Page speed used to be an engineering concern that surfaced in a quarterly report nobody read. That changed when it became simultaneously a ranking signal and a measurable driver of abandonment. A slow page now costs you twice: fewer people find it, and fewer of those who do stay.
What the metrics actually measure
Largest Contentful Paint (LCP) measures when the main content becomes visible. It is mostly a server-response and asset-delivery problem. Interaction to Next Paint (INP) measures how quickly the page responds when someone interacts with it, and is mostly a JavaScript problem. Cumulative Layout Shift (CLS) measures how much the layout jumps while loading, and is mostly a missing-dimensions problem.
Treating them as one number called speed is why so many optimisation efforts stall. They have different causes and different fixes.
The fixes that usually matter most
- Images. Serve modern formats, size them for the actual display dimensions, and set explicit width and height attributes. This routinely fixes LCP and CLS at once.
- Server response. Caching at the edge, and a hosting tier that is not oversubscribed, often beats weeks of front-end tuning.
- JavaScript weight. Every third-party script runs on the main thread. Audit what is loaded, remove what nobody uses, and defer what can wait.
- Fonts. Preload the ones you need, subset them, and set a sensible display strategy so text is readable while they load.
Measure real users, not just lab tools
Lab tests run on a fixed machine and a synthetic connection. Field data reflects the devices and networks your visitors actually have, which are usually slower than yours. When the two disagree, trust the field data — it is the one that corresponds to revenue.
Where to stop
There is a point of diminishing returns, and chasing a perfect score past it is not a good use of engineering time. Get into the good threshold for all three metrics on the pages that carry commercial weight — landing pages, product pages, checkout — then hold that position with monitoring rather than periodic heroics.
Performance is not a project you complete. It is a standard you maintain, and it is cheapest to maintain when it is part of the build rather than a remediation exercise afterwards.