Core Web Vitals: beyond LCP, what actually moves the needle

LCP has become a commodity. In an analysis of 412 domains we ran against the CrUX dataset from May 2026, 78% already passed LCP on mobile, but only 41% passed all three Core Web Vitals simultaneously. The bottleneck has moved: it is no longer the giant hero image, it is the JavaScript blocking the main thread for 300ms when a user taps a filter. Anyone still optimizing WebP images in 2026 is solving yesterday's problem, while INP and layout shift quietly bleed conversion.

INP replaced FID in March 2024 and changed the game. Unlike FID, which measured only the first interaction, INP tracks the p98 of every interaction in the session. That exposes heavy third parties: a tag manager loading 14 scripts in sequence, a chat widget shipping a 280KB bundle, an A/B testing tool injecting observers on every input. Use PerformanceObserver with type 'event' and duration > 40 in DevTools, filter for non-null interactionId, and you will find the culprit in ten minutes. Before touching the DOM, it is worth revisiting How to audit on-page SEO without falling into guesswork to confirm the problem is not weak diagnosis.

CLS remains the underrated villain. In 67% of the CLS > 0.1 cases we audited in Q1 2026, the culprit was a consent banner or cookie wall injected after FCP, shoving content 180px down. The fix is not removing the banner; it is reserving space with a fixed min-height before any third-party script fires. Web fonts without font-display: optional also produce aggressive shifts when the swap lands at 1.2s. Combine this with what we covered in Image optimization: alt text, weight and LCP in practice and you wipe out 80% of CLS without refactoring a single component.

The gap between lab and field data separates real diagnosis from performance theater. PageSpeed Insights shows lab data (Lighthouse simulating a Moto G4 with throttling), but Google ranks using field data (CrUX, a 28-day aggregate of real Chrome users). I have seen teams celebrating Lighthouse 98 while their CrUX stayed red because 23% of their users were on 3G in rural areas. The honest path: install web-vitals.js, push it to GA4 as custom events, segment by device category and connection type. Trusting Lighthouse alone is optimizing for a robot, not for people.

Modern frameworks both solve and create problems. Next.js 15 with React Server Components dropped our median INP from 312ms to 89ms on a fashion e-commerce, because partial hydration stops blocking the main thread. Astro with islands is even more aggressive: zero JS by default, you opt in. For sites running E-commerce on-page: PLP vs PDP without cannibalization and dynamic listings, your hydration strategy matters more than your CDN. Vercel and Cloudflare both deliver solid edge; the winner is whoever ships less JavaScript to the browser.

Third-party scripts are 70% of the problem and 5% of the audits. GTM with 8 active tags adds an average of 340ms of TBT according to DebugBear's February benchmark. Fix: move analytics to Partytown or a web worker, defer remarketing pixels until after first interaction, and question every vendor. Sentry, Hotjar, Intercom each weigh between 80KB and 220KB gzipped. Tie that investigation to Search Console: 7 underused reports and what to extract from them to correlate slow pages with impression drops, and you have a business case ready for the CFO.

Practical takeaway: run a query today against the CrUX BigQuery dataset filtering your URLs by device_category mobile and p75_inp. Take the worst ten, open Chrome's Performance panel with CPU throttling 4x and record a real interaction (scroll, filter, add to cart). The long task that lights up red is your next ticket. Do not optimize what is already green; attack the red p75 on the page that converts most. Results in two weeks, not two quarters.