Why speed test results differ
Every speed test measures a different path in a different way. Each test uses its own servers, some open several connections at once while ours sends one request at a time, and each turns many samples into one number differently. Wi-Fi, the time of day and other traffic in the house move all of them. Compare tests run back to back, on a cable, and look at the trend, not one number.
The short version
| This site | Speedtest by Ookla | fast.com | Google's test | |
|---|---|---|---|---|
| Servers | Nearest Cloudflare location | Ookla's own server network | Netflix servers | Measurement Lab (M-Lab) servers |
| Connections | One request at a time | Scales to several, across several servers | Set by a "Parallel connections" setting | One (M-Lab's NDT test) |
| Fastest line it can measure | About 1 Gbps | Up to 10 Gbps, per Ookla | Not stated | Up to 700 Mbps, per Google |
| Ping under load | Yes (the bufferbloat grade) | Yes | Yes | Not stated |
Every figure about another test comes from its maker's own pages or M-Lab's, listed under Sources. We name these products only to identify them; this site isn't connected to any of them.
Different servers
A speed test measures the path between your device and one particular server. Change the server and you change the path.
- This site measures to the nearest Cloudflare location, which is usually in or near your metro area.
- Speedtest by Ookla runs its own network of test servers. Ookla describes it as the largest network of test servers, with servers in every country and major population center.
- fast.com runs its downloads and uploads against Netflix servers.
- Google's test is run with M-Lab. M-Lab says its servers sit outside internet providers' last-mile networks, so tests cross the points where providers connect to the rest of the internet. It also says many popular tests are hosted at the edge, often inside providers' networks, which usually gives higher numbers.
A server inside your provider's network skips the busiest connection points. A server outside it measures them. Both answers are honest; they answer different questions.
One connection or several
A single connection takes a moment to speed up, and on a long or busy path it may never fill a fast line. Several connections at once fill it sooner.
- Speedtest by Ookla says its engine raises the number of connections, to several servers, until the line is full.
- fast.com has a "Parallel connections" setting with a minimum and a maximum.
- M-Lab's NDT test, the one Google's search integration uses, runs over a single TCP connection. M-Lab argues that one stream better shows a path's problems, and that a multi-stream result is an "aggregate capacity."
- This site sends one request at a time, in growing sizes.
So on the same line, a multi-connection test often shows a higher download than a single-connection one, especially on fast plans or far-away servers. Neither is wrong.
How many samples become one number
A test collects many samples and has to pick one figure to show.
- Our ping is the median (middle value) of about 20 round trips. Our download and upload are the 90th percentile of the per-request speeds: fast, but not the single best moment.
- Ookla's methodology describes several latency figures, including a minimum latency (the lowest of several pings) and an interquartile mean (the average of the middle half of the pings).
A lowest-ping figure will always look better than a median. A test that reports a minimum and one that reports a median can both be right about the same connection.
Ping is measured in different ways
Browsers can't send ICMP, the message the classic ping command uses, so our ping is an HTTP round trip: a tiny web request and its reply. It usually reads a few milliseconds higher than an ICMP ping to the same place. Other tests time their own kinds of messages. Compare ping between tests only roughly.
Wi-Fi, the device and the browser
- Wi-Fi is the biggest variable. Moving one room, or a neighbor's network getting busy, can change the result from one run to the next. A cable takes Wi-Fi out of the comparison: see Wi-Fi vs. Ethernet.
- The device. An old phone or a laptop on battery saver may not keep up with a fast line.
- Browser or app. A test in a browser tab shares the browser with everything else open. A test app, or your router's built-in test, avoids some of that overhead and can read higher on very fast lines.
The time of day and the rest of the house
Results change from minute to minute. Evenings are busy on many networks. Inside the house, a game update, a cloud backup or a 4K stream uses the same line while you test. Google's help page itself warns that results can change in a short time and can differ from other tests.
The fastest speed each test can measure
We measure up to about 1 Gbps. On a faster line the result says "at least" rather than a number. Google says its test measures up to 700 Mbps. Ookla says its engine measures connections up to 10 Gbps. If you pay for a multi-gigabit plan, use a test built for that speed to check the plan, and use this site for ping, jitter, packet loss and bufferbloat.
How to compare fairly
- Use a cable and close other apps and tabs.
- Pause big downloads, uploads and streams in the house.
- Run the tests back to back, a few times each, and note the time.
- Compare like with like: download with download, ping at rest with ping at rest.
- Look at the trend. If every test shows ping rising under load, or packet loss, that's real, whatever the exact numbers.
Our own methods, thresholds and limits are on how it works.
Sources
All checked November 2026.
- Ookla, Speedtest Methodology (last updated October 15, 2025): server network, connection scaling, 10 Gbps, latency definitions.
- Netflix, FAST.com: its questions and answers and its settings panel (Netflix servers, parallel connections, loaded and unloaded latency).
- Google Search Help, Test your internet speed: run with M-Lab, up to 700 Mbps, results can differ.
- M-Lab, Introducing ndt7: one TCP connection.
- M-Lab, How fast is my Internet? (March 2019): single-stream versus multi-stream, and where test servers sit.
- M-Lab, Introducing Data Transfer Limits to NDT (June 2023): Google Search's NDT integration.