← Back to blog

CrUX History Is a Rolling Window, Not a Weekly Snapshot

CrUX History updates weekly, but each point summarizes 28 days. Read overlapping windows, p75 values, and missing periods correctly before calling a field regression.

A CrUX History chart is not a sequence of independent weekly measurements. Each point summarizes a rolling 28-day window, so treating adjacent points as separate weeks can make ordinary variation look like a regression—or hide when a change actually reached users.

What one point represents

The Chrome UX Report (CrUX) History API returns time series for real-user experience metrics. It updates on Mondays, with data through the previous Saturday. Each point covers a 28-day collection period, and there is one period per week. Consequently, two adjacent points share 21 days of observations. A month of chart points is not a month of disjoint samples.

That overlap affects how a release appears. If you deploy a change on Tuesday, the next point still includes three weeks from before the release and only a few days after it. The pre-release experiences then leave the window gradually, rather than disappearing at the next weekly update. A curve that improves slowly may reflect this windowing, not a gradual rollout or a delayed fix.

Use the API's collectionPeriods dates to label the intervals behind values. A displayed date by itself is ambiguous: it may look like a measurement taken that day, even though the value summarizes four weeks. Keep the start and end dates visible in internal charts and incident notes.

Treat p75 as a distribution summary

CrUX reports distributions, including the 75th percentile (p75): the value at or below which 75% of recorded experiences fall in that collection. It is not the average experience, the worst case, or a score assigned to a particular visitor. The History API also notes that percentile values are derived from the underlying data and need not be values any one user actually experienced.

Do not average p75 values across weeks to estimate a monthly p75. Percentiles cannot be combined that way: the percentile of a combined population depends on the underlying observations, not just each window's p75. Likewise, don't average a desktop p75 and a phone p75 and call it the site's overall p75. Compare like-for-like dimensions, or use the distribution the reporting system provides.

A small p75 change is a signal to investigate, not proof of a statistically significant regression. CrUX deliberately does not expose exact traffic counts, and the points overlap. Look at multiple distribution indicators, the size and persistence of the movement, and whether it aligns with a code or infrastructure change. Avoid inventing a confidence interval from a chart that does not contain the sample counts needed to calculate one.

Separate performance change from audience mix

An origin-level record aggregates eligible experiences across pages on that origin; a URL-level record covers a specific page. Those answer different questions. A site-wide p75 can move because the traffic mix shifted toward a slower page, even if the page-level metrics did not worsen.

The same applies to device and geography dimensions. A greater share of phone traffic, or a change in which countries contribute eligible experiences, can move an overall distribution. When diagnosing a release, compare the same URL or origin, metric, form factor, and collection periods. Then inspect other segments to see whether the aggregate movement is a broad regression or a change in who or what is represented.

Coverage changes at the dataset level too: the August 2026 CrUX monthly release listed 18,294,881 origins, up 1.3% from the month before. That does not show that any one site improved; it means a cross-origin aggregate is not a fixed panel.

Missing data is not a passing result

CrUX has eligibility criteria for both users and pages. Pages must be publicly discoverable and sufficiently popular; Google does not publish an exact minimum visitor count. The dataset may also omit a page or origin when eligible data is insufficient for a particular collection period. In the History API, a period with no eligible data is represented by null percentile values and NaN histogram densities.

So “no CrUX result” does not mean “fast,” “slow,” or “unchanged.” It means the dataset cannot provide a result for that identifier and period. CrUX also represents eligible Chrome experiences, not every visitor to your site. If you use first-party real-user monitoring alongside it, document the population and aggregation window separately; the two datasets are complementary, not interchangeable.

A safer trend-reading routine

For every comparison, record the identifier (origin or URL), metric, form factor, and exact start and end dates. Compare a period with prior periods that use the same dimensions, and remember that adjacent periods share most of their observations. Treat a short-lived wobble as a prompt to inspect the distribution and deployment timeline, not as a verdict.

Use field data to answer what eligible visitors experienced, and a controlled lab run to investigate why. A lab result is a reproducible test under chosen conditions; it does not replace the field distribution. The distinction matters when a chart is used in a release review: one explains the observed population, the other helps isolate a cause.

Rolling aggregates are useful precisely because they smooth noisy field experiences. Read their dates, distributions, and eligibility limits as part of the metric—not as footnotes—and the chart becomes evidence about users rather than a weekly scorecard.

Sources

Comments

Loading comments…