Mean, Median, Min and Max in Latency Tests Explained
The mean (average) adds all readings and divides by their number; the median is the middle reading; the minimum is the best and the maximum the worst. Latency data has occasional large spikes, so the mean can be pulled up by a single bad reading while the median and minimum stay stable. GlobeLatency shows the average of successful passes together with the best and worst pass, so you can see both the typical value and the spread.
On this page
- Why one latency reading is not enough
- What is the mean (average) latency?
- What is the median latency?
- What are minimum and maximum latency?
- Comparison: which statistic answers which question?
- When the mean misleads, and when the median does
- How GlobeLatency summarises readings
- Worked examples with GlobeLatency-style numbers
- How many readings should you take?
- Other measures of spread
- Common mistakes and myths
- What to do next
Quick answer: when you measure latency several times you get several numbers, and you must summarise them. The mean (average) is the total divided by the count. The median is the middle value when the readings are sorted. The minimum is the lowest (best) reading and the maximum the highest (worst). Latency often has a few large spikes, so the mean can look worse than the typical experience, the median shows the typical case, the minimum shows what the route can do in ideal conditions and the maximum shows how bad it gets. Looking at all four together is far more informative than any single number.
Why one latency reading is not enough
If you measure the latency to the same server five times, you will almost never get five identical numbers. Each request travels through different queues, may land on a slightly different path, may or may not need a fresh connection, and may collide with other traffic on your home network. A single reading therefore tells you very little: it might be a lucky best case or an unlucky spike.
The solution is to take several readings and summarise them with statistics. Which statistic you pick changes the story you tell, so it is worth understanding what each one measures.
What is the mean (average) latency?
The mean is the sum of all readings divided by the number of readings. It is what most people mean by "average". Every reading contributes equally, so it uses all the information, and that is both its strength and its weakness: an extreme value pulls it strongly in its direction.
Example: five readings of 48, 50, 51, 52 and 400 ms. The sum is 601 ms; divided by 5, the mean is 120.2 ms. Four of the five readings were about 50 ms, but the average says 120 ms, a number that none of the readings is close to.
What is the median latency?
The median is the middle reading after sorting from lowest to highest. Half the readings are below it and half above. With an odd number of readings it is the exact middle one; with an even number it is usually taken as the mean of the two middle ones (some tools use the lower middle one instead, as in the nearest-rank method on the percentiles page).
Example: the same five readings sorted are 48, 50, 51, 52, 400. The median is 51 ms. It ignores how extreme the last reading was, so it describes the typical request well.
The median is the 50th percentile, usually written p50.
What are minimum and maximum latency?
The minimum (min, best) is the single lowest reading. The maximum (max, worst) is the single highest. In the example above, min is 48 ms and max is 400 ms.
- The minimum estimates the floor of the path: distance and fixed processing, with little or no queuing. It is the most stable of the four across repeated tests, and it is useful to compare routes or locations because it removes most of the noise.
- The maximum shows the worst moment seen. It is the least stable statistic, since it depends on a single event, but it matters for real-time uses: a call or game does not care that your average is good if there is a 400 ms stall every minute.
- The range (max minus min) is a crude measure of spread. In the example it is 352 ms. A similar idea is jitter, the variation between consecutive packets.
Comparison: which statistic answers which question?
| Statistic | Answers | Affected by spikes? | Best used for |
|---|---|---|---|
| Minimum (best) | What can this path do at its best? | No (reversed: it ignores them) | Comparing routes and locations, spotting physical limits |
| Median (p50) | What is a typical reading? | Barely | Describing everyday experience |
| Mean (average) | What is the overall level including bad moments? | Strongly | Quick summary, trend tracking, capacity planning |
| High percentile (p95, p99) | How bad is it for the unlucky requests? | Yes, by design | Real-time quality, service-level goals |
| Maximum (worst) | What was the single worst moment? | Entirely | Detecting instability and stalls |
When the mean misleads, and when the median does
Latency distributions are not symmetric. There is a hard floor (physics), a cluster of typical readings just above it and a long tail of occasional slow readings that can be many times larger. This shape is why the mean usually sits above the median: the tail pulls it up.
The mean misleads when there are outliers
Take ten readings: 20, 21, 21, 22, 22, 23, 23, 24, 25 and 900 ms. The mean is (20+21+21+22+22+23+23+24+25+900) ÷ 10 = 1,101 ÷ 10 = 110.1 ms. The median is the average of the 5th and 6th sorted values: (22+23) ÷ 2 = 22.5 ms. The median describes nine of the ten requests; the mean describes none of them. If the spike was a one-off, the median is the better summary.
The median misleads when spikes are frequent
Suppose a third of the readings are slow. Ten readings: 20, 21, 22, 23, 24, 25, 300, 320, 340, 360. The median is (24+25) ÷ 2 = 24.5 ms, which sounds great, while the mean is 145.5 ms and a third of requests really are terrible. Here the median hides a real problem. Always look at the maximum or a high percentile next to the median.
Rule of thumb
- Mean close to median and max not far from min: stable connection.
- Mean well above median: occasional spikes are pulling the average up.
- Median itself high, min low: the path can be fast but is often congested or contended.
- Min high: the route is long or indirect, or a slow access link adds a constant delay; look at distance and routing.
How GlobeLatency summarises readings
GlobeLatency uses real HTTPS requests from your browser and measures a destination several times (passes). In a Test you choose 1 to 20 passes per destination (default 3). For each destination the result shows:
- Average: the mean of the successful passes. Passes that time out or fail are not counted in the average, so a destination with 3 passes and 1 timeout is averaged over the 2 that answered.
- Best: the lowest successful pass (the minimum).
- Worst: the highest successful pass (the maximum).
- 95th and 99th percentile: nearest-rank percentiles over that destination's passes. With few passes they simply equal the worst pass; see percentiles p50, p95 and p99.
The summary cards also show Your average, Your best pass and Your worst pass across the selected destinations, how many destinations were reached, and how many were flagged Unstable. A destination is flagged Unstable when its worst pass is more than 3 times its average, which is a built-in way of saying "the maximum is far from the mean". It is flagged Slow when the average exceeds 2,000 ms and Down when no pass succeeded.
In the live monitor and map, each destination keeps its last 60 readings in your browser and the card shows the average, best and worst of those readings plus the percentage that got no reply. For how the tool takes its measurements, see how GlobeLatency measures latency.
Worked examples with GlobeLatency-style numbers
Example 1: three passes, one slow
Passes to one destination: 62, 58 and 190 ms. Average = (62+58+190) ÷ 3 = 310 ÷ 3 = 103.3 ms. Best = 58 ms. Worst = 190 ms. Is it Unstable? The threshold is 3 × 103.3 = 310 ms and the worst is 190 ms, so no. Interpretation: the route's floor is about 58 ms, and the first pass or an unlucky one was three times slower. The first request to a host often includes DNS and TLS setup, so a higher first pass is common. The average of 103 ms overstates what you would see on later requests.
Example 2: a failed pass
Five passes: 45, 47, no reply, 46, 400. Only four succeeded: 45, 47, 46, 400. Average = 538 ÷ 4 = 134.5 ms; best 45; worst 400. Unstable test: 3 × 134.5 = 403.5 ms; worst 400 is below it, so the flag does not trigger even though the destination clearly had two bad moments (a timeout and a spike). This shows why you should read the best, worst and reach count together, not just the flag.
Example 3: a rolling window
A monitored destination has an average of 80 ms over its last 60 readings, a best of 62 ms and a worst of 840 ms. The best and the average agree, which means the path normally performs near its floor. The 840 ms is a single spike in 60 readings. Whether it matters depends on the use: for browsing, ignore it; for a call, check whether it repeats at particular times.
Example 4: reading a tile against the world
If your average to a destination is 120 ms and the world average for that destination is 100 ms, the ratio is 1.2, which GlobeLatency labels as slightly slower (yellow). Because the comparison uses averages, a few spikes in your passes can move you from "good" to "slightly slower". Look at the best pass: if it is close to or below the world average, the path is fine and the spikes are the issue. See reading your results for how colours are assigned.
How many readings should you take?
| Readings | What you can trust |
|---|---|
| 1 | Almost nothing. It is one sample, possibly an outlier. |
| 3 to 5 | A rough average, best and worst. One spike can dominate the mean. |
| 10 to 20 | A fairly stable mean and median, a meaningful minimum and a first useful look at the 95th percentile. |
| Dozens, over minutes or hours | A good picture of typical, worst and time-of-day patterns. |
Spread out your readings in time as well as count. Twenty readings in 20 seconds miss a congestion event that happens every evening.
Other measures of spread
The standard deviation summarises how far readings typically stray from the mean; it is larger for noisy connections. The range is max minus min. Jitter, in the sense used for voice and video, is the variation in delay between consecutive packets (see jitter explained). Percentiles describe the tail directly. For most readers, mean, median, min, max and a high percentile are enough.
Common mistakes and myths
- Quoting only the average. It hides both the best case and the bad moments.
- Treating the best reading as what you will experience. The minimum is the floor, not the norm.
- Treating the worst reading as proof of a broken connection. One spike is normal; repeated spikes are a pattern.
- Comparing averages from different numbers of readings. A 3-pass and a 20-pass test are not equally reliable.
- Including timeouts as zeros or ignoring them. Failed readings are not 0 ms. Track the failure rate separately (see packet loss explained).
- Comparing one tool's average with another's. Different measurement layers give different baselines; see ping vs HTTPS latency.
- Forgetting the browser and Wi-Fi. Overhead, extensions, VPNs and background tabs can add noise; see accuracy and limitations.
What to do next
Run a latency test with 10 or more passes on a few destinations and compare the average, best and worst for each. If the worst is far above the average, read jitter explained, then go deeper into the tail with percentiles p50, p95 and p99. To interpret the colours and verdicts you see, continue with reading your results.
Frequently asked questions
Should I use mean or median for latency?
Use both when you can, and look at the minimum and maximum too. The mean reflects everything, including spikes; the median shows the typical reading and ignores outliers. If they are close together the data is steady. If the mean is much higher than the median, a few slow readings are dragging it up and the connection is inconsistent.
What does the minimum latency tell me?
The minimum, or best reading, approximates the floor of the path: distance plus unavoidable processing with no queuing or contention. It is a good indicator of the physical route quality. It is not a good indicator of everyday experience, because you rarely get the best reading every time.
Why is my maximum latency so much higher than my average?
Occasional spikes are normal on real networks: a Wi-Fi retry, a busy router, a background download, a new connection that needs DNS and TLS setup, or a server briefly under load. One spike in a handful of readings can be several times the average. It only matters if spikes are frequent or large for your use.
What is a normal difference between best and worst?
On a quiet wired connection to a nearby server, the best and worst readings are often within a few milliseconds to a few tens of milliseconds of each other. Wi-Fi, mobile data and long routes have a wider spread. As a rough guide, a worst value over three times the average is unusual and suggests instability.
Does GlobeLatency show the median?
GlobeLatency's result tile shows the average of the successful passes, the best (lowest) and worst (highest) pass, and the 95th and 99th percentile. It does not label a median. You can approximate the typical value by looking at the best pass, the average and the 95th percentile together.
How many readings do I need for a reliable average?
More is better, but a handful is enough to spot big problems. With 3 readings one outlier can dominate the result. Ten to twenty give a more stable average and make high percentiles meaningful. For ongoing monitoring, a rolling window of dozens of readings, such as the last 60, smooths noise while still reacting to change.