Field data, not lab scores
A Lighthouse run on a developer laptop over office broadband tells you about that laptop. Your users are on mid-range Android devices on congested mobile networks, and the gap between the two is where the lost revenue lives.
Start with real-user data segmented by device, country and template. It almost always reveals that one template on one device class is responsible for most of the damage.
Three causes, most of the time
Across dozens of audits the same three causes dominate: images shipped at the wrong size and format, third-party scripts loaded synchronously, and a server response time nobody has measured since launch.
Fixing those three properly beats a long list of micro-optimisations, and it is work you can explain to a non-technical stakeholder — which matters when you need permission to remove a marketing tag.
Why it compounds
A faster site converts more of the same traffic, so every future marketing pound buys more. It also crawls more efficiently, so new content is discovered faster. And it lowers infrastructure cost per visit.
None of those effects are dramatic in isolation. Together, over two years, they are usually larger than the redesign that got all the attention.
Budgets or it decays
Without a gate in CI, performance regresses. Someone adds a hero video, a chat widget, a font weight, a tag. Each is defensible; collectively they undo the work.
A performance budget that fails a pull request turns that from a quarterly cleanup into a five-minute conversation at the moment of the change.
Where the time actually goes
In audits across commerce and B2B sites, the distribution is remarkably consistent. Images and video account for the largest share of transferred bytes; third-party JavaScript accounts for the largest share of main-thread blocking; and server response time is the quiet third that nobody looks at because it needs backend access.
That consistency is good news, because it means the first pass is predictable work rather than an investigation. Fix the media pipeline, audit the tags, measure the server, and you will have captured most of the available gain before you begin the interesting optimisations.
Only then is it worth arguing about hydration strategies and bundle splitting — which is roughly the reverse of the order most performance conversations happen in.
Questions people ask about this
Largest Contentful Paint under 2.5 seconds, Interaction to Next Paint under 200 milliseconds and Cumulative Layout Shift under 0.1 — measured at the 75th percentile of real users, not on a developer laptop.
It is a genuine but modest ranking input. The larger effect is commercial: faster pages convert more of the traffic you already have, and are crawled more efficiently so new content is discovered sooner.
Performance budgets enforced in CI. Without a gate that fails a pull request, expect regression within two quarters as tags, fonts and widgets accumulate — each individually defensible.