Tools
Check what you are getting right now.
The test downloads a payload of known size from this server and times it. Then it puts your number next to the measured median for the country you are in, so the figure means something.
The test downloads about 3 MB, or about 13 MB if your connection is fast enough that the larger sample runs. Nothing is uploaded and nothing is stored.
This is one download from one server at one instant, so treat it as approximate. A single stream is limited by the round trip time to the server as well as by your connection. Two runs a minute apart commonly differ by a third. Upload and latency are not measured.
T-Mobile, AT&T
What this test actually does
It runs a short warm up download first, then the real sample. The warm up exists because the first moments of any connection are spent on TLS setup and TCP slow start, and including them would report a number lower than what you actually have.
The sample is 3 MB. If that finishes in under 1.5 seconds, your connection is fast enough that 3 MB was too small to measure properly, so the test repeats with 10 MB and uses that result instead.
Throughput is calculated as bytes received times eight, divided by seconds elapsed. Nothing is smoothed, weighted or adjusted. The payload is random bytes, because a compressible payload would measure the compressor rather than the link.
The response is served with no store caching headers and a unique query string, so a repeat test cannot quietly measure your own disk.
Why a browser speed test is approximate
This is one download, from one server, over one connection, at one instant. A proper measurement uses several parallel streams to several endpoints and reports a distribution rather than a point.
That matters because a single stream is limited by the round trip time between you and the server, not only by the capacity of your connection. If the server is far away, the number will read low even on a good link.
The result also depends on what else is happening. A phone syncing photos in the background, a laptop on the same hotspot pulling an update, and a busy cell at the same moment will all pull the figure down.
So read it as a rough check that the connection is working and roughly where it should be, rather than as a verdict. Two runs a minute apart on the same connection commonly differ by a third.
Upload and latency are not measured here. Latency is what makes a video call feel bad, and it is often the real problem when a connection with a good download figure still feels broken.

What number do you actually need
Below 1 Mbps, maps will load slowly and messaging still works. Streaming and video calls will not. This is roughly what a throttled unlimited plan gives you after the daily cap.
At 2 to 5 Mbps, maps, messaging, email and standard definition video all work comfortably. Video calls hold at this range but drop resolution.
At 5 to 25 Mbps, everything a traveller does on a phone works, including high definition streaming and stable video calls. This is the range that matters.
Above 25 Mbps, almost nothing you do on a phone gets better. A 190 Mbps connection and an 80 Mbps connection feel identical in the hand. The difference shows up when you tether a laptop and pull a large file.
That is why the coverage tables print the tenth percentile alongside the median. The bad tenth is what decides whether a country feels fine or feels broken, and the headline figure hides it.
Related pages
Questions people ask
Does running this test use my data?
Why is my result lower than the median for this country?
Why does it sometimes fail?
Does it measure upload or latency?
Is 1 Mbps enough for anything?
Where do the comparison figures come from?
If the number is bad, the plan may not be the problem
Check the measured median for where you are, then decide. Coverage pages carry the carriers in priority order and the local traps.