Percentile Definition: Meaning, Formula, Examples, and Website Metrics
A percentile shows where a value sits in an ordered data set, making it useful for interpreting website performance, benchmarks, rankings, and audit results.
· 13 min read
A p75 Largest Contentful Paint of 2.5 seconds means 75% of measured page visits reached their largest visible content in 2.5 seconds or less—not that every visitor did. This percentile definition helps SEOs, agencies, and site owners turn a long list of website measurements into a clear answer about typical, poor, and unusually slow user experiences.
Percentile definition: what a percentile means
A percentile is a value below which a specified percentage of observations in an ordered data set falls. The 50th percentile is the median; the 75th percentile is the value at or below which roughly 75% of observations occur; and the 95th percentile identifies the cutoff experienced by all but the slowest or highest 5% of observations. NIST describes percentiles as values that divide ordered data into hundredths. (itl.nist.gov)
The most useful plain-English reading is this:
- p50: half of observations are at or below this value.
- p75: three quarters of observations are at or below this value.
- p90: nine out of ten observations are at or below this value.
- p95: only the worst 5% are above this value.
- p99: only the worst 1% are above this value.
For website metrics where lower is better—such as loading time, server response time, or Interaction to Next Paint—a lower percentile value is usually preferable. A p95 API response time of 1.8 seconds means 95% of requests finished in 1.8 seconds or faster, while the slowest 5% took longer.
Percentiles are not percentages of improvement, pass rates, or grades. If a page has a 90th-percentile load time of 4 seconds, it does not mean the page is 90% fast. It identifies a position in that page-load-time distribution.
Why percentiles are more useful than one average
An average can be useful, but it can conceal uneven experiences. Consider ten page-load measurements in seconds: 1, 1, 1, 1, 1, 1, 1, 1, 1, and 20. Their mean is 2.9 seconds, even though nine of ten visits loaded in one second.
The percentile view makes the distribution visible:
| Metric | Result in this example | What it reveals |
|---|---|---|
| Mean | 2.9 seconds | One severe outlier pulls the average upward |
| p50 | 1 second | The middle visitor experienced a fast load |
| p75 | 1 second | At least 75% of visits were fast |
| p90 | 1 second under a nearest-rank convention | Nine of ten visits were fast |
| p95 | depends on the calculation method | The slow tail needs closer inspection |
The opposite situation can also occur: an acceptable average can mask a meaningful number of slow visits. That is why monitoring systems commonly expose p50, p95, and p99. Google’s Site Reliability Engineering guidance notes that latency percentiles reveal whether 50%, 5%, or 1% of requests are too slow, whereas an arithmetic mean only signals that request time changed without showing who was affected. (sre.google)
For audit reporting, averages answer “what was the overall arithmetic result?” Percentiles answer “how consistently did users or pages experience this?” Both can belong in a report, but they should not be treated as interchangeable. For a deeper comparison, see percentile vs mean for website audits.
Percentile vs. percentile rank
The terms percentile and percentile rank are closely related, yet they begin with different inputs and give different outputs.
A percentile starts with a percentage and returns a value. For example, “the 75th percentile of page-load time is 3.2 seconds” returns a time value.
A percentile rank starts with a value and returns a percentage position. For example, “a 3.2-second page-load time has a percentile rank of 75 in this data set” describes where the observation sits relative to the rest of the sample.
A practical distinction
Suppose an agency audits 100 product pages for the number of broken internal links:
- The 90th percentile might be 12 broken links. This is a cutoff expressed in the original unit: links.
- A particular product page with 12 broken links might sit at the 90th percentile rank. This is a relative position expressed as a percentage.
That distinction matters in client reports. “p90 is 12 broken links” tells the reader where the upper-end issue threshold is. “This URL is in the 90th percentile” identifies how unusually affected that specific URL is.
In casual writing, people often say “in the 90th percentile” to mean either idea. A careful audit should state both the metric and unit: for example, “the URL’s 4.1-second LCP is at the 90th percentile among sampled category pages.”
How to calculate a percentile
Calculating a percentile always starts by sorting values from smallest to largest. The complication is that a requested percentile often lands between two observations rather than exactly on one of them.
NIST gives one approach based on the position p(N + 1), where p is expressed as a proportion and N is the number of observations. When the result falls between positions, the method interpolates between the surrounding values. NIST also explicitly notes that there is no universally accepted interpolation method. (itl.nist.gov)
Worked p75 example
Take these ten Lighthouse Largest Contentful Paint results, already ordered in seconds:
1.2, 1.4, 1.8, 2.0, 2.2, 2.7, 3.1, 3.8, 4.5, 6.0
Using the simple nearest-rank method:
- Multiply the desired percentile by the number of values:
0.75 × 10 = 7.5. - Round up to the next whole position: position 8.
- Select the eighth value: 3.8 seconds.
Under that convention, p75 is 3.8 seconds. In plain language, 75% of the sampled tests had an LCP at or below 3.8 seconds.
Using NIST’s p(N + 1) position method, the position is 0.75 × 11 = 8.25. That falls one-quarter of the way between 3.8 and 4.5 seconds, producing 3.975 seconds. Neither calculation is inherently deceptive; they use different rules. The report should name the software or percentile method when exact reproducibility matters.
Why percentile tools can show different numbers
Two tools can report slightly different p75 or p95 figures from the same underlying data. Differences usually come from one or more of these factors:
- Interpolation rule: Spreadsheet, analytics, and statistical tools can select or interpolate positions differently.
- Ties: Repeated values require a convention for assigning ranks.
- Rounding: A dashboard might round milliseconds to tenths of a second before display.
- Sampling period: A 28-day field-data window and a one-day RUM window do not contain the same visits.
- Segmentation: Mobile, desktop, country, browser, URL, and URL-group distributions can differ sharply.
- Histogram estimation: High-volume monitoring platforms may store counts in ranges rather than every raw measurement.
Google Cloud Monitoring explains that distribution metrics can use histogram buckets, and that bucket width, sample count, and measurement distribution influence estimated percentile values. It also notes that with fewer than 20 samples, p95 and p99 can land in the same bucket. (docs.cloud.google.com)
The practical lesson is not to chase tiny discrepancies across dashboards. Instead, use the same tool, population, device segment, window, and calculation method for before-and-after comparisons. If a performance report changes from p75 LCP 2.4 seconds to 2.7 seconds using the same method, that movement is easier to interpret than two numbers generated under different conditions.
Percentiles in website performance and Core Web Vitals
Percentiles matter directly in technical SEO because Google’s Core Web Vitals assess real-world loading, responsiveness, and visual stability. Google documents three primary targets: Largest Contentful Paint within 2.5 seconds, Interaction to Next Paint below 200 milliseconds, and Cumulative Layout Shift below 0.1. (developers.google.com)
The widely used assessment point is the 75th percentile of real-user page-load data. Therefore, a good p50 is not enough when p75 remains poor.
What p75 means for a page experience metric
If a page’s p75 LCP is 2.4 seconds, at least 75% of the relevant recorded visits had an LCP of 2.4 seconds or less. If it is 3.1 seconds, a substantial part of the visitor population has an experience above Google’s 2.5-second LCP target.
This makes p75 a reasonable operational threshold: it does not allow the slowest quarter of visits to disappear inside a median, but it also does not let a tiny number of extreme outliers define the whole page’s status. Google Search Central recommends achieving good Core Web Vitals as part of delivering a strong user experience and aligning with what its ranking systems seek to reward. (developers.google.com)
Lab testing remains useful for diagnosing an individual page build, but it is not a substitute for field percentiles. A lab test may show one simulated run; a field percentile summarizes many actual visits across devices, networks, and locations.
Choosing p50, p75, p95, or p99 for an audit
There is no universally correct percentile. The right choice depends on what the metric measures and what decision the audit needs to support.
| Percentile | Best use in a website audit | Example question |
|---|---|---|
| p50 | Typical experience | What does the middle visitor experience? |
| p75 | Broad consistency and Core Web Vitals | Is the experience acceptable for at least three quarters of visits? |
| p90 | Significant upper-end issues | How bad is performance for the slowest 10%? |
| p95 | Reliability and serious friction | Are a meaningful minority of checkout or API requests failing users? |
| p99 | Extreme-tail incidents | Are rare but severe timeouts or bottlenecks occurring? |
For an ecommerce site, p50 can show whether the storefront generally feels fast, p75 can track Core Web Vitals, and p95 can expose a poor mobile-network or cache-miss experience that harms a smaller but still valuable group of customers.
For crawl-based SEO metrics, percentiles can prioritize remediation. An auditor might calculate p90 for HTML response time across 5,000 URLs, then inspect the roughly slowest 10% first. Or they may use p75 for total page weight to set a realistic improvement target for templates rather than focusing on one unusually heavy campaign landing page.
A useful rule is to pair a percentile with a count. “p95 is 4.8 seconds across 8,400 product-page visits” is more informative than “p95 is 4.8 seconds” alone. Sample size determines how much confidence a reader should place in the tail.
Percentiles for website benchmarks and competitive context
Percentiles can also describe relative performance against a peer group. In that use, the result should be read carefully because direction matters.
If a conversion rate is at the 75th percentile versus comparable businesses, it usually means it is better than about three quarters of that peer group. But if page-load time is at the 75th percentile, it may mean the site is slower than many peers because larger time values are worse. The report should always label the favorable direction instead of assuming “higher percentile” means “better.”
Google Analytics provides benchmarking percentiles including the median, 25th percentile, and 75th percentile for peer groups of similar businesses. It uses industry categorization and, for absolute metrics, estimates comparison ranges that account for a property’s active-user count. (support.google.com)
Benchmarking has limits. Peer definitions, traffic mix, region, seasonality, and measurement configuration all influence the comparison. A local service business and a national ecommerce brand can both sit in a broad category while having very different customer journeys. Percentile benchmarks are best treated as a diagnostic prompt, not a universal target.
Applying percentiles to SEO and site-audit findings
Percentiles make large audit datasets easier to prioritize without pretending every URL has equal business impact. An audit can combine percentile thresholds with template type, conversion importance, crawlability, and revenue relevance.
Useful percentile-based audit checks
- HTML response time: Review URLs above p90 or p95 to locate origin, caching, database, or redirect issues.
- Page weight: Compare p50 and p90 transferred bytes by page template to distinguish a general bloat problem from a few asset-heavy pages.
- Internal-link depth: Find the 90th percentile of click depth, then assess whether valuable pages are concentrated in the deepest part of the architecture.
- Accessibility issue count: Use p75 or p90 issue counts per template, but investigate high-severity failures individually rather than relying only on aggregate statistics.
- AI answer-engine visibility checks: Compare the share of prompts or entities where pages are cited, surfaced, or absent, while keeping prompt set, country, model, and test date consistent.
A local audit tool can make this workflow more practical because it can crawl a site, retain the supporting URL-level evidence, and turn distribution-level findings into specific fixes. Audra combines AI visibility checks with technical SEO, performance, accessibility, best-practice, and link auditing so an auditor can identify the pages behind a p90 or p95 result rather than stopping at one summary statistic. Teams that need a site-wide performance workflow can also review how to run Lighthouse on a whole site.
For client reporting, state the metric as a complete sentence: “At p75, category-page LCP is 2.9 seconds on mobile; the target is 2.5 seconds.” Then name the likely affected audience: “approximately the slowest quarter of recorded visits is above the target.” That wording is much clearer than a color-coded score alone.
Common percentile mistakes to avoid
A percentile is powerful only when interpreted with its population and units. These recurring mistakes produce misleading recommendations.
- Calling p75 the average: p75 is a positional cutoff, not the arithmetic mean.
- Assuming p75 describes every user: It describes a threshold for 75% of the measured observations, not a guarantee for all visits.
- Ignoring directionality: A higher percentile is favorable for revenue and engagement but often unfavorable for latency, errors, or issue counts.
- Averaging percentiles: Averaging p95 values from two hosts, URLs, or time periods is not generally the same as calculating p95 from the pooled raw observations.
- Comparing unmatched segments: Desktop p75 from one month should not be compared casually with mobile p75 from another month.
- Hiding the sample size: A p99 from 15 visits is far less stable than p99 from 150,000 visits.
- Treating a tool’s output as exact: Interpolation, rounding, and histogram buckets can make a percentile an estimate.
A clear explanation of the term can prevent these errors. Readers who want alternate wording and more examples can see percentile meaning for website audits or this broader percentile explanation with formulas and use cases.
FAQ
What is a percentile in simple terms?
A percentile shows the value below which a chosen share of observations falls after the data is sorted. For example, if p75 page-load time is 3 seconds, then 75% of the measured page loads took 3 seconds or less. It is a way to describe the position of a value within a distribution.
Is the 50th percentile the same as the median?
Usually, yes. The 50th percentile is the median: the middle point of an ordered data set. Half of observations are at or below it and half are at or above it, subject to the calculation convention used for even-sized samples or repeated values. NIST identifies the 50th percentile as the median. (itl.nist.gov)
What does the 75th percentile mean for Core Web Vitals?
For Core Web Vitals, p75 represents the point at which 75% of relevant real-user observations are at or below the reported value. A p75 LCP of 2.5 seconds means three quarters of observed visits reached LCP within 2.5 seconds. Google’s published target for good LCP is within 2.5 seconds. (developers.google.com)
Why do Excel, analytics tools, and monitoring platforms show different percentiles?
Percentile formulas are not completely standardized. Tools can use different interpolation rules, ways of treating ties, rounding choices, sampling windows, and histogram approximations. The differences are often most noticeable with small data sets or high percentiles such as p95 and p99. For reliable trend analysis, use one consistent tool and method.
Should a website audit use p75 or p95?
Both can be useful. Use p75 when assessing broad consistency and Core Web Vitals-style experience across most users. Use p95 when diagnosing more severe tail problems, such as slow checkout requests, intermittent server delays, or the slowest page templates. The report should include sample size and the affected segment in either case.
Sources
- https://www.itl.nist.gov/div898/handbook/prc/section2/prc262.htm
- https://www.itl.nist.gov/div898/software/dataplot/refman2/auxillar/percenti.htm
- https://developers.google.com/search/docs/appearance/core-web-vitals
- https://support.google.com/analytics/answer/16388466?hl=en
- https://docs.cloud.google.com/monitoring/api/v3/distribution-metrics
- https://sre.google/workbook/monitoring/