statisticspercentileswebsite performancecore web vitalsseo reporting

Percentile Definition: Examples, Ranks, and Website Metrics

A percentile identifies the value at or below which a chosen share of an ordered data set falls, helping teams interpret scores, performance distributions, and outliers.

· 13 min read

With 10 page-load measurements, the nearest-rank P90 is the ninth value—not an interpolated estimate between values. This percentile definition guide shows how to calculate and interpret percentiles consistently, so SEOs, marketers, and site owners can explain scores, slow page loads, and outliers without confusing a percentile with a percentage or an average.

A percentile result can vary by calculation convention, particularly in a small data set. The examples below use the nearest-rank method throughout: sort values from low to high, calculate the rank with ceil((k / 100) × n), and select the value at that rank.

What is a percentile?

A percentile tells you the value at or below which a given percentage of observations in an ordered data set fall. For example, the 70th percentile is the value at or below which 70% of observations fall under a stated calculation convention. This is the core percentile definition used in statistics, test-score reporting, and performance monitoring. (NIST)

Percentiles are commonly written as P50, P75, P90, and P95. The number names the location in the distribution; the reported result is a value in the metric’s own unit, such as seconds, milliseconds, dollars, or points.

PercentileWhat the percentile value indicatesTypical use
P50The value at the middle of the distributionTypical result; often called the median
P75A value that covers the lower 75% of observationsBroad consistency checks
P90A value near the slower or higher endInvestigating an upper tail
P95A value near the most extreme 5% of a large sampleFinding severe delays or high counts

A percentile is not itself a score out of 100. “P75 LCP is 2.4 seconds” means that 2.4 seconds is the 75th-percentile value. It does not mean the page earned 75% or that 2.4 seconds has a percentile rank of 75 in every possible comparison group.

The same concept works for many website-audit measures: response time, image size, number of accessibility issues per page, redirect-chain length, or an AI visibility score calculated across a defined set of pages. The metric and the population being measured must always be stated alongside the percentile.

A 10-value website-performance example

Consider these 10 page-load times in seconds, already sorted from fastest to slowest:

1.1, 1.3, 1.4, 1.6, 1.8, 2.0, 2.2, 2.7, 3.4, 5.8

Using nearest rank, P50 is the fifth value, 1.8 seconds. P90 is the ninth value, 3.4 seconds. P95 is the 10th value, 5.8 seconds, because ceil(0.95 × 10) equals 10.

Requested percentileNearest-rank calculationReported percentile value
P50ceil(0.50 × 10) = 51.8 seconds
P70ceil(0.70 × 10) = 72.2 seconds
P90ceil(0.90 × 10) = 93.4 seconds
P95ceil(0.95 × 10) = 105.8 seconds

What this small sample can and cannot say

For P90, exactly 9 of the 10 observations are at or below 3.4 seconds, and 1 of 10 is above it. That aligns neatly with 90% because the sample size is 10.

P95 requires more care. Under nearest rank, the P95 value is 5.8 seconds, but all 10 observations—100%, not literally 95%—are at or below that value. A data set of 10 observations cannot represent every one-percentage-point position as a distinct observed rank. It can only represent steps of 10%.

Therefore, the accurate interpretation is: under the nearest-rank convention, P95 for this 10-observation sample is 5.8 seconds. It is not evidence that precisely 95% of visits were 5.8 seconds or faster. With a larger sample, such as 1,000 visits, P95 has a much finer resolution.

The arithmetic average of the 10 times is 2.33 seconds. That average is useful, but it hides that one recorded load took 5.8 seconds. Percentiles describe positions in the ordered distribution, while the mean summarizes values after adding and dividing them. NIST distinguishes these types of descriptive statistics. (NIST SP 1500-28)

Percentile vs percentage vs average

Percentage, percentile, and average all use numbers, but they answer different questions. Mixing them produces misleading reporting.

A percentage is a part of a known total. If 72 of 100 URLs return a 200 HTTP status, 72% of the URLs meet that condition. The denominator is the total number of URLs.

A percentile value is a threshold from a sorted set. If P72 page-load time is 2.9 seconds, then 2.9 seconds is the selected point in the observed distribution under the chosen method. It is not a 72% performance score.

An average is a calculated central value. For the 10 load times above, the mean is 2.33 seconds, found by adding all 10 times and dividing by 10. Averages can be pulled upward by a few unusually slow values, while percentiles make it easier to inspect specific parts of the distribution.

TermQuestion it answersWebsite example
PercentageWhat share of a total meets a condition?72 of 100 URLs pass a check = 72%
Percentile valueWhat value marks a point in ordered data?P90 load time = 3.4 seconds
Percentile rankWhat share of a comparison group is at or below a score?A score ranks above roughly 75% of a cohort
AverageWhat is the arithmetic central level?Mean load time = 2.33 seconds

A student can score 72% on an assessment and have a percentile rank of 90 if most other students scored below 72%. Percentage measures achievement against available points; percentile rank measures standing against a comparison group. (GeeksforGeeks)

How to calculate a percentile from ordered data

There is no single universal percentile formula used by every spreadsheet, analytics package, or programming library. NIST documents multiple calculation methods, including methods that interpolate between observations. (NIST Dataplot Reference Manual)

For transparent manual work, the nearest-rank method is easy to document and reproduce.

Nearest-rank calculation steps

For a requested percentile k and a data set containing n values:

  1. Sort the values from lowest to highest.
  2. Calculate rank = ceil((k / 100) × n).
  3. Select the value at that rank, counting the first value as rank 1.
  4. Report the method, sample size, metric unit, and period with the result.

Using the 10 page-load values, P90 is calculated as follows:

rank = ceil((90 / 100) × 10) = ceil(9) = 9

The ninth value is 3.4 seconds, so P90 is 3.4 seconds under nearest rank. P95 is ceil(9.5) = 10, so its value is 5.8 seconds.

Why tools may return another number

A spreadsheet may interpolate and return a value between 3.4 and 5.8 for P95. Another tool may use a formula based on n - 1, use inclusive or exclusive endpoints, or handle ties differently. None of those outputs is automatically wrong; they simply use different conventions.

The practical requirement is consistency. A monthly report should not compare a nearest-rank P90 from January with an interpolated P90 from February and call the difference a performance change. For large datasets, use the reporting tool’s documented method consistently and retain the raw sample size.

What P70, P90, and P95 mean in practice

The percentile number identifies the requested cutoff. The percentile value is what must be judged as good, bad, high, or low for the underlying metric.

If P70 image-transfer size is 450 KB, 450 KB is the percentile value. Under a method that supports this interpretation, roughly 70% of measured images are 450 KB or smaller, while the remaining portion is larger. For image weight, lower values generally indicate less data to transfer.

If P90 server response time is 1.8 seconds, 1.8 seconds is the value around the 90th-percentile cutoff. It helps reveal a slower portion of requests that a P50 result might conceal. It does not mean the server earned 90% or has a “90th-percentile score” independent of the chosen dataset.

If P95 response time is 4.5 seconds in a sufficiently large dataset, it indicates that the upper tail reaches 4.5 seconds. The next step is not to assume a single cause. Segmenting by page template, country, device, cache status, or third-party dependency may show whether a specific condition drives the tail.

For a 10-observation nearest-rank sample, however, P95 selects the maximum value. Analysts should say “nearest-rank P95 from 10 observations” rather than implying a precise 95/5 split. Sample size changes how finely any percentile can describe the distribution.

Percentile, percentile rank, and percentile range

These three terms are related but should not be used as substitutes.

A percentile calculation starts with a requested location and returns a metric value. “P75 Largest Contentful Paint is 2.4 seconds” starts with 75 and produces 2.4 seconds.

A percentile rank starts with a score or observation and returns a relative standing. If a test score has a percentile rank of 75, approximately 75% of the comparison group scored at or below that score. Percentile rank requires a defined comparison population; changing the cohort can change the rank even if the score stays the same.

A percentile range can have two meanings. On some assessment reports, it is a range of plausible percentile ranks reflecting measurement uncertainty. In analytics, it commonly means the span between two percentile values, such as P25 through P75.

For example, if P25 response time is 0.8 seconds and P75 is 1.9 seconds, the middle 50% spans 1.1 seconds. The formal interquartile range is P75 minus P25. NIST defines it as a measure of spread based on the middle half of the data. (NIST)

A clear report can therefore say: “P75 response time is 1.9 seconds; the P25–P75 range is 0.8–1.9 seconds.” It should not call 1.9 seconds a “75th-percentile rank.”

Website performance: why P75 is the primary benchmark

Website visits occur under different conditions: mobile and desktop devices, fast and slow networks, different countries, varied page templates, and changing server states. Field performance data is therefore a distribution rather than a single universal load time.

Google’s Core Web Vitals assessment uses the 75th percentile of field data. Its published thresholds classify Largest Contentful Paint as good at 2.5 seconds or less, needing improvement above 2.5 through 4 seconds, and poor above 4 seconds at P75. For Interaction to Next Paint, the good threshold is 200 milliseconds or less and poor is above 500 milliseconds, also assessed at P75. (web.dev)

P90 and P95 are not Core Web Vitals pass/fail criteria. They can still be useful internal diagnostic views of the slow tail, but their value depends on the audience, sample size, and segmentation. Google’s field-measurement guidance recommends considering the distribution and monitoring regressions rather than relying on one isolated value. (web.dev)

A practical performance review can use three checkpoints:

  1. P50 for a typical observed experience.
  2. P75 for the Core Web Vitals assessment point and broad consistency.
  3. P90 or P95 for exploratory analysis of unusually slow experiences, clearly labelled with the calculation method and sample size.

Lab tools and field datasets serve different purposes. A single Lighthouse run is controlled and repeatable, while field data represents real visitor conditions. Google explains the distinction between lab and field data in its performance documentation. (web.dev) For sitewide testing context, see this guide to running Lighthouse across a whole site.

Is a higher percentile better or worse?

The answer depends on whether “higher” refers to a percentile rank or a percentile value, and on what the metric measures.

A higher percentile rank is generally favorable for a desirable outcome. A test score at the 90th percentile rank is higher than the scores of roughly 90% of the comparison group. A conversion rate that sits at a high rank relative to comparable campaigns may likewise be favorable.

A higher percentile value is favorable only when larger metric values are desirable. For example, a P90 cache-hit rate of 98% may be welcome if higher hit rates are better. But a P95 response-time value of 5 seconds is worse than a P95 response-time value of 1 second because lower latency is better.

Use this rule of thumb:

  • For benefit metrics such as revenue, completed conversions, or successful cache-hit rate, higher values may be better.
  • For cost metrics such as milliseconds, bytes, errors, redirects, or accessibility issues per page, lower values are usually better.
  • For metrics with a target band, such as a temperature or inventory range, compare the percentile value with the target rather than assuming either direction is desirable.

The wording “high percentile” alone is ambiguous. A reliable statement names both items: “The page has a high percentile rank for conversion rate,” or “The P95 latency value is high at 5 seconds.”

Using percentiles in an SEO or audit report

Percentiles are most useful when they lead to a focused investigation rather than become a decorative statistic. A report that says “P95 is 4.5 seconds” lacks enough context to guide a decision.

A stronger report includes the metric, unit, population, time window, sample size, and method. For example: “For 1,000 mobile visits to product pages during October 2026, nearest-rank P75 LCP was 2.8 seconds.” That statement can be compared responsibly with another report using the same conditions.

A practical checklist is:

  • Name the metric and unit, such as P75 LCP in seconds or P90 image weight in KB.
  • Define the population: mobile visits, HTML URLs, product templates, or a specified country.
  • Include the date range and number of observations.
  • State the percentile convention where small samples or cross-tool comparisons matter.
  • Show P50 and P75 together when typical and broadly consistent performance both matter.
  • Add P90 or P95 as a diagnostic tail view, not as a substitute for Core Web Vitals assessment.
  • Segment before assigning a cause: template, device, route, country, or cache condition can change the result.

Website auditing also involves more than performance distributions. Technical SEO, accessibility, link health, and answer-engine visibility each require their own metrics and definitions. Teams evaluating local crawling workflows can compare approaches in Audra versus Sitebulb for technical SEO and AI search workflows. For related interpretation examples, see Percentiles Explained: Formula, Examples and Practical Uses.

FAQ

Is a higher percentile better or worse?

A higher percentile rank is generally better when higher scores represent better outcomes, such as test performance or conversion rate. A higher percentile value is not automatically better. For delay, error, byte-size, and issue-count metrics, lower P75, P90, or P95 values are usually preferable. Always judge the value according to what the metric measures.

What does the 70th percentile mean?

The 70th percentile is the value at or below which 70% of an ordered dataset falls, subject to the chosen calculation method. If P70 page-load time is 2 seconds, the result places 2 seconds at that distribution cutoff. It is not a claim that the page scored 70% or that 70% is inherently good.

What does the 95th percentile mean?

P95 is a value near the upper end of a distribution, associated with the point where about 95% of observations are at or below it in a large dataset. With nearest rank and only 10 observations, P95 selects the 10th value, so all observations are at or below it. State the method and sample size to avoid false precision.

How is a percentile different from a percentage?

A percentage is a share of a total: 80 passing URLs out of 100 is 80%. A percentile is a location in ordered observations: P80 response time is the selected response-time value at the 80th-percentile cutoff. A percentile rank instead describes where one score stands relative to a comparison group.

How do you calculate a percentile from an ordered data set?

First sort values from low to high and choose a calculation convention. Under nearest rank, calculate ceil((k / 100) × n), where k is the requested percentile and n is the number of observations, then select that ranked value. Other software may interpolate, so document the method when comparing results.

Sources