How do you optimize Core Web Vitals on a production site?

Core Web Vitals are the three metrics Google uses for both UX and search ranking: LCP (perceived load speed), INP (interactivity), CLS (visual stability). Each measures a different user-felt experience and each has its own optimization playbook. Trying to optimize all three with one trick (e.g., "ship less JS") is necessary but not sufficient.

LCP — Largest Contentful Paint (target ≤ 2.5s)

The time at which the largest visible element (often the hero image or hero text block) is painted. Browsers expose it via PerformanceObserver({ type: 'largest-contentful-paint' }). Find it in Chrome DevTools → Performance → "LCP" marker, or in Lighthouse's report.

The four phases of an LCP measurement: (1) TTFB — server response time, (2) load delay — time from response to LCP resource request, (3) load time — fetching the LCP resource, (4) render delay — time to actually paint. Most LCP wins come from (1) and (2):

• Serve from a CDN with edge caching; aim for TTFB <400ms.
• Use modern image formats: AVIF (smallest), then WebP, with PNG/JPEG fallbacks. Use srcset + sizes for responsive serving.
• Mark the hero with <img fetchpriority="high"> or Next.js <Image priority> so the browser preloads it.
• Don't lazy-load above-the-fold images — that delays LCP.
• Inline critical CSS and defer the rest. Render-blocking CSS adds directly to LCP.
• Preconnect to your image / font CDNs; preload the LCP image (<link rel="preload" as="image">).
• Use self-hosted fonts with font-display: swap or optional; avoid late-arriving @font-face rules.
• Avoid huge client JS before paint — prefer SSR / streaming (Next.js App Router, Remix).
• Use <link rel="preload" as="fetch" fetchpriority="high"> for above-the-fold JSON data so the network request starts during the HTML download.

INP — Interaction to Next Paint (target ≤ 200ms)

Replaced FID in March 2024. Measures the worst (P98) interaction latency across the session: tap, click, key press → next paint. Unlike FID, it captures interactions after first input. This is the metric most apps fail.

The contributors: input delay (main thread busy), processing time (event handler + React render), presentation delay (paint).

Optimizations:

• Break up long tasks (>50ms blocks). Use scheduler.yield() (Chromium), setTimeout(_, 0), or requestIdleCallback to insert yield points. The isInputPending API tells you to bail out early.
• useTransition + useDeferredValue in React: keep input handlers cheap, defer expensive derived updates.
• Web Workers for CPU-heavy work (JSON parsing of large blobs, search indexing, image processing). The main thread should be a renderer, not a compute engine.
• Hydration cost: defer non-critical hydration with next/dynamic({ ssr: false }) or RSC. Less hydration = less main-thread work during early interactions.
• Event delegation instead of attaching listeners to thousands of rows.
• Debounce input handlers that derive expensive views.
• Avoid forcing layout in click handlers (any read of offsetWidth after a write triggers sync layout).
• Memoize heavy components so click handlers don't trigger giant re-renders.

CLS — Cumulative Layout Shift (target ≤ 0.1)

Sum of all unexpected layout shifts across the page lifetime. Each shift's impact fraction × distance fraction contributes; the worst session 5-second window scores.

Optimizations:

• Always set width/height (or CSS aspect-ratio) on <img> and <video> so the browser reserves space before the asset loads.
• Reserve space for ads, embeds, banners with fixed-height containers. Late-loading content pushing the page down is the #1 cause of CLS regressions.
• Don't insert content above the fold after load. Cookie banners, "subscribe to our newsletter" toasts, late personalization headers — all common CLS killers. Inject them at the bottom of the layout, or pre-allocate space.
• Font swaps. Use font-display: optional (no swap if not in 100ms) or preload the font so it arrives in time. A "swap" from a system font to a custom font with different metrics causes line-height shifts.
• CSS transforms instead of layout for animations (transform: translateY(...) doesn't shift).
• Skeletons with the same final dimensions as the loaded content.

Measurement. Use both:

• Lab data (Lighthouse, WebPageTest, DevTools throttled mobile) — controlled, repeatable, tells you what the metric could be.
• Field / RUM data (Chrome User Experience Report / web-vitals JS library reporting to your analytics) — tells you what real users experience. Field is what Google uses for ranking. Always optimize toward the P75.

Workflow. Read field data → identify which metric fails → run a lab profile of a slow user's journey → fix the worst contributor → re-measure in field for 28 days. Don't chase Lighthouse scores; chase the P75 field metric for each route.