Skip to main content

Web QoE Scoring Model

Note

This model is in beta and under evaluation. We welcome feedback on the scoring methodology and curve parameters.

The Web QoE Score combines web performance metrics into a page-load quality score on a 0–100 scale. It is provided as the statistic value web_qoe_score.

By default, the score is calculated from three metrics: First Contentful Paint (FCP), Largest Contentful Paint (LCP), and Time to First Byte (TTFB). These are the metrics that can be reliably measured in automated tests, which is how Surfmeter collects web performance data. Other metric combinations are supported as well, and are described under alternative metric sets below.

How it works​

Each web performance metric is scored individually on a 0–100 scale using a log-normal scoring curve. The individual scores are then combined into a single weighted average — the Web QoE Score.

The S-shaped scoring curve maps metric values, such as a first paint time of 1.2 seconds, to scores. Fast values score close to 100 and slow values close to 0. Two parameters define each curve:

  • p10 — the metric value at which the score reaches approximately 90 (the "good" threshold). Values better than this produce diminishing returns in the score.
  • median — the metric value at which the score is approximately 50. This represents the boundary between acceptable and poor performance.

This approach is based on Google's Lighthouse performance scoring methodology, which describes the process as follows:

The Lighthouse scoring curve model uses HTTPArchive data to determine two control points that then set the shape of a log-normal curve. The 25th percentile of HTTPArchive data becomes a score of 50 (the median control point), and the 8th percentile becomes a score of 90 (the good/green control point). Between 0.50 and 0.92, there's a near-linear relationship between metric value and score. Around a score of 0.96 is the "point of diminishing returns" as higher, the curve pulls away, requiring increasingly more metric improvement to improve an already high score.

— Lighthouse Performance Scoring, Google Chrome Developers

Google also provides a Lighthouse scoring calculator and a graphing function for exploring the log-normal curve behavior.

Score interpretation​

The Web QoE Score can be interpreted using a simple traffic-light scheme:

Score rangeRatingMeaning
90–100🟢 GoodThe page loaded quickly. Users are unlikely to notice any performance issues.
50–89🟡 Needs improvementThe page load was noticeably slow. Users may be mildly frustrated.
0–49🔴 PoorThe page load was significantly below expectations. Users are likely to abandon the page or have a bad experience.

These thresholds follow the same 90/50 boundaries used by Lighthouse for its performance score.

Input metrics​

The default metric set uses the following three metrics:

TTFB is partly contained in the other two metrics — a slow server delays both the first and the largest paint. It is still scored separately, because it isolates the network and server part of the page load from the rendering part. A page that renders slowly because of heavy JavaScript and a page that renders slowly because the connection is bad both show poor FCP and LCP values, but only the second one shows a poor TTFB. In this sense, TTFB acts as a proxy for network conditions, which makes it valuable for network-focused monitoring.

The default set excludes two further Web Vitals metrics:

  • INP – Interaction to Next Paint: measures responsiveness to user interactions. It requires a real user clicking, tapping, or typing, and cannot be measured in automated tests at all. Measuring it requires Real User Monitoring (RUM).
  • CLS – Cumulative Layout Shift: measures visual stability during loading. Many measurement environments do not report it, and when they do, it is frequently zero, which contributes little information to the score.

Where these two metrics can be measured, they are still recorded and reported as statistic values of their own — they are simply not part of the score. Keeping them out of it also keeps scores comparable: because a CLS of 0 is a valid measurement rather than a missing one, letting CLS into the calculation would mean that the same page scores differently depending on whether the browser happened to report a layout shift at all.

Total Blocking Time (TBT) is not used either, as it is not directly measurable in all environments. Similarly, the Speed Index (SI) is not used, as it is difficult to measure.

Note

The default score reflects loading performance only. It cannot capture interactivity or visual stability problems: a page with fast paint times but sluggish responsiveness or jumping content will still score well.

Scoring curves​

The curves for the default metric set are aligned to the Web Vitals thresholds: the p10 value of each curve is the Web Vitals "good" threshold, and the median is the "poor" threshold.

Metricmedian (= "poor" threshold)p10 (= "good" threshold)Weight
FCP3000 ms1800 ms0.20
LCP4000 ms2500 ms0.50
TTFB1800 ms800 ms0.30

The weights are the relative Lighthouse weights of these three metrics (0.10, 0.25, and 0.15), renormalized to sum to 1. LCP carries the most weight, as it is the metric that best represents when the page becomes useful to the user.

These curves are identical for PC and mobile device types, since the Web Vitals thresholds do not distinguish between the two.

Aligning the curves to the Web Vitals thresholds means that a page which just meets all "good" thresholds scores 90, and a page at the "poor" thresholds scores 50 — matching the score interpretation above. The Lighthouse curves are not used here, because they are calibrated for a set in which CLS and INP — which tend to score high — compensate for the much stricter paint-time curves. Applying the Lighthouse PC curves to FCP, LCP, and TTFB alone would give a page that meets every Web Vitals "good" threshold a score of only 58, which is not a useful result.

Effect of TTFB on the score​

The following table illustrates how much the server response time moves the overall score. FCP and LCP are held at their "good" thresholds (1.8 s and 2.5 s), and only TTFB varies:

TTFBTTFB metric scoreWeb QoE Score
0.2 s10093
0.8 s9090
1.2 s7485
1.8 s5078
2.5 s3072
4.0 s1066

Example scores​

The following table shows scores for the PC device type at the Web Vitals tier boundaries:

ScenarioFCPLCPTTFBWeb QoE ScoreRating
Excellent0.5 s1.0 s0.2 s100Good
Good (Web Vitals "good" thresholds)1.8 s2.5 s0.8 s90Good
Needs improvement (Web Vitals "poor" thresholds)3.0 s4.0 s1.8 s50Needs improvement
Poor5.0 s6.0 s3.0 s15Poor

And here are scores for a set of typical websites:

ScenarioFCPLCPTTFBWeb QoE ScoreRating
Fast CDN-served landing page0.6 s0.9 s0.15 s100Good
Well-optimized news site1.2 s2.0 s0.4 s98Good
Typical corporate website2.0 s3.0 s0.9 s82Needs improvement
Heavy SPA (React/Angular)2.8 s3.5 s0.6 s72Needs improvement
Ad-heavy media site3.5 s5.0 s1.2 s43Poor
Slow shared hosting blog4.5 s7.0 s2.5 s15Poor
Overloaded e-commerce site5.5 s8.0 s3.5 s7Poor
Broken/failing site12 s18 s6.0 s1Poor

The heavy SPA is a good illustration of what the score expresses: its server responds quickly (TTFB 0.6 s, scoring 96 on its own), but the page takes a long time to render. Its overall score is still in the middle of the "needs improvement" range.

Alternative metric sets​

Besides the default set, the scoring model supports a few other metric combinations, for platforms that provide different data. Surfmeter measurements do not use them: their score is always calculated from FCP, LCP, and TTFB, even when CLS was measured. The sets below therefore only come into play when the scoring function is called directly with other metrics, and any combination not listed here produces no score at all.

FCP and LCP only​

When TTFB is not available, the score is calculated from FCP and LCP alone, using the same Web Vitals threshold-aligned curves with equal weights of 0.5 each. At the tier boundaries this produces the same scores as the default set (90 at "good", 50 at "poor"). The two sets are therefore broadly comparable. Away from the boundaries, results differ where server response time and rendering time diverge — the heavy SPA above scores 72 with TTFB included and 61 without it.

Sets including CLS and INP​

Three further sets exist for platforms that report CLS, and optionally INP. These use the original Lighthouse v10 curves and weights:

MetricPC medianPC p10Mobile medianMobile p10Weight
FCP1600 ms934 ms3000 ms1800 ms0.10
LCP2400 ms1200 ms4000 ms2500 ms0.25
CLS0.250.10.250.10.25
INP500 ms200 ms500 ms200 ms0.25
TTFB1800 ms800 ms1800 ms800 ms0.15

The three combinations are:

  • Full (5 metrics): FCP, LCP, CLS, INP, and TTFB. Requires Real User Monitoring data for INP.
  • 4 metrics without INP: FCP, LCP, CLS, and TTFB, as produced by Lighthouse in lab mode. INP's weight of 0.25 is redistributed proportionally among the remaining four, giving effective weights of approximately FCP 0.13, LCP 0.33, CLS 0.33, TTFB 0.20.
  • 4 metrics without TTFB: FCP, LCP, CLS, and INP. TTFB's weight of 0.15 is redistributed the same way, giving approximately FCP 0.12, LCP 0.29, CLS 0.29, INP 0.29.

Unlike the default set, these sets use more lenient FCP and LCP curves for mobile than for PC, reflecting the typically slower rendering on mobile devices. The same page therefore scores higher when measured as a mobile device: the typical corporate website above scores 66 as PC and 83 as mobile in the full set.

Warning

Scores from these sets are not directly comparable to the default one. Because the Lighthouse curves are much stricter for FCP and LCP, a page that meets all Web Vitals "good" thresholds scores 90 in the default set but only 74 in the full set — a page rated 🟢 Good by one set is rated 🟡 Needs improvement by the other. Compare scores only within the same metric set.

The following table shows the same websites as above, scored with every available metric set (PC, with CLS and INP values added):

ScenarioCLSINPDefault (3)FCP + LCPFull (5)4 (no INP)4 (no TTFB)
Fast CDN-served landing page00.03 s100100999999
Well-optimized news site0.050.12 s9898878485
Typical corporate website0.120.25 s8281666163
Heavy SPA (React/Angular)0.030.40 s7261626256
Ad-heavy media site0.350.80 s4331282920
Slow shared hosting blog0.200.60 s1511312831
Overloaded e-commerce site0.451.20 s75101010
Broken/failing site1.503.00 s10110

The slow shared hosting blog shows why the sets are not interchangeable: its paint times are bad, but its CLS and INP values are acceptable. The Lighthouse-based sets rate it around 30, while the default set — which only looks at loading performance — rates it at 15.