Quick answer
The 75th percentile (p75, upper quartile or Q3) is the value that 75% of measurements are at or below; the other 25% are higher. Google's Core Web Vitals judge pages on p75: at least 75% of page views must meet each metric's good threshold.
Formula: p75 = value at sorted position (n − 1) × 0.75 + 1, interpolated (linear method); equivalent: at least 75% of values ≤ p75
What the 75th percentile means
The 75th percentile (written p75, also called the upper quartile or third quartile, Q3) is the value that 75% of your measurements are at or below. The other 25% are higher. If a page's p75 load time is 2.4 seconds, then three out of four visits loaded in 2.4 seconds or less, and one in four took longer.
It sits between the median (the 50th percentile, the typical visit) and the 95th or 99th percentile (the slow tail). That position is why performance people like it: it describes the experience of "most visitors, including some slow ones" without being dragged around by a handful of extreme values. The same number can be read two ways, and they are equivalent:
- "The p75 is at most the threshold."
- "At least 75% of visits are at or under the threshold."
How to calculate the 75th percentile: a 10-value recompute
Sample data: ten largest contentful paint (LCP) times, in milliseconds, for one page, already sorted: 1,200; 1,350; 1,500; 1,650; 1,900; 2,100; 2,300; 2,800; 3,900; 6,500.
With n = 10 values and p = 0.75 there is no value that sits exactly at the 75% mark, so every method has to pick a rule. Three common ones give three answers:
| Method | Rule | Position in the sorted list | p75 |
|---|---|---|---|
| Nearest rank | Take the value at rank ceil(n × p) | ceil(7.5) = 8th value | 2,800 ms |
Linear interpolation (default in NumPy; Excel PERCENTILE.INC) | Position (n − 1) × p + 1, then interpolate | 7.75th: 2,300 + 0.75 × (2,800 − 2,300) | 2,675 ms |
Weibull / NIST handbook (Excel PERCENTILE.EXC) | Position p × (n + 1), then interpolate | 8.25th: 2,800 + 0.25 × (3,900 − 2,800) | 3,075 ms |
Same ten numbers, three p75 values from 2,675 to 3,075 ms. With 10 samples the method matters; with thousands of samples the three converge to within a rounding error. The NIST/SEMATECH handbook writes the interpolation rule as p(N + 1) = k + d, then Y(p) = Y[k] + d × (Y[k + 1] − Y[k]). NumPy documents linear as its default method and lists the others by name.
For comparison on the same data: the mean is 2,520 ms, the median is 2,000 ms, p90 is 4,160 ms and p95 is 5,330 ms (linear method).
Why Core Web Vitals use the 75th percentile
Google's Core Web Vitals judge a page or site on its 75th percentile of real page loads, measured separately for mobile and desktop. A metric counts as "good" when at least 75% of page views meet that metric's good threshold. The thresholds are 2,500 ms or less for LCP, 200 ms or less for interaction to next paint (INP) and 0.1 or less for cumulative layout shift (CLS).
web.dev explains the choice: the percentile has to make sure a majority of visits had the target experience, and the 75th percentile is less likely than higher percentiles to be swung by outliers. In plain terms, p75 is demanding enough that a site cannot pass by serving only its fastest half, but not so extreme that one flaky connection decides the result.
Apply the rule to the sample page. Seven of the ten loads are at or under 2,500 ms, so 70% are good, below the 75% bar. Every p75 in the table above (2,675 to 3,075 ms) is also above 2,500 ms. The two views agree: the page fails LCP, whichever interpolation rule you pick. Sample data, recomputed in Python.
What p75 hides and what it keeps
Now replace the slowest load, 6,500 ms, with a 20,000 ms timeout and recompute.
| Statistic | Original sample | With a 20,000 ms outlier |
|---|---|---|
| Mean | 2,520 ms | 3,870 ms |
| Median (p50) | 2,000 ms | 2,000 ms |
| p75 (linear) | 2,675 ms | 2,675 ms |
| p95 (linear) | 5,330 ms | 12,755 ms |
One extreme value moved the mean by more than 1.3 seconds and p95 by more than 7 seconds, and left the median and p75 untouched. That is the practical case for p75: it ignores the worst quarter. It also means p75 will not warn you about that worst quarter. If the slowest 5% of visits are failing completely (a payment page that times out for some regions), p75 stays green while p95 or an error rate turns red. Use p75 as the pass or fail line, and look at p95 and errors to see the tail.
Fix the slow quarter, and watch p75 move
p75 only improves when the loads at and above the 75% mark get faster. Speeding up your fastest loads does nothing to it.
Sample data: suppose two fixes (smaller hero image, preloaded font) pull the 8th, 9th and 10th loads down to 2,400; 2,450 and 3,000 ms. The sorted list becomes 1,200; 1,350; 1,500; 1,650; 1,900; 2,100; 2,300; 2,400; 2,450; 3,000. The linear p75 falls from 2,675 to 2,375 ms (7.75th position: 2,300 + 0.75 × 100 = 2,375), and 9 of 10 loads are now at or under 2,500 ms, so 90% are good. The page now passes, and the pass came from moving three slow loads; the median did not change at all.
A decision rule you can use: if p75 is above the threshold, find what the slowest quarter of visits have in common (device, country, page, browser) before changing anything for everyone.
Common mistakes with percentiles
- Averaging percentiles. The p75 of two days, two pages or two countries is not the average of their p75 values. Recompute from the combined samples, or compare the groups side by side.
- Mixing mobile and desktop. Core Web Vitals are assessed per device type. A fast desktop majority can hide a failing mobile audience in a pooled p75.
- Trusting small samples. With 10 loads, one value moves p75 by hundreds of milliseconds (see the three methods above). Judge a page on hundreds of loads, not dozens.
- Reading p75 as "75% of users". It is a share of visits or page loads, not of people. One person with a slow phone contributes many slow loads.
- Treating "p75" as a lab number. Lab tests run one load on one machine; a percentile only means something over many real visits.
Reading p75 in TapCub
Take the sample page above: you want to know whether it passes LCP and where its slow quarter comes from.
- Open Monitoring › Performance. The five cards (LCP, CLS, INP, first contentful paint, time to first byte) are measured in real visitors' browsers, not in a lab run. Switch the card statistic among P95, P75, median and average, and read the P75 first.
- Compare it with the good and poor thresholds shown for each metric (LCP: good at most 2.5 seconds, poor above 4 seconds).
- Use the Slowest pages list, which is sorted by P75 largest contentful paint, to pick the page to fix first.
- Break the trend down by page, country, device, browser or operating system, and set the percentile to P50, P75, P90 or P99 to see whether the slow quarter is concentrated in one segment.
- Check whether those slow pages also lose people with the funnels and pages reports before spending engineering time.
Limits to keep in mind: the numbers come from visitors' browsers, so blocked scripts or an old snippet leave a page empty (server time is not substituted), and TapCub notes that percentiles jump when a page has few samples. For a step-by-step walkthrough, see check Core Web Vitals; for the metrics themselves, see Largest Contentful Paint and the general definition of a percentile.
Sources
- web.dev, How the Core Web Vitals metrics thresholds were defined (published May 21, 2020; updated May 7, 2025).
- web.dev, Web Vitals (75th percentile, segmented by mobile and desktop).
- NIST/SEMATECH e-Handbook of Statistical Methods, Percentiles.
- NumPy documentation, numpy.percentile (default
linearmethod).
Was this explanation clear?
We use your answers to fix and expand this page.
Couldn’t send. Please try again in a minute.
Thanks for the feedback. We’ll use it to improve this page.You’ve already sent feedback on this page. Thanks.

Data Analyst
Explores traffic sources, conversion funnels, and retention, using clear metric definitions and concrete examples to explain what the numbers mean.