percentileswebsite auditscore web vitalsseo reportingweb performance

What Does a Percentile Mean? Definition, Examples, and Website Metrics

A percentile shows where a value sits within an ordered comparison group, making it useful for interpreting test scores, health charts, and website performance data.

· 14 min read

A page can have a fast-looking median load time while 25% of its real user visits are much slower. That is why Google evaluates Core Web Vitals at the 75th percentile, not just the average: percentiles show how an experience is distributed across a group. (developers.google.com)

For anyone asking what does a percentile mean, the practical answer is simple: it tells them where a value stands relative to other values after the data has been ordered. This guide explains how to read percentiles correctly, calculate them with a worked example, avoid common mistakes, and use them to make better website audit decisions.

What does a percentile mean?

A percentile is a position measure. The *k*th percentile is a value at or below which roughly k% of observations in a comparison group fall. The 50th percentile is the median: the middle point of an ordered dataset. NIST describes percentiles as values that split ordered data into hundredths. (itl.nist.gov)

For example, consider a test score at the 90th percentile. It means the score is equal to or higher than approximately 90% of scores in the relevant comparison group. About 10% of the group scored higher.

Percentiles are useful because a raw number often lacks context. A page with a 3.2-second Largest Contentful Paint (LCP) needs comparison against a target or a set of user visits before an SEO or developer can decide whether it is typical, concerning, or exceptional.

The key idea is this:

  • A value might be 3.2 seconds, 82 points, or 60 kilograms.
  • A percentile explains that value's relative position in a defined dataset.
  • The comparison group matters as much as the number itself.

A percentile is not automatically a quality rating. Being at a high percentile can be good when higher is desirable, such as an exam score. It can be bad when lower is desirable, such as page-load time, error count, or customer-support wait time.

Percentile meaning in plain English

The easiest way to interpret a percentile is to translate it into “how many observations are at or below this point?”

PercentilePlain-English interpretation
10th percentileAbout 10% of values are at or below this value; about 90% are above it.
25th percentileRoughly one quarter of values are at or below it. This is also the first quartile, or Q1.
50th percentileHalf the values are at or below it. This is the median, or Q2.
75th percentileRoughly three quarters of values are at or below it. This is the third quartile, or Q3.
90th percentileAround 90% of values are at or below it, leaving about 10% above it.

The 25th, 50th, and 75th percentiles have special names because they divide a dataset into four broad sections: quartiles. The spread between the 25th and 75th percentiles is the interquartile range (IQR), a useful way to understand the middle half of a dataset without letting a handful of extreme values dominate the picture. (statisticsbyjim.com)

For a simple example, a child at the 25th weight-for-age percentile weighs the same as or more than 25% of the relevant reference population and less than 75% of it. The CDC uses percentiles as a clinical indicator of children’s size and growth patterns, but also stresses that growth charts should contribute to an overall picture rather than serve as a diagnosis by themselves. (cdc.gov)

That same caution applies to business and web data. A percentile is evidence, not a complete explanation.

A percentile is not the same as a percentage

One of the most common interpretation errors is treating a percentile as a percentage score. They answer different questions.

A percentage describes a share of a possible total. If a person answers 42 out of 50 questions correctly, their score is 84%.

A percentile describes how that result compares with other results. An 84% test score might be the 60th percentile in one difficult cohort and the 95th percentile in another easier cohort. The raw percentage can stay the same while the percentile changes because the comparison group changes.

Percentage versus percentile: a worked distinction

Suppose 100 candidates take a technical assessment:

  1. Candidate A gets 84% correct.
  2. In a particularly strong candidate pool, 40 people score higher.
  3. Candidate A is therefore around the 60th percentile.

The candidate’s result is still 84%. The percentile does not revise the result; it adds context about relative performance.

The same distinction matters in website reporting. “72% of accessibility checks passed” is a percentage of checks. “The page’s Interaction to Next Paint is at the 75th percentile” describes a point in the distribution of user experiences. Combining the two carelessly can create misleading client reports.

For a deeper comparison of a percentile with a mean, see Audra’s guide to percentile vs mean in website audits. The mean summarizes arithmetic average performance; a percentile describes a position in ordered performance data.

How percentiles are calculated

To calculate a percentile, analysts first sort the data from lowest to highest. They then find the position associated with the requested percentile and either select an observed value or interpolate between two observed values.

There is an important complication: there is no single universally accepted percentile calculation method. NIST documents multiple practical interpolation methods, while the nearest-rank method selects an actual observation rather than creating a value between observations. Different tools can therefore return slightly different answers for the same small dataset. (itl.nist.gov)

The nearest-rank method

A straightforward method uses this formula:

rank = ceiling((p / 100) × n)

Where:

  • p is the requested percentile, such as 75.
  • n is the number of observations.
  • ceiling means round up to the next whole-number position.

The answer is the value at that position in the ordered list.

Worked example: finding the 75th percentile

Imagine eight page-load measurements, in seconds, already sorted:

1.2, 1.5, 1.7, 2.0, 2.1, 2.5, 3.4, 5.8

To find P75 with nearest rank:

ceiling((75 / 100) × 8) = ceiling(6) = 6

The sixth value is 2.5 seconds. Under this method, the 75th percentile is 2.5 seconds: about 75% of the measurements are at or below 2.5 seconds.

If the rank calculation lands between positions, some tools interpolate. NIST’s default illustrated method uses a position based on p(N + 1) and interpolates between adjacent sorted values when appropriate. That is statistically legitimate, but it means an analyst should record the tool and method when comparing small datasets or reproducing an analysis. (itl.nist.gov)

Percentile, percentile rank, and quartile differences

People often use “percentile” and “percentile rank” interchangeably, and ordinary conversation usually survives that shortcut. In reporting and analysis, separating the terms makes the result clearer.

A percentile value is the cutoff value at a specified point in a distribution. For example, “P75 LCP is 2.5 seconds” identifies the performance value at the 75th percentile.

A percentile rank describes where an individual observation falls within a group. For example, “this page’s speed is at the 75th percentile relative to comparable pages” describes its relative rank. Definitions vary around whether equal values count as below, at, or below-and-equal, especially where ties occur. (itl.nist.gov)

A quartile is simply a named percentile:

  • Q1 = 25th percentile
  • Q2 = 50th percentile, or median
  • Q3 = 75th percentile

The distinction between percentile values and ranks becomes especially useful in audit dashboards. “P90 response time is 1,800 ms” is a threshold value. “This URL is in the 90th percentile for response time among crawled URLs” is a relative comparison. They are not interchangeable claims.

Readers who need examples of these labels in SEO reports can also review percentile definitions, ranks, and website metrics.

Why the comparison group changes the answer

A percentile has no useful meaning without a defined population, sample, or segment. A 90th percentile result only says where a value falls relative to a particular collection of observations.

For example, the CDC’s growth-chart percentiles compare a child against a reference population matched to attributes such as age and sex. Comparing a five-year-old’s measurement against all children of all ages would produce an unhelpful percentile. (cdc.gov)

Website metrics have the same requirement. A performance percentile can differ materially depending on whether the data is segmented by:

  • mobile or desktop users;
  • individual URL or site origin;
  • country, connection type, or device class;
  • the past 28 days or another reporting window;
  • first-time versus returning visitors; and
  • all page views versus a selected template group.

Google’s documentation specifically recommends evaluating Core Web Vitals at the 75th percentile of page loads, with mobile and desktop considered separately. PageSpeed Insights can aggregate field data at page or origin level when enough data is available. (developers.google.com)

A client report should therefore not merely say “the site is at P75.” It should say what metric, period, device segment, URL grouping, data source, and percentile method were used. That turns a vague statistic into an auditable finding.

Why P75 matters in website performance audits

Google uses P75 for Core Web Vitals because it describes a user experience nearer the slower end of the normal range without focusing exclusively on the most extreme outliers. In practical terms, a P75 result asks: “Can at least three out of four user visits achieve the target?”

As documented by Google, a page or origin with enough data passes the Core Web Vitals assessment only when the 75th percentiles of LCP, INP, and CLS are all in the “Good” range. The published thresholds are LCP of 2.5 seconds or less, INP of 200 milliseconds or less, and CLS of 0.1 or less. (developers.google.com)

Example: median can hide a real user problem

Consider 100 visits to a product page:

  • Median LCP: 1.8 seconds
  • P75 LCP: 2.7 seconds
  • P90 LCP: 5.1 seconds

The median says the typical visit may look healthy. But P75 shows that at least one in four visits takes longer than 2.7 seconds, above Google’s 2.5-second “Good” LCP threshold. P90 shows a more serious experience for the slowest 10%.

This is why an audit should review several percentiles rather than rely on one average. In Audra, a local-first audit workflow can pair performance distributions with technical SEO, accessibility, best-practice, link, and AI answer-engine visibility checks so a consultant has evidence for prioritizing the URLs that need attention first.

For a fuller explanation of percentile-based audit interpretation, see percentile meaning and website audit uses.

When high and low percentiles are good or bad

A percentile does not carry a built-in positive or negative meaning. Direction depends on the metric.

When higher is usually better

Higher percentile ranks are generally favorable for metrics where more is desirable:

  • test scores and conversion rates;
  • organic visibility relative to competitors;
  • completed accessibility checks;
  • uptime percentage; and
  • number of authoritative referring domains, with quality caveats.

If a site is in the 90th percentile for conversion rate within a credible benchmark set, it outperforms roughly 90% of that set on that metric.

When lower is usually better

Lower values are generally preferable for friction or risk metrics:

  • LCP and INP in milliseconds or seconds;
  • server response time;
  • JavaScript error rate;
  • crawl depth to important pages; and
  • count of broken internal links.

Here, report wording needs care. “P90 LCP is 4 seconds” does not mean the page performs better than 90% of visits. It means 90% of observed visits loaded at or below 4 seconds, while the slowest 10% took longer. Whether 4 seconds is acceptable depends on the relevant target.

A useful client-facing convention is to state the direction explicitly: “Lower is better; P75 is 2.7 seconds.” This small addition prevents a common reading error.

Common mistakes when interpreting percentiles

Percentiles are simple in concept, but reports often overstate what they prove. The following mistakes are especially common in SEO, analytics, and performance reviews.

  1. Treating a percentile as a score out of 100. A 75th percentile is not “75% correct” or “75 out of 100.” It is a relative position.
  2. Leaving out the comparison group. A percentile derived from 12 URLs is not directly comparable with one derived from 10,000 mobile user visits.
  3. Assuming all software uses the same formula. Small datasets can produce different percentile cutoffs depending on nearest-rank, interpolation, and tie handling. NIST explicitly notes multiple methods in use. (itl.nist.gov)
  4. Using only the mean. Averages can conceal a cluster of slow or failed experiences. P75 and P90 make the tail of the distribution visible.
  5. Equating correlation with cause. A poor percentile identifies a distributional problem; it does not by itself prove whether images, third-party scripts, hosting, or rendering behavior caused it.
  6. Ignoring sample size. P95 from 20 visits is based on very little information. The percentile may be mathematically calculable, but it is not necessarily stable enough for a confident decision.

A good audit uses percentiles alongside raw samples, medians, trend data, segmentation, and page-level diagnostics. That approach helps teams distinguish a persistent problem from a one-off outlier.

How to use percentiles in a practical audit report

Percentiles become more useful when they lead directly to prioritization. Rather than reporting a long table of isolated values, an agency or site owner can build an action-oriented view.

A five-step reporting approach

  1. Define the metric and direction. State whether lower or higher is better. For example: “Lower LCP is better.”
  2. Name the dataset. Specify page views, crawled URLs, query samples, or benchmark competitors, plus the date range and device segment.
  3. Show at least two distribution points. Median and P75 are often a strong baseline; add P90 when tail risk matters.
  4. Compare with a real threshold. Google’s P75 Core Web Vitals thresholds are concrete examples for user-experience performance work. (developers.google.com)
  5. Connect the finding to a fix. Group slow URLs by template, investigate the resources and rendering pattern they share, then retest after changes.

For example, an audit might state: “Across 42 mobile product URLs tested during the reporting period, median LCP was 1.9 seconds, P75 was 2.8 seconds, and P90 was 4.9 seconds. Lower is better. The P75 result misses Google’s 2.5-second good threshold, so image delivery and render-blocking dependencies on the product template should be investigated first.”

That language is more useful than “site speed needs improvement” because it identifies scope, severity, benchmark, and next action. For another practical walkthrough, see percentile explained with formulas and website audit use cases.

FAQ

What does it mean to be in the 75th percentile?

Being in the 75th percentile generally means a value is equal to or higher than about 75% of the values in its comparison group. For a metric where lower is better, such as page-load time, P75 instead means 75% of measurements are at or below that reported time. Always state the metric direction.

Is the 50th percentile the same as the average?

No. The 50th percentile is the median, the middle value in ordered data. The average, or mean, is found by adding all values and dividing by their count. A few unusually large values can pull the mean upward, while the median is less sensitive to those extremes. (itl.nist.gov)

How is a percentile different from percentile rank?

A percentile is usually the cutoff value associated with a percentage of ordered observations, such as P75 LCP of 2.5 seconds. Percentile rank usually describes an individual observation’s relative standing in a group. Everyday usage overlaps, but distinguishing them produces more precise reporting, particularly in analytics and audit work.

Why can two tools give different percentile results?

Tools can use different valid conventions for rank position, rounding, interpolation, and tied values. The difference is most visible in small datasets. NIST describes several percentile estimation methods used in practice, so reports should identify the source tool or calculation method when exact reproducibility matters. (itl.nist.gov)

Why does Google use the 75th percentile for Core Web Vitals?

Google evaluates Core Web Vitals at the 75th percentile to judge whether a good experience reaches most real-world visits rather than only the median visit. For data with sufficient coverage, PageSpeed Insights assesses LCP, INP, and CLS using their P75 values at the page or origin level. (developers.google.com)

Sources