afterwards for the result, so the figure is what crossed the network, not what your browser handed to its own send buffer. The same ramp rule as the download applies.
What is latency under load, and what is RPM?
While each transfer runs, the page keeps sending small QUIC datagrams and timing their echoes. A connection whose router queues too much data gets slower to respond when it is busy (bufferbloat): the increase over the idle round trip is graded from A+ (under 5 ms) to F (400 ms or more). RPM, round trips per minute, is 60,000 divided by the median round trip under load, as the IETF responsiveness draft defines it: higher is better.
How are latency and jitter measured over QUIC and TCP?
Over QUIC, as datagram round trips: the server echoes each one at once, and the page times it in a background worker, away from page rendering. Over TCP, as HTTP request round trips on a connection kept open, read from the browser's Resource Timing (request start to first response byte), so the browser's own request handling is not counted. Jitter is the mean difference between consecutive round trips, the same definition for both.
How is packet loss measured?
The page sends 400 QUIC datagrams of about 1,000 bytes, one every 5 ms, and counts the echoes that come back within four times the 95th-percentile round trip (at least one second). Loss is round trip, either direction. The result carries a 95% confidence interval: 0 lost of 400 means under about 1%, not exactly 0%. TCP has no loss figure: it retransmits whatever is lost.
Why compare QUIC and TCP?
They reach the same server over the same path, so a difference is the network treating them differently. QUIC runs over UDP, which some networks rate-limit or block; TCP rarely is. The servers run BBR congestion control on both (BBRv3 for QUIC, the kernel's BBR for TCP), so the comparison is not two different controllers. Each transfer runs alone, one after another.
Which server should I pick?
The page times every location from your browser when it loads and starts on the one that answers fastest; you can choose any other. The servers are in Kansas City, Kansas City 2 and St. Louis.
What does Compare show?
Where your result falls among other tests in the last 30 days that were measured the same way: the share of them that did worse, per metric. It compares against tests of the same server when there are at least ten, and says when it had to use every server instead. Results from before this way of measuring are not compared.