Percentiles: the 10th tells you who's suffering.

In Averages lie I learned from the best to lead with the median, and I still stand by that one! The median tells you what a typical person gets. The catch is that the median hides the people having a rough time.

Those folks are down at the bottom of the pile. That's where the 10th percentile comes in.

What the 10th percentile is

Line up 100 speed tests from slowest to fastest. The 10th one from the slow end is your 10th percentile. One test in ten came in at or below that number, and the other nine did better.

That's the whole thing. No fancy math, just a different spot in the same line.

Two towns with the same median

Here's a made-up example. Town A and Town B both have a median download speed of 100 Mbps. Put that on a slide and they look identical.

In Town A, even a bad test is pretty usable. In Town B, one test in ten comes in around 2 Mbps, which is rough territory for a house where somebody's on a video call and somebody else is streaming. Town B has a problem, and the median will never tell you about it.

Signal strength does this too

The same thing happens with signal strength (RSSI or RSRP). Say most readings in an area sit around -90 dBm, which is fine. The bottom 10% might be down near -115 dBm, deep inside buildings or out at the edge of a cell. Those are the people with dropped calls and the spinning wheel, and they're invisible if all you report is the middle.

Before we go further, let's define "bad internet," because a single speed test can fool you. A test is a snapshot. It measures one moment, on one device, from one spot in your house.

Let's run a quick thought experiment. Maria runs a test from her couch on a Tuesday night and gets 4 Mbps. What she can't see is that her son is downloading a big game update, and she's two rooms away from the router. Next door, Sam tests at 6 a.m. right next to his router and gets 900 Mbps. Yet his video calls freeze in the evenings.

So who has bad internet? From those two tests alone, you can't tell. One slow result doesn't prove Maria's connection is bad, and one fast result doesn't prove Sam's is good. To actually know, you need lots of tests over time, and that's the job percentiles do.

Here's how that plays out. If Maria's 10th percentile over a month looks healthy, Tuesday was a one-off. If Sam's 10th percentile keeps coming in low, his problem is real, and that one great test was covering for it.

How I pull it

It's one line in SQL:

select
  percentile_cont(0.5) within group (order by download_mbps) as median_mbps,
  percentile_cont(0.1) within group (order by download_mbps) as p10_mbps
from speed_tests;

And in pandas:

df["download_mbps"].quantile([0.1, 0.5])

(When should someone use SQL or Python? Coming to a theater near you!)

These days the 10th percentile sits right next to the median whenever I share speeds. If the gap between them is big, that gap is usually the story.

The median tells you how most people are doing. The 10th percentile tells you who's suffering. I report both.

Update: I found Town A and Town B

After this went up, I ran the same idea on real data: Ookla's open speed test tiles for every Texas county in Q4 2025. I was looking for two counties with nearly the same median and a very different 10th percentile.

They turned up right next to each other, west of Houston.

The medians are within 20 Mbps of each other. But one test in ten in Waller County comes in around 106 Mbps or slower, versus 310 in Fort Bend. The median says these two counties are the same. The bottom 10% says they really aren't.

A quick note on the numbers: these are percentiles of tests, so busy tiles count more. Count every tile once instead and you get lower numbers, which is the tile-versus-test question I dig into in SQL or Python? Yes. The code and both charts are on GitHub if you want to rerun it.