statisticspercentileswebsite auditstechnical seoweb performance

Percentile vs Mean: What the Difference Means for Website Audits

A percentile shows relative position within a distribution, while a mean shows the arithmetic average—two measures that can lead to very different conclusions in website audits.

· 13 min read

Google evaluates Core Web Vitals at the 75th percentile rather than by average performance, because an average can conceal a meaningful share of slow visits. (web.dev) Understanding percentile vs mean helps SEOs, agencies, and site owners report website performance more honestly: they can distinguish a typical experience from the experience at the slow end of the distribution.

A percentile measures where a value sits relative to other observations. A mean is the arithmetic average of all observations. They can be similar in balanced data, but they can diverge sharply when a few unusually high or low values pull the average away from the middle.

What does percentile mean in statistics?

A percentile is a cutoff value below which a stated percentage of observations fall. For example, if a page-load-time dataset has a 90th percentile of 4.8 seconds, approximately 90% of measured loads are at or below 4.8 seconds and the slowest 10% are above it.

Percentiles describe relative position, not a share of points earned. A test score of 85% means 85 out of every 100 possible points were earned. Being in the 85th percentile means the score is higher than roughly 85% of scores in the comparison group. Those are different claims.

Quartiles are commonly used percentile cutoffs:

  • First quartile (Q1): the 25th percentile.
  • Second quartile (Q2): the 50th percentile, also called the median.
  • Third quartile (Q3): the 75th percentile.
  • Interquartile range (IQR): Q3 minus Q1, which summarizes the middle 50% of values.

NIST describes the IQR as the difference between the 75th and 25th percentiles and notes that it is less affected by extreme values than standard deviation. (itl.nist.gov) That makes percentile-based reporting especially useful for website metrics where a small number of failed requests, slow devices, or congested connections can create a long right tail.

Percentile vs mean: the essential difference

The mean and percentile answer different questions. Neither is automatically better; the right choice depends on whether the decision requires one overall average or a view of how results are distributed.

MeasureWhat it answersCalculationSensitivity to outliersWebsite-audit example
MeanWhat is the arithmetic average?Sum all values, then divide by the number of valuesHighAverage page load time across 10 test runs
Median / p50What is the middle observation?Sort values and find the midpointLowTypical page load time
p75What value covers 75% of observations?Sort values and locate the 75th-percentile cutoffLower than meanPerformance level experienced by the slower quarter of users
p95How bad is the upper tail for most users?Sort values and locate the 95th-percentile cutoffDesigned to expose tail performanceSlow-load or slow-crawl risk

The mean is a measure of central location. Standard deviation measures how widely values spread around that mean; because it squares deviations, a very large observation can have a disproportionate effect on the result. (itl.nist.gov) A percentile does not average distances from the center. It identifies a location in ordered data.

For client reporting, a mean can still be helpful. It is often compact and intuitive. But pairing it with p50, p75, or p95 prevents a report from implying that every visitor, URL, or crawl request had the average experience.

Is the 50th percentile the same as the mean?

No. The 50th percentile is the median, not necessarily the mean. In a perfectly symmetric normal distribution, the mean, median, and mode coincide. NIST lists all three at the location parameter, μ, for a normal distribution. (itl.nist.gov) Real-world audit data, however, is often skewed.

Consider seven page-load observations in seconds, sorted from fastest to slowest:

1.2, 1.3, 1.4, 1.5, 1.7, 2.0, 9.0

Using the nearest-rank approach:

  • p50 / median: 1.5 seconds.
  • Mean: (1.2 + 1.3 + 1.4 + 1.5 + 1.7 + 2.0 + 9.0) / 7 = 2.59 seconds.
  • p75: 2.0 seconds.
  • p95: 9.0 seconds in this very small sample.

One 9-second result pulls the mean up by more than one second above the median. The p50 still shows the center of the ordered observations. This is why a dashboard that reports only a 2.59-second average can make a mostly fast set of loads look slower than typical, while also failing to communicate how severe the single slow case was.

The 50th percentile equals the mean only under particular data shapes, such as a symmetric normal distribution. It can also happen by coincidence in other datasets, but it should not be assumed.

How to calculate a percentile from raw data

A percentile calculator can use more than one valid convention. Differences are most visible in small datasets, because software may use nearest rank, linear interpolation, or another quantile definition. NIST notes that percentile methods have not been fully standardized across software implementations. (itl.nist.gov) Therefore, reports should state the method when exact reproducibility matters.

The nearest-rank percentile formula

For a straightforward manual method:

  1. Sort the dataset from smallest to largest.
  2. Convert the desired percentile to a decimal, such as 0.75 for p75.
  3. Calculate the rank: r = ceil(p × n).
  4. Select the value at rank r, where n is the number of observations.

Suppose an audit records 10 Lighthouse performance scores:

48, 55, 61, 63, 67, 71, 74, 79, 83, 92

For p75:

  • p = 0.75
  • n = 10
  • r = ceil(0.75 × 10) = ceil(7.5) = 8
  • The eighth value is 79.

Under nearest rank, the 75th-percentile performance score is 79. This means at least 75% of the measured scores are 79 or below—not that 79% of anything happened.

Interpolation methods can return a value between observations

Spreadsheet functions and statistical tools often use interpolation. One common method calculates a zero-based position as:

position = (n − 1) × p

If the position falls between two sorted values, the method estimates a value between them. A calculator using an Excel-style inclusive approach may therefore return a slightly different p75 from nearest rank for the same 10 observations. The difference is methodological, not necessarily an error. (standarddeviationcalculator.net)

For ongoing SEO reporting, the practical rule is simple: use one method consistently, label it, and avoid comparing numbers produced by mismatched tools as if they were identical.

Percentile rank is different from a percentile value

The terms sound similar but work in opposite directions.

  • A percentile value starts with a percentage and returns a cutoff. Example: “What is the p90 page-load time?”
  • A percentile rank starts with a score or observation and returns its relative standing. Example: “What percentile rank is a 3.2-second load time?”

With raw data, percentile rank is generally based on the proportion of observations below, at, or below-and-equal-to the chosen value. The exact treatment of ties varies by convention, so a report should document the chosen rule for repeated values.

For example, if 85 of 100 sampled URLs have a lower crawl response time than 1.1 seconds, then 1.1 seconds is around the 85th percentile by a strictly-below convention. If several URLs tie at 1.1 seconds, a different calculator may assign a somewhat different rank.

This distinction matters during audits. A p95 response time answers, “How slow are the slowest 5% of responses?” A percentile rank for one URL answers, “How does this URL compare with the rest of the crawl?” Both can be useful, but they should not be interchanged in a client report.

How to find percentile using mean and standard deviation

Mean and standard deviation can estimate percentiles only when a normal distribution is a reasonable approximation. A normal distribution is the symmetric bell-shaped model defined by a mean, μ, and standard deviation, σ. (itl.nist.gov) Website performance data often has a right-skewed tail, so raw observed percentiles are usually safer when enough measurements are available.

Find a percentile value from a normal distribution

To estimate the value at percentile p:

x = μ + zₚσ

Where:

  • μ is the mean.
  • σ is the standard deviation.
  • zₚ is the z-score associated with percentile p.

Common z-scores include:

  • p50: z = 0
  • p75: z ≈ 0.674
  • p90: z ≈ 1.282
  • p95: z ≈ 1.645
  • p99: z ≈ 2.326

If a metric is reasonably normal with a mean of 2.0 seconds and a standard deviation of 0.5 seconds, its estimated p90 is:

2.0 + (1.282 × 0.5) = 2.64 seconds

Find percentile rank from a score

To estimate the percentile rank of a known score x:

z = (x − μ) / σ

Then use a standard-normal cumulative distribution table or software to convert z to a percentile. A z-score of +1.0 corresponds to about the 84.13th percentile under the normal-distribution assumption. (everycalculators.com)

The caveat is critical: mean plus standard deviation does not reveal actual percentiles for arbitrary data. Two distributions can share the same mean and standard deviation but have very different tails. Check a histogram, an empirical cumulative distribution, or the raw observations before using a normal-distribution percentile estimate.

A compact decision path: raw observations or normal approximation?

The following decision path helps select a percentile method without overstating precision.

  1. Do the raw measurements exist? Use them. Sort the values and calculate empirical percentiles with a documented convention.
  2. Is the sample very small? Report the sample size and treat p95 carefully. With only 10 observations, nearest-rank p95 is simply the largest value.
  3. Is the distribution visibly skewed or has clear outliers? Prefer median, quartiles, and empirical percentiles over normal estimates.
  4. Is a normal model defensible? Use mean and standard deviation with z-scores for an estimated percentile, clearly labeling it as an estimate.
  5. Are results compared across tools? Confirm each tool's percentile convention, data window, device mix, and population before declaring a regression.

An empirical cumulative distribution function can show the observed distribution directly and help assess whether a fitted normal curve is appropriate. (statisticsbyjim.com) This is more rigorous than assuming every metric forms a bell curve because the mean and standard deviation are available.

Why percentiles matter in website performance audits

Core Web Vitals provides a practical example of percentile thinking. Google guidance says tools assessing Core Web Vitals should consider a page passing when it meets recommended targets at the 75th percentile for all three metrics. (web.dev) Field data is a distribution of real-user experiences, rather than one synthetic test result. (web.dev)

Imagine 10 real-user Largest Contentful Paint measurements in seconds:

1.4, 1.5, 1.5, 1.6, 1.6, 1.7, 1.8, 2.0, 3.4, 8.9

Using nearest rank:

  • Mean: 2.54 seconds.
  • p50: 1.6 seconds.
  • p75: 2.0 seconds.
  • p95: 8.9 seconds.

The mean says “2.54 seconds,” which is useful but incomplete. The p50 says a typical measured experience is 1.6 seconds. The p75 says the slower quarter starts beyond 2.0 seconds. The p95 exposes a serious tail issue that could affect users on weak networks, older devices, or specific routes.

Google also recommends looking beyond p75 to p90 or p95 when diagnosing regressions, while historical FID guidance specifically pointed to p95–p99 for particularly poor first experiences. (web.dev) The broader audit lesson remains valid: report the target percentile for compliance, then inspect higher percentiles to find the experience of the users who are struggling most.

Applying percentiles to SEO, crawling, and accessibility work

Percentiles are not only for page speed. Agencies can use them across a technical audit to identify a problematic tail without letting a sitewide average obscure it.

Useful applications include:

  • Crawl response time: p50 describes the normal server response; p95 highlights the pages most likely to slow a large crawl.
  • HTTP errors: a mean is usually not meaningful for a binary 404 or 500 status. Use counts, rates, and segmentation instead.
  • Redirect chains: p75 or p95 chain length can reveal whether a small set of URLs is creating excessive hops.
  • Accessibility checks: report the median issues per page and p90 or p95 to identify templates with unusually high defect density.
  • Performance scores: use quartiles to show whether weak scores are concentrated in a minority of pages or spread across the site.

For crawl diagnostics, a p95 response-time spike paired with a growing rate of HTTP 429 responses may indicate rate limiting rather than universally slow hosting. The practical next steps differ: reduce concurrency, respect Retry-After, review server logs, and separate temporary throttling from persistent server problems. See the related 429 errors during crawls diagnostic playbook for that workflow.

A local-first audit process can preserve the underlying URL-level measurements needed for these checks. Instead of presenting only one audit score, agencies can segment percentiles by template, device class, status code, or directory and show clients where the tail risk actually sits.

How to report p50, p75, and p95 without confusing clients

A useful report should translate percentiles into operational language. Avoid assuming that every stakeholder knows that “p95” refers to a cutoff rather than an average.

A concise reporting format is:

Metricp50p75p95Interpretation
Page load time1.6s2.0s8.9sMost loads are fast, but the slowest 5% need investigation
Crawl response time240ms410ms2,800msServer is usually responsive; a tail of slow URLs may constrain crawl efficiency
Accessibility issues per page2518Typical pages have few issues, while some templates likely need remediation

Always include four pieces of context:

  1. Population: Which pages, visits, URLs, or test runs were included?
  2. Time window: Was the data collected during one crawl, a 28-day field-data period, or a release comparison?
  3. Method: Nearest rank, interpolation, or the software's named method.
  4. Segmentation: Desktop versus mobile, template group, geography, authenticated state, or other meaningful cohorts.

This context prevents false comparisons. For example, PageSpeed Insights field data and a single lab test can disagree because field data captures real users across devices and networks while lab data uses controlled conditions. (web.dev) A report should treat that as different evidence, not automatically as a tool failure.

For additional examples of turning percentiles into audit decisions, see Audra's guide to percentile formulas, examples, and website-audit use cases.

FAQ

What does percentile mean in statistics?

A percentile is a value below which a given percentage of observations falls. The 75th percentile is the value at or below which about 75% of the data occurs. It measures relative position in an ordered dataset, unlike a percentage, which expresses a proportion out of 100.

Is percentile the same as mean or average?

No. The mean is the arithmetic average of every value. A percentile is a cutoff in the sorted distribution. The mean can be heavily affected by unusually large or small values, while the median and other percentiles show where observations sit without averaging their distances from the center.

What does it mean to be in the 90th percentile?

Being in the 90th percentile generally means a value is higher than about 90% of the comparison group. For a metric where lower is better, such as page load time, the 90th percentile is a warning sign: it marks the boundary beyond which the slowest 10% of observations fall.

How do you calculate a percentile from a data set?

Sort the values, choose a percentile method, and locate the corresponding rank. In the nearest-rank method, use ceil(p × n), where p is the percentile as a decimal and n is the number of observations. Other tools interpolate between observations, so small differences are normal when methods differ.

How do you find percentile using mean and standard deviation?

Only under a reasonable normal-distribution assumption. First find the percentile's z-score, then calculate x = mean + (z × standard deviation). For instance, p95 uses a z-score of about 1.645. For skewed web-performance data, calculate percentiles from the raw observations instead of relying solely on this estimate.

Sources