Speed work goes wrong in a predictable way: someone runs a lab test, chases the score, applies a plugin, and the number improves while actual visitors experience nothing different. Lab tests run on a fast machine on a good connection. Your users are on mid-range phones on variable networks, and that's the data Google uses to assess you.
So we start with field data — real user measurements from Chrome UX Report and your own analytics — and identify which specific interactions are slow, on which devices, in which regions. Then we fix in order of measured impact. That usually means images and media first, because they're the largest thing most sites ship. Then JavaScript: what's loading, what's blocking, what's third-party and whether it earns its cost. Then fonts, caching, CDN configuration, and where relevant the database queries and hosting underneath.
Every change is measured before and after, so you can see what each one bought. And because performance decays — new scripts, new plugins, unoptimised uploads — we finish with monitoring and a performance budget, so regressions surface as alerts rather than as a slow decline someone notices two quarters later.