Performance is a feature you have to protect. I keep a checklist that takes about 30 minutes to run and catches 90 percent of regressions. It has saved three launches I can name — a storefront whose hero video added four seconds to mobile LCP, a dashboard shipping two chart libraries for one chart, and a marketing site loading five font weights for two used. Every one was caught by the same routine below.
Set a budget before you optimize
A budget turns vague concern into enforceable numbers. Mine for typical marketing and SaaS pages: under 170KB of compressed JavaScript per route, under 1MB total page weight on mobile, zero render-blocking requests beyond the critical CSS and fonts, and Core Web Vitals passing on a mid-range Android device over 4G — not just desktop Chrome.
Enforce it in CI where possible. Bundle analyzers and Lighthouse budgets fail the build when a dependency pushes a route over limit, which forces the conversation at pull-request time instead of after launch. Manual discipline fades; automated gates do not.
Images, fonts and media first
Images are the biggest lever on most pages. Serve AVIF with WebP fallback, size images to their display slots instead of uploading 4000-pixel originals, set explicit width and height to prevent layout shift, and lazy-load everything below the fold. One e-commerce audit cut 2.1MB to 380KB from these steps alone.
Fonts come second. Use display swap so text never stays invisible, subset to the languages you serve, and ship at most two weights per family. Every additional weight is a render-blocking download on first visit. For video, never autoplay a multi-megabyte file as a hero — use a poster image with click-to-play, which is what finally fixed that four-second storefront LCP.
- Audit the largest image on each template — convert to AVIF/WebP, resize to display size, add dimensions
- Subset fonts, keep two weights, verify display swap and preconnect to the font host
- Lazy-load below-the-fold media and defer offscreen iframes
- Replace hero video autoplay with poster plus click-to-play
- Defer third-party scripts past interactivity and remove tags nobody owns
JavaScript and third parties
Audit every kilobyte of client JavaScript by asking what breaks without it. Marketing pages have shipped with A/B testing, chat widgets, session replay and four analytics tools simultaneously — each one taxing INP long after load. Keep one analytics path, lazy-load chat and replay on interaction, and code-split route-level dependencies so the homepage never pays for the dashboard charts.
Third-party scripts deserve a named owner and a removal date. My rule: any tag without an owner gets removed in the next sprint. Consent-mode analytics that fires only after a choice also keeps the initial load honest, which pairs well with a clear cookie notice.
Measure like your users browse
Always measure with real device profiles — a mid-range Android on throttled 4G, not your development laptop. Track LCP under 2.5 seconds, INP under 200 milliseconds and CLS under 0.1 as the passing bar, and record the numbers per template so regressions show up as deltas instead of vibes.
Reviewing the checklist every sprint keeps performance discussions concrete and prevents the slow creep that makes sites feel heavy a year later. Run it alongside the technical SEO audit before every launch — the two overlap heavily on images, scripts and crawl efficiency — and treat any regression as a bug, not a tradeoff, until proven otherwise.


