Percentile Explained: Formula, Examples, and Website Audit Use Cases
A percentile shows where a value sits in an ordered dataset, helping website teams interpret performance, SEO, accessibility, and AI visibility patterns that simple averages can hide.
· 13 min read
A website can report an average Largest Contentful Paint (LCP) of 1.8 seconds while its 75th-percentile visitors wait more than 2.5 seconds. Understanding a percentile helps SEO teams calculate and interpret those slower experiences, so they can turn audit exports into more accurate priorities rather than relying on a misleading average.
What is a percentile?
A percentile is a value that indicates an observation’s relative position within an ordered dataset. The 75th percentile, commonly written as p75, is the value at or below which approximately 75% of observations fall; the remaining 25% are above it.
For example, imagine 100 recorded page loads. If p75 LCP is 2.4 seconds, roughly 75 loads were 2.4 seconds or faster and roughly 25 loads took longer. It does not mean the page achieved a score of 75%.
The values must be sorted from smallest to largest before a percentile can be calculated. The 50th percentile is the median, while the 25th and 75th percentiles are also called the first and third quartiles. NIST notes that percentiles identify points in a distribution rather than a proportion of a single page, user, or score.
In a website audit, one observation might be:
- An LCP, Interaction to Next Paint (INP), CLS, or HTML-response-time measurement
- A crawl depth, redirect-hop count, internal-link count, or image-byte total
- An accessibility-error count across audited URLs
- A numeric AI-visibility measure, such as competitor citations per prompt
A percentile therefore describes the dataset’s shape. It helps explain how the middle, the fast end, and the slow or risky tail differ.
Percentile vs percentage vs percentile rank
The words percentage, percentile, and percentile rank are often used interchangeably in reports, but they answer different questions.
A percentage describes a share of a whole. If 92 of 100 crawled pages return an HTTP 200 status, then 92% of the sample has that outcome. It is a rate, not a relative position.
A percentile identifies a cutoff value in an ordered distribution. A p75 response time of 620 milliseconds means approximately 75% of measured response times are at or below 620 ms. A percentile rank describes the relative standing of one observation against a named comparison group.
| Term | Question it answers | Website-audit example |
|---|---|---|
| Percentage | How much of the total? | 18% of URLs have missing meta descriptions |
| Percentile | What value marks a point in the distribution? | p75 LCP is 2.4 seconds |
| Percentile rank | Where does one item sit among peers? | A page’s 3.8-second LCP is slower than 94% of pages in the sample |
The comparison group is essential. A page in the 90th percentile for response time among 200 product URLs is not automatically in the 90th percentile across an industry, a country, or the public web.
For a broader explanation of the language used in client reporting, see Audra’s percentile meaning guide. The practical rule is simple: state the dataset before stating the percentile.
Why averages can hide website problems
An average compresses every observation into one number. That makes it convenient, but a mean can disguise a slow tail caused by low-end devices, unstable connections, heavy templates, third-party scripts, or a small collection of problematic URLs.
Consider these 12 mobile LCP observations in seconds, sorted from fastest to slowest:
0.8, 1.0, 1.1, 1.3, 1.5, 1.7, 1.8, 2.0, 2.1, 2.6, 3.0, 4.2
Their average is approximately 1.93 seconds. That sounds strong. However, three experiences take 2.6 seconds or more, including one at 4.2 seconds. A team that reports only the average could overlook a template, location, or device segment producing poor experiences.
Percentiles reveal several useful parts of that same distribution:
- p25 shows a relatively fast experience.
- p50 shows the middle observation, or median.
- p75 shows the level experienced by most users while retaining attention on the slowest quarter.
- p95 exposes the point where the slowest 5% begin.
Google’s Web Vitals guidance distinguishes lab data from field data for this reason. Lighthouse can repeatedly test a page under controlled conditions, while field data reflects real devices, networks, locations, and browsing contexts. Field data is a distribution, not a single universal load time.
How to calculate a percentile
There is no one universal percentile formula. Software products can use different conventions, particularly when a requested percentile lies between two data points in a small sample. NIST documents multiple percentile methods, so two spreadsheets can return slightly different p75 results from the same 12 values.
A practical, transparent method is inclusive linear interpolation:
- Sort all observations from smallest to largest.
- Count the observations as n.
- Convert the percentile to a decimal, p. For p75, use 0.75.
- Calculate the location: L = 1 + (n - 1) × p.
- If L is not a whole number, interpolate between the two surrounding values.
Worked p75 formula example
Using the 12 LCP values above:
L = 1 + (12 - 1) × 0.75 = 9.25
Position 9 is 2.1 seconds and position 10 is 2.6 seconds. The decimal part of the position is 0.25, so the p75 result is:
2.1 + 0.25 × (2.6 - 2.1) = 2.225 seconds
Rounded for a report, p75 LCP is 2.23 seconds.
Using this same method, p25 is 1.25 seconds, p50 is 1.75 seconds, and p95 is approximately 3.54 seconds. For more examples of formulas, ranks, and reporting conventions, see percentiles explained.
The best practice is not to claim that one interpolation convention is universally correct. It is to identify the method, the sample size, the date range, and the rounding used in the report.
The most useful percentiles: p25, p50, p75, and p95
Different percentiles answer different operational questions. Using only one can create blind spots, particularly when a site has many page templates or widely varied traffic conditions.
p25: the faster quarter
The 25th percentile shows a value achieved by the faster quarter of observations. If p25 LCP is 1.25 seconds, one quarter of measured loads occur at or below that mark. It can identify what is possible under favorable conditions, but it is rarely an appropriate headline performance metric.
p50: the typical middle
The 50th percentile is the median. Half of observations are at or below it, and half are at or above it. In the 12-value example, p50 is 1.75 seconds. It is more resistant to extreme values than an average, but it can still understate problems affecting the slowest quarter of visitors.
p75: performance for most users
The 75th percentile is a useful management threshold because it focuses on the experience of most users without ignoring the slow tail. It is the convention used in Core Web Vitals reporting.
p95: the risky tail
The 95th percentile marks the level reached by all but the slowest or highest 5% of observations. A high p95 response time, error count, or redirect count can uncover a serious segment-level issue even when p75 is acceptable.
A site can therefore pass a p75 threshold while still having a p95 worth investigating. That is not a contradiction; it is evidence that the experience is uneven.
Why p75 matters in a website performance audit
As of November 2026, Google’s Core Web Vitals use the 75th percentile of page views to classify real-user experience. The current “good” thresholds are LCP of 2.5 seconds or less, INP of 200 milliseconds or less, and CLS of 0.1 or less. Google also defines poor thresholds of over 4.0 seconds for LCP, over 500 ms for INP, and over 0.25 for CLS.
That framework explains why an excellent one-off Lighthouse result is not the same as a passing Core Web Vitals assessment. Lighthouse is valuable for diagnosis, but it is lab data collected with predefined device and network settings. Search Console’s Core Web Vitals report uses real-world field data, groups similar URLs, and separates mobile and desktop reporting.
A practical interpretation process is:
- Check the p75 field result by device category.
- Identify whether the issue affects a URL group, such as product pages or category pages.
- Use lab testing to reproduce likely causes, such as a late-loading hero image or render-blocking script.
- Review p95 or segmented RUM data to understand the worst affected visitors.
- Recheck field data after sufficient collection time; CrUX-based data commonly reflects a rolling 28-day period.
For example, a category-page group with p75 LCP of 2.4 seconds may meet the good threshold, but a p95 of 5.1 seconds still suggests a meaningful slow tail. That could warrant investigation into image delivery, regional latency, consent banners, or third-party tags before conversion losses become more visible.
Using percentile data for SEO and crawl audits
A percentile is not limited to speed metrics. It can make a crawl of 500 or 5,000 URLs more useful by separating normal variation from the pages that need attention first.
Useful technical SEO distributions include:
- HTML response time across indexable URLs
- Crawl depth from a selected navigation root
- Number of redirect hops before a final 200 response
- Internal links received by each canonical URL
- JavaScript requests, transferred bytes, or image payload per page
- Accessibility errors detected per URL or template
Metric direction determines whether a high percentile is favorable. A high p95 for redirect hops is usually bad because it means some URLs require many redirects. A high p95 for relevant internal-link counts may be positive, but it can also reveal an uneven linking structure if p25 pages receive very few links.
For instance, suppose 500 indexable pages have a p50 crawl depth of 3, p75 of 5, and p95 of 9. The p95 result does not mean 95% of the site is nine clicks deep. It means only about 5% of pages are deeper than nine, and those pages deserve a targeted review for navigation, internal links, orphaning, or indexation value.
Audra’s local-first audits can preserve the relevant crawl context with the finding: URL, page type, response, links, performance checks, accessibility checks, and exportable evidence. That makes percentile summaries more actionable than a standalone dashboard number. Teams needing remediation steps can also use Audra’s website audit fix guides.
Applying percentiles to AI visibility data
AI-answer-engine visibility requires stricter definitions than many reports provide. A percentage can show the share of a fixed prompt set where a brand appears. A percentile can be useful only when the measurement and comparison set are numeric and explicit.
For example, an agency might test 50 commercial prompts across 10 brands and record the number of answers mentioning each brand. If a brand appears in 31 of 50 answers, its mention rate is 62%. That is a percentage, not a percentile.
Percentile-style reporting may be appropriate for measures such as citation position, competitor mention count, answer length, or visibility scores across a defined market sample. A claim such as “the brand is in the 95th percentile for AI visibility” needs all four of these details:
- The engine or engines tested and the collection dates.
- The exact prompt set, language, location, and device or account assumptions.
- The competitor or market dataset used for comparison.
- The scoring rule and whether higher values indicate better visibility.
Without those definitions, the statement sounds precise but cannot be independently interpreted. A repeatable audit should retain the prompt, answer evidence, observed brand mention, competitor result, and date—not just a final score.
Common percentile mistakes in reports
Percentiles are easy to calculate incorrectly and even easier to explain poorly. A credible audit avoids these recurring errors.
First, do not confuse a percentile with a percentage. “p75 LCP is 2.3 seconds” does not mean a page is 75% loaded.
Second, do not omit the comparison group. “This URL is in the 92nd percentile” means very little without stating whether the comparison is the site, a template group, a set of competitors, or another defined population.
Third, do not combine incompatible observations. Mixing mobile and desktop visits, logged-in and logged-out states, lightweight articles and script-heavy product pages, or multiple countries can produce a percentile that describes no meaningful user segment.
Fourth, do not imply false precision. A p75 of 2.225 seconds calculated from only 12 observations should usually be reported as 2.23 seconds, alongside the sample size and method. It is an estimate from a point in time, not a promise for every future visit.
Finally, do not treat Core Web Vitals as a ranking guarantee. Google recommends achieving good Core Web Vitals as part of good page experience, but its Search documentation is clear that strong metrics alone do not guarantee prominent rankings. Content usefulness, relevance, crawlability, and the broader site experience still matter.
A practical percentile reporting template
A compact table can give stakeholders the context to act without drowning them in raw crawl rows. The following example makes the dataset, direction, and next action visible.
| Metric | Dataset | p50 | p75 | p95 | Recommended action |
|---|---|---|---|---|---|
| Mobile LCP | 12 category-page measurements | 1.75 s | 2.23 s | 3.54 s | p75 is below 2.5 s; investigate the p95 slow tail |
| Redirect hops | 500 crawlable URLs | 0 | 1 | 3 | Review chains with three or more hops |
| Internal links | 500 indexable URLs | 14 | 31 | 86 | Review pages below p25 for linking opportunities |
| Accessibility errors | 500 audited pages | 2 | 7 | 24 | Prioritize templates driving the p95 count |
The table also makes a key point clear: p75 is not always the only percentile that matters. For user experience, Google’s p75 framework is highly useful. For crawl risk, outages, accessibility debt, and unusually heavy templates, p95 can provide the stronger prioritization signal.
For additional website-focused examples, Audra’s percentile definition guide explains how percentile ranks and cutoff values can be used in audit metrics.
FAQ
What is a percentile in simple terms?
A percentile shows where a value sits in a sorted group of values. The 75th percentile is the value at or below which roughly 75% of the data falls. For website performance, a p75 LCP of 2.5 seconds means about three quarters of measured page loads reached LCP in 2.5 seconds or less.
Is percentile the same as percentage?
No. A percentage is a share of a total, such as 18% of pages having missing titles. A percentile is a position in an ordered dataset, such as p75 response time being 600 milliseconds. Percentile rank describes how one score compares with others in a defined comparison group.
How do you calculate the 75th percentile?
Sort the values, choose a percentile method, and find the rank for 0.75 of the distribution. One common inclusive method uses L = 1 + (n - 1) × p. When the resulting position falls between observations, interpolate between them. Different spreadsheets and analytics tools may use different methods, especially for small samples.
Is a 95th percentile good or bad?
It depends on the metric direction. A low p95 is desirable for LCP, INP, response time, redirect hops, and accessibility errors because it means the problematic tail is controlled. A high p95 can be desirable for revenue, conversion value, test scores, or visibility scores where higher values are better.
Why does Google use the 75th percentile for Core Web Vitals?
Google uses p75 to evaluate performance for most real users rather than relying on an average or a single lab run. As of November 2026, the good thresholds are p75 LCP at or below 2.5 seconds, INP at or below 200 ms, and CLS at or below 0.1, assessed separately for mobile and desktop data.
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://web.dev/articles/defining-core-web-vitals-thresholds
- https://support.google.com/webmasters/answer/9205520?hl=en
- https://web.dev/articles/lab-and-field-data-differences
- https://web.dev/articles/vitals-tools
- https://web.dev/articles/crux-and-rum-differences