Ryan Whitcomb

I start by separating two things that get reported as one. Field data comes from real visits and describes what people got. Lab data comes from a synthetic load and describes what a controlled run produced. They answer different questions and disagree often, so every measurement here is labelled. Three metrics describe most of the perceived experience. Largest Contentful Paint is when the main content appeared, Interaction to Next Paint is whether the page responded when touched, and Cumulative Layout Shift is whether things moved while it loaded. The thresholds are percentiles over real users, which means a good average hides a bad tail and the number to watch is the one at the 75th percentile. The causes are narrower than the tool output suggests. A slow LCP is usually an image or a font, and usually because it was not prioritised. A poor INP is usually a long task on the main thread, often a hydration step. A CLS is usually an image or an advertisement without reserved space. The three fixes that cover most cases are reserving space for anything that loads late, deferring work that is not needed for the first view, and serving images at the size they are displayed. I write them up in that order because they are in that order of return. It is worth saying that a third-party script is frequently the largest single cause, and that removing one usually beats optimising around it. A measurement without a percentile attached is not a measurement, and a target without a date attached is not a plan.