Native or Cross-Platform? A Practical Way to Decide in 2026
The honest answer depends on four specific factors. Here is the decision framework we use with clients, and where each...
Read article
We took a client site from an eleven-second load to 1.1 seconds. Here is the ordered list of changes and roughly what each one contributed.
Performance advice is usually a long undifferentiated checklist. In practice a small number of changes deliver most of the improvement. Here is what we did on a recent rebuild, in order of impact.
Images - roughly 60 percent of the total gain
The original site served full-resolution JPEGs, some over 3 MB, scaled down in the browser. We converted to AVIF with WebP fallbacks, generated responsive sizes, added correct width and height attributes to eliminate layout shift, and lazy-loaded everything below the fold. Page weight fell by 84 percent. This one category of work delivered more than everything else combined.
Render-blocking resources - about 20 percent
Six stylesheets and eleven synchronous scripts sat in the head, including two chat widgets and three analytics tags. We inlined critical CSS, deferred the rest, delayed third-party scripts until after first interaction, and removed two tags nobody could identify an owner for.
Fonts - about 10 percent
Four families in nine weights, self-hosted without preloading. We cut to two families in four weights, preloaded the critical ones, subsetted to Latin, and used font-display: swap. Largest Contentful Paint improved by nearly a second on its own.
Server and caching - about 10 percent
Time to First Byte was 900 milliseconds because every page ran uncached database queries. Object caching in Redis, full-page caching for anonymous visitors, and a CDN in front brought TTFB under 200 milliseconds.
What barely mattered
Minifying already-gzipped CSS saved a few kilobytes. Removing unused CSS helped less than expected. Neither was worth the effort compared to the image work.
Measure on the right device
Test on a mid-range Android phone on a throttled connection, not on your own laptop over office fibre. Lab tools are useful for diagnosis, but field data from real users is what Google actually scores you on - and the difference between the two is frequently large.
Share this article