A perfect score is not the goal
Performance work has an unhelpful gravitational pull towards a number in a testing tool. Teams spend weeks getting from 78 to 96 and conversion does not move, because the remaining eighteen points were in metrics nobody experiences.
The useful question is narrower: how long before a visitor can see and use the thing they came for, on the device and connection they actually have?
The three that matter and why
- Largest contentful paint — when the main content appears. Under 2.5 seconds is the working target. This is the one that correlates most visibly with bounce.
- Cumulative layout shift — whether things move as the page loads. Shifting layouts cause mis-taps, which cause abandonment and, on checkout pages, real anger.
- Interaction responsiveness — whether the page reacts promptly when tapped. A page that looks ready but ignores taps for a second feels broken.
Everything else is diagnostic. Useful for finding causes, not worth optimising as an end in itself.
Measure on real devices, not your laptop
Your site is fast on a developer's machine on office fibre. Your customers are on a mid-range Android on a train. Test on throttled connections and modest hardware, and use field data from real users rather than lab tests alone.
Lab scores tell you what could happen. Field data tells you what does. When they disagree, the field data is the one your customers experienced.
The fixes with the best return
- Images. Still the biggest single cause. Correct dimensions, modern formats, lazy loading below the fold, and an explicit width and height so nothing shifts.
- Fonts. Preload the one or two you actually use, subset them, and set a swap behaviour so text appears immediately rather than after the download.
- Third-party scripts. Audit them. Most sites carry tags for tools nobody has used in a year, each costing real milliseconds.
- Render-blocking CSS and JS. Inline what is needed for the first screen, defer the rest.
- Caching and a CDN. Unglamorous, effective, and often a configuration change rather than a project.
Third-party scripts deserve their own audit
Analytics, chat widgets, heatmaps, ad pixels, review widgets, cookie banners. Each was added for a reason and few are ever removed. Together they frequently account for more load time than the entire site.
List every one, name an owner, and delete anything nobody claims. Then load what remains after the page is usable rather than before. A chat widget that appears two seconds late costs nothing; one that blocks the page costs conversions.
Decide what a second is worth before you spend
Take current conversion rate and traffic, apply a conservative uplift assumption to the segment that is currently slow, and compare with the cost of the work. Often the honest answer is that a fortnight of image and script work captures most of the value and further optimisation does not pay.
Performance work also decays. Someone adds a hero video, someone else adds a tag, and six months later you are back where you started — so budget a quarterly check rather than a one-off heroic effort.