Capability: Core Web Vitals trend (CrUX)

Understanding how a site’s performance evolves over time is essential for maintaining a fast, reliable user experience. The browser telemetry dataset that aggregates real‑world measurements from millions of visitors provides a longitudinal view of the three Core Web Vitals metrics—Largest Contentful Paint (LCP), Interaction to Next Paint (INP), and Cumulative Layout Shift (CLS). This article explains the capability triggered by the prompt “Show my Core Web Vitals trend”, covering what it is, when to use it, where it runs, how it works, and answers common questions.

The Core Web Vitals trend capability extracts historical data from the browser telemetry dataset for a given origin or URL pattern and presents the monthly (or weekly) evolution of LCP, INP, and CLS. Rather than a single snapshot, the output shows how each metric has shifted across time, enabling teams to spot regressions, validate optimizations, and align performance goals with actual user conditions.

Key elements of the output:

Element	Description	Typical unit
LCP	Time from page start to the largest visible element rendering	seconds (s)
INP	Latency of the most delayed user interaction (click, tap, keypress)	milliseconds (ms)
CLS	Sum of all unexpected layout shift scores during the page lifespan	unitless (ratio)
Period	Aggregation window (monthly by default,可调 to weekly)	–
Sample size	Number of page views contributing to each period’s percentile	–
Percentiles	75th‑percentile (the value Google’s search ranking uses) and optionally 50th/95th	–

The capability does not rely on synthetic lab tests; it reflects the actual conditions experienced by visitors across devices, network types, and geographies.

When to use it

Performance regression detection

When a new feature release coincides with a noticeable dip in the 75th‑percentile LCP or a rise in INP, the trend view makes the regression obvious. Teams can correlate the timing with deploy logs, third‑party script updates, or changes in image delivery pipelines.

Validation of optimizations

After implementing lazy‑loading, server‑side rendering, or font‑display adjustments, the trend line should show a downward movement in LCP or CLS. Seeing the metric improve over successive periods provides concrete evidence that the change benefited real users, not just a lab environment.

Setting realistic performance budgets

Because the dataset reflects the actual distribution of user conditions, the observed 75th‑percentile values become a data‑driven baseline for budgeting. For example, if the median LCP over the last three months is 2.4 s, a team might set a target of ≤2.2 s for the next quarter, knowing that the current user base already experiences better than the median.

Communicating impact to stakeholders

Executive reports benefit from a simple line chart that shows “performance over time” rather than a barrage of raw numbers. The trend capability supplies the data needed for such visualizations, making it easier to justify continued investment in performance work.

Cross‑team alignment

SEO, UX, and infrastructure groups often speak different languages. A shared trend chart that displays the same Core Web Vitals numbers creates a common reference point, reducing friction when prioritizing work.

Where does it run

The capability is executed within our specialized AI orchestration platform, which pulls the required slices from the browser telemetry dataset on demand. The underlying data resides in a publicly accessible, privacy‑safe aggregate that strips individual user identifiers while preserving metric distributions.

Because the dataset is updated daily, the trend can be refreshed as frequently as the user desires—monthly for most reporting cadences, or weekly for fast‑moving product teams. The processing occurs in a secure, isolated compute environment; no raw user‑level data leaves the aggregation service, ensuring compliance with privacy regulations such as GDPR and CCPA.

How it works

Below is a step‑by‑step outline of the internal workflow that turns the prompt “Show my Core Web Vitals trend” into a ready‑to‑publish chart.

1. Input parsing

• The prompt is interpreted to extract the target scope (origin, URL pattern, or specific page).
• Optional parameters such as aggregation window (monthly/weekly), desired percentiles, and date range are identified; defaults are applied when omitted.

1. Data retrieval

• Queries are sent to the browser telemetry dataset’s API endpoint for the specified origin.
• The request asks for the monthly (or weekly) 75th‑percentile values of LCP, INP, and CLS, plus the sample size (number of page views) for each period.
• The dataset returns a time‑ordered series, typically covering the last 12–24 months, depending on data availability for the origin.

1. Data validation & cleaning

• Periods with insufficient sample size (e.g., < 100 page views) are flagged and either omitted or annotated with a warning icon, preventing over‑interpretation of noisy data.
• Outlier spikes caused by known events (e.g., a major network outage) are optionally smoothed using a rolling median if the user elects to enable noise reduction.

1. Metric transformation

• Raw values are kept in their native units (seconds for LCP, milliseconds for INP, unitless for CLS).
• For reporting convenience, the system can also compute derived indicators such as “percentage change month‑over‑month” or “trend slope via linear regression.”

1. Visualization preparation

• A multi‑line chart is generated, with the X‑axis representing time and the Y‑axis showing metric values.
• Each Core Web Vital receives its own line, colored consistently (e.g., blue for LCP, orange for INP, green for CLS).
• Shaded bands depict the 50th–95th percentile range when requested, giving a sense of variability.

1. Output delivery

• The chart is rendered as an SVG or PNG embedded in the response, accompanied by a downloadable CSV of the underlying numbers.
• A brief textual summary highlights the most significant observation (e.g., “LCP improved from 2.9 s to 2.3 s over the last six months, a 21 % reduction”).

1. Feedback loop

• Users can annotate specific points on the chart (e.g., “launch of new image CDN”) to create a shared narrative that persists across sessions.

First‑hand observation

In a recent internal test, we applied the capability to a media‑rich news site covering the period January – June 2024. The 75th‑percentile LCP dropped from 2.8 s in January to 2.2 s in June after the team enabled aggressive Brotli compression and switched to a next‑generation image format. Simultaneously, INP remained stable around 45 ms, indicating that the main thread was not blocked by the new asset pipeline. The trend view made it straightforward to present these results to the editorial leadership, who then approved a budget for further core‑web‑vitals work.

FAQ

Q1: Does the trend data include mobile and desktop separately? A: Yes. The browser telemetry dataset breaks down metrics by form factor. When requesting the trend, you can specify “phone”, “desktop”, or “all devices”. The output will show separate lines or a combined view depending on the selection.

Q2: How far back does the data go? A: Coverage varies by origin. For sites that have received a measurable amount of traffic for at least six months, the dataset typically provides 12–24 months of history. Newer origins may show fewer months until sufficient sample size accumulates.

Q3: What if my site has low traffic and the sample size is small? A: The system flags periods where the sample size falls below a configurable threshold (default 100 page views). In those cases, the point is displayed with a lighter opacity and a tooltip warns of higher uncertainty. You can also widen the date range or switch to weekly aggregation to increase the sample per bucket.

Q4: Can I compare two origins side‑by‑side? A: Absolutely. By submitting two separate prompts—or a single prompt with a comma‑separated list of origins—the platform returns aligned time series, enabling a direct visual comparison (e.g., staging vs. production).

Q5: Is the data affected by privacy thresholds? A: The browser telemetry dataset applies differential privacy techniques to ensure that no individual user can be re‑identified. This may cause slight smoothing at the extremes of the distribution, but the 75th‑percentile—which is used for search ranking considerations—remains stable and comparable across periods.

Q6: How does this differ from synthetic lab tools like Lighthouse? A: Lab tools measure performance under a controlled environment (specific device, network throttling, no variance). The trend capability reflects the actual distribution of conditions experienced by real visitors, capturing influences such as varying connection speeds, device capabilities, and geographic location. Both perspectives are valuable; synthetic data is ideal for debugging, while field data reveals true user impact.

Q7: Are there any costs associated with running the trend query? A: Computation is performed on-demand within our specialized AI orchestration environment. Costs are calculated dynamically based on the complexity of the request (date range, number of origins, selected percentiles) and are reflected in the usage metrics of your account. No per‑query token pricing is disclosed.

Q8: Can I export the raw numbers for internal reporting? A: Yes. Each trend response includes a downloadable CSV file containing the timestamp, metric values, sample size, and any applied flags (e.g., low‑sample warning).

Takeaway

The Core Web Vitals trend capability transforms raw, privacy‑safe field measurements into a clear, time‑based narrative of how real users experience your site’s speed, responsiveness, and visual stability. By turning the prompt “Show my Core Web Vitals trend” into a regularly refreshed chart, teams gain an objective baseline for detecting regressions, validating optimizations, and setting performance goals that reflect actual user conditions—without the guesswork inherent in lab‑only testing. Use it to align engineering, product, and leadership around a shared, data‑driven view of performance health.