Lighthouse vs RUM: how do you measure performance properly?

Question

IT Interview Practice · Accepted Answer

Measure with **both** lab and field data — they answer different questions. **Lighthouse** is lab (synthetic, reproducible, for debugging); **RUM** is field (real users, for truth). If you only trust one, you optimize the wrong thing.

## Lab vs field

| | Lab (Lighthouse) | Field / RUM |
|--|------------------|-------------|
| Source | One synthetic run, fixed device/network | Real visits, all devices/networks |
| Good for | Debugging, CI regression gates, repeatable | What users *actually* experience |
| INP | Can't measure (no real interactions) | Measured across the whole visit |
| Weakness | Doesn't reflect your real traffic | Noisy, needs volume, harder to debug |

Lighthouse runs a single throttled load on a clean machine — great for *reproducing* an issue, useless for knowing your p75. **INP especially** can only be captured in the field, because it needs real interactions over a session.

## Collecting field data with the web-vitals library

```js
import { onLCP, onCLS, onINP } from "web-vitals";

function send(metric) {
  // Attribution build tells you WHICH element/interaction caused it
  navigator.sendBeacon("/rum", JSON.stringify({
    name: metric.name, value: metric.value, id: metric.id,
  }));
}
onLCP(send); onCLS(send); onINP(send); // report each metric as it finalizes
```

This is the same methodology as Google's **CrUX** (Chrome User Experience Report), which powers Search ranking — CrUX is real-user p75, not lab.

## How to use them together

- **RUM** to detect real regressions and know your p75 (what ranking uses).
- **Lighthouse in CI** as a guardrail to catch regressions *before* they ship.
- Segment RUM by device/country/connection — an averaged number hides that low-end Android is your slow tail.

## Why it matters

Senior signal is knowing lab ≠ field, that ranking uses field p75, and that INP is only real in the field. Optimizing a Lighthouse score while real p75 stays poor is a classic trap this question is designed to expose.

Source	One synthetic run, fixed device/network	Real visits, all devices/networks
Good for	Debugging, CI regression gates, repeatable	What users actually experience
INP	Can't measure (no real interactions)	Measured across the whole visit
Weakness	Doesn't reflect your real traffic	Noisy, needs volume, harder to debug

Lab vs field

Collecting field data with the web-vitals library

How to use them together

Why it matters

Related questions