Quick answer
The 95th percentile (P95) is the value that 95% of measurements fall at or below, so only 5% are higher. It describes the slow tail an average hides: on 20 sample page loads the mean is 2.51 s but P95 is 6.8 to 9.4 s, depending on the method.
Formula: Nearest rank: P95 = value at rank ⌈0.95 × n⌉ in sorted data. Interpolated (NIST): position 0.95 × (n + 1), then interpolate between neighbours.
What does the 95th percentile mean?
The 95th percentile (P95) of a set of measurements is the value that 95% of them fall at or below, so only 5% are higher. If P95 page load time is 6.8 seconds, then 19 of every 20 loads were 6.8 seconds or faster and 1 of 20 was slower. NIST defines a percentile this way: a value such that at most 100p% of the measurements are less than it (NIST/SEMATECH e-Handbook).
A percentile is a position in sorted data, not a score out of 100. "In the 95th percentile" for a test means 95% of test takers scored the same or lower, not that the person answered 95% of questions correctly. The same idea in performance work answers a question the average cannot: how bad is it for the slowest 5% of users? That is why P95 appears in latency targets, page-speed reports and network billing. The broader idea is in percentile, and the neighbouring cut used by Core Web Vitals is 75th percentile.
How do you calculate the 95th percentile?
Sort the values from smallest to largest, then pick the one at the 95% position. There is more than one way to define that position, and the choice matters on small samples. Sample data: 20 page-load times in seconds, sorted:
1.1, 1.2, 1.2, 1.3, 1.4, 1.5, 1.5, 1.6, 1.7, 1.8, 1.9, 2.0, 2.1, 2.3, 2.4, 2.6, 2.9, 3.4, 6.8, 9.5
| Method | Rule | Rank for n = 20 | P95 |
|---|---|---|---|
| Nearest rank | Value at rank ⌈0.95 × n⌉ | ⌈19.0⌉ = 19 | 6.8 s |
| Linear interpolation (spreadsheet default) | Position 1 + 0.95 × (n − 1) | 19.05 | 6.935 s |
| NIST method (rank p(N+1)) | Position 0.95 × (n + 1), interpolate | 19.95 | 9.365 s |
Recompute the NIST row: p(N+1) = 0.95 × 21 = 19.95, so k = 19 and d = 0.95. The estimate is the 19th value plus 0.95 times the gap to the 20th: 6.8 + 0.95 × (9.5 − 6.8) = 6.8 + 2.565 = 9.365. The NIST handbook describes this as taking the k-th ordered value plus d times the difference to the next (NIST/SEMATECH e-Handbook).
The three answers span 6.8 to 9.4 seconds on the same 20 numbers. The reason is simple: with 20 values, the top 5% is a single observation (the 9.5), so the "P95" depends on how the method treats that one point. The methods agree closely on large samples. The practical rule is to know which method your tool uses and not compare numbers across tools with different methods.
Why is P95 better than the average for speed?
The average hides the slow tail. In the same 20 loads, the mean is 50.2 ÷ 20 = 2.51 s and the median is 1.85 s. Both look fine. Yet two loads took 6.8 and 9.5 seconds, 10% of visits, and 25% of loads (5 of 20) were slower than 2.5 seconds. A report that shows only the mean tells the team the page is fast.
| Statistic | Value on the 20 loads | What it tells you |
|---|---|---|
| Mean | 2.51 s | Pulled up by two slow loads, so not typical |
| Median (P50) | 1.85 s | The typical visit |
| P75 | 2.45 s (interpolated) | The speed three out of four visits met |
| P90 | 3.74 s (interpolated) | The slowest 10% start here |
| P95 | 6.8 to 9.4 s depending on method | The slow tail |
Google's Site Reliability Engineering book makes the point with a quantity: an average latency of 100 ms at 1,000 requests per second can coexist with 1% of requests taking 5 seconds, and if you do not look at the distribution, assuming most requests are near the mean is hopeful thinking (Google SRE Book, Monitoring Distributed Systems). A 5-second tail is 50 times the average, and 1% of 1,000 requests a second is 10 users every second.
Which percentile should you use for what?
| Percentile | Typical use | Why |
|---|---|---|
| P50 (median) | The typical experience | Robust to outliers |
| P75 | Core Web Vitals pass or fail | Google measures the 75th percentile of page loads, segmented across mobile and desktop |
| P90 | Service goals that tolerate a small tail | Cheaper to meet than P95 |
| P95 | Latency targets, capacity, network billing | The slow tail without the noise of extremes |
| P99 | Strict latency goals | Catches the worst 1%, needs far more data |
A page passes Core Web Vitals only if it meets the recommended targets at the 75th percentile for all three metrics: Largest Contentful Paint at 2.5 seconds or less, Interaction to Next Paint at 200 milliseconds or less and Cumulative Layout Shift at 0.1 or less (web.dev, Web Vitals). So when someone says "the P75 is good", they mean three of four visitors had a good experience; P95 asks about the fourth.
How many data points does a stable P95 need?
A percentile is only as stable as the tail behind it. P95 is decided by the top 5% of the data, so the number of observations in that tail is what matters:
| Samples (n) | Observations above P95 | How stable |
|---|---|---|
| 20 | 1 | One value decides it, as in the example |
| 100 | 5 | Moves noticeably with a few slow loads |
| 400 | 20 | Reasonably stable |
| 2,000 | 100 | Stable |
As a rule of thumb (an engineering judgement rather than a standard), you want about 20 observations in the tail, which means roughly 400 samples. Drop the 9.5 s load from the 20-load sample and the mean falls from 2.51 s to 2.14 s, a 15% change from removing one value in 20, while the nearest-rank P95 collapses onto the new maximum (6.8 s); with so few values nothing is safe. If a page or a country has only a handful of samples, report the median and the count instead.
Can you average percentiles from different groups?
No. A percentile of the combined data is not the average of the groups' percentiles. Sample data: page A has 100 loads from 1.00 to 1.99 seconds, page B has 300 loads from 2.00 to 4.99 seconds.
| Quantity | Value |
|---|---|
| P95 of page A (nearest rank 95) | 1.94 s |
| P95 of page B (nearest rank 285) | 4.84 s |
| Simple average of the two P95s | 3.39 s |
| Load-weighted average of the two P95s | (100 × 1.94 + 300 × 4.84) ÷ 400 = 4.12 s |
| P95 of all 400 loads (rank 380) | 4.79 s |
Both averages are too low because the combined tail comes mostly from page B. To get a correct combined percentile you need the raw values or a histogram, not the groups' summaries. Dashboards that average per-minute percentiles into an hourly figure make the same mistake.
What is 95th percentile billing?
Network providers often bill bandwidth on the 95th percentile. Usage is typically sampled every 5 minutes, the samples are sorted for the month, the top 5% are discarded and the next highest becomes the billable rate. A 30-day month has 30 × 24 × 12 = 8,640 samples, so the top 432 (36 hours) are ignored, which lets a customer burst for up to about 72 minutes a day without extra charge (Wikipedia, Burstable billing). A sustained peak longer than that does count, so a single short spike is free but a long one raises the bill.
What mistakes do people make with P95?
- Quoting P95 from a handful of samples. The example above shows one data point deciding it.
- Comparing numbers from tools that use different methods. 6.8 and 9.4 seconds can describe the same data.
- Averaging percentiles. Use raw data or histograms, as the aggregation table shows.
- Reporting only the average. It hides the slow tail that real users feel.
- Reading P95 as "95% are fine". It means 5% are as bad as that or worse, which at scale is a lot of people.
- Forgetting the metric. "P95 is 3 seconds" means little without saying whether it is load time, response time or something else.
Where do you see percentiles in TapCub?
Performance reports in TapCub are built from real loads in the visitor's browser, not lab runs. In Monitoring › Performance, each metric card (Largest Contentful Paint, Cumulative Layout Shift, Interaction to Next Paint, First Contentful Paint and Time to First Byte) can show P95, P75, median or average, while the trend and breakdown can use P50, P75, P90 or P99. The breakdown dimensions are page, country, device, browser and operating system, and the Slowest pages list is sorted by P75 Largest Contentful Paint.
The documentation's advice matches this article: start with P75 and the share rated good, then break down by page to find slow paths, and switch the card to P95 to see the tail. Percentiles jump when a page has few samples, and the page has no previous-period comparison, so note last month's value yourself. Metrics are reported by the visitor, so a blocked script leaves the page empty rather than substituting server time. For how to read the results by page, see the check Core Web Vitals guide.
Sources
- NIST/SEMATECH e-Handbook of Statistical Methods: percentiles, checked 2026-11-21
- Google SRE Book: Monitoring Distributed Systems, checked 2026-11-21
- web.dev: Web Vitals, checked 2026-11-21
- Wikipedia: Burstable billing, checked 2026-11-21
Was this explanation clear?
We use your answers to fix and expand this page.
Couldn’t send. Please try again in a minute.
Thanks for the feedback. We’ll use it to improve this page.You’ve already sent feedback on this page. Thanks.

Customer Experience Consultant
Writes about live chat, visitor conversations, and self-service, sharing practical ways to make customer communication clearer and more effective.