How to Improve INP: Fix Core Web Vitals Fast

INP replaced FID in March 2024 and it's the Core Web Vital most sites fail. Here's how to improve INP with real measurements: the web-vitals library, long-task splitting with scheduler.yield, third-party audits — plus LCP and CLS fixes in the same pass.

Of the three Core Web Vitals, INP is the fastest win left in 2026. It replaced First Input Delay in March 2024, and it's the metric most sites fail — because it measures *every* interaction across the page's life, not just the first one. The good news: learning how to improve INP is mostly mechanical. Kill long tasks, audit third-party scripts, measure real users — and the score follows. This guide does all of it with code, then folds LCP and CLS fixes into the same pass. What you'll learn The 2026 thresholds for LCP, INP, and CLS (and which data Google actually ranks on) Which measurement tool to use when (lab vs field decision table) How to measure Core Web Vitals on real users with the web-vitals library Why third-party scripts are the #1 INP killer — and how to audit them Splitting long tasks with scheduler.yield() (with a safe fallback) Fixing LCP (preload, priority, formats) and CLS (dimensions, fonts, reserved space) Three mistakes that waste performance effort Measure first: pick your tool Guessing is the enemy. Before changing anything, choose the right instrument — lab and field data answer different questions: Tool What it measures Best for Watch out Lighthouse (lab) Synthetic navigation; uses TBT as the INP proxy CI gating, regression checks Not real users; no real interactions in a default run DevTools Performance panel (lab) One traced session, Interactions lane Finding the exact long task behind a slow click One session ≠ field data web-vitals JS library (field/RUM) Real users, real interactions, full page lifetime Diagnosing production INP by page and interaction Needs real traffic; aggregate, don't eyeball single sessions PageSpeed Insights + CrUX (field) 28-day rolling real-user data, 75th percentile The number Google actually ranks…

Explore