The field-data distinction matters. A page can score well in a synthetic test run on fast hardware and still fail Core Web Vitals, because the score reflects real users on real devices and networks. Optimising against a lab tool without checking field data is a common way to spend effort and see no change.
The usual causes are unglamorous: oversized images, render-blocking scripts, fonts that swap late, and content injected above existing elements after load. Fixing layout shift in particular is often just reserving space for images and embeds before they arrive — cheap work with a disproportionate effect on the score.
Codazz builds this in production — Site Speed Optimisation.