Speed work goes wrong when it starts with a score instead of a measurement. A tool prints 68, someone asks for 90, and the next three weeks go into whatever the tool listed first. I work the other way round: find which of the three Core Web Vitals is failing, find what is causing it on that specific template, fix that, and re-measure.
The three measurements, and their thresholds
Google publishes exactly what "good" means, so there is nothing to argue about:
- Largest Contentful Paint measures loading. Good is 2.5 seconds or less.
- Interaction to Next Paint measures responsiveness. Good is 200 milliseconds or less.
- Cumulative Layout Shift measures visual stability. Good is 0.1 or less.
The number that counts is the 75th percentile of real page views, not your own load on your own connection. If three quarters of visits to a page meet the threshold, the page passes.
Lab data and field data are different questions
PageSpeed Insights gives you two things at once and they answer different questions. The field data at the top is what real visitors experienced over the last 28 days. The lab run underneath is a single simulated load on a throttled mobile connection.
Field data tells you whether you have a problem. Lab data helps you find the cause. I never report a lab score as an outcome on its own, because a lab run can improve while real visitors see nothing change.
The order I work in
1. Largest Contentful Paint first. It is the one that usually fails, and it is usually caused by something structural rather than by image weight alone. What I check: how long the server takes to send the first byte, whether the largest element is discoverable in the initial HTML or only appears after JavaScript runs, whether the hero image is preloaded or lazy loaded by mistake, and whether render-blocking CSS and fonts sit in front of it.
The worst case I have measured was a yacht charter fleet listing at 20.6 seconds on mobile, against Google's 2.5 second threshold. No amount of image compression was going to fix that. The page served Googlebot a "Loading yachts..." placeholder and built the real content in the browser afterwards. That is an architecture problem wearing a speed problem's clothes.
2. Layout shift second, because it is usually cheap. Images and embeds without reserved dimensions, fonts that swap and reflow the text, and banners injected above existing content account for nearly all of it.
3. Interaction to Next Paint last, unless the page is obviously heavy on script. Long tasks on the main thread, oversized third party tags and event handlers that do too much work in one go are the usual causes. Tag managers are worth auditing on their own here.
What this looks like on a real site
On the Klimastål build I took desktop performance from 68 to 96 across four measured checkpoints, and mobile Largest Contentful Paint sat at 2.4 seconds, inside Google's good band. That took server response work, correct image handling and removing render blocking, in that order. It did not take a rewrite.
What I hand over
Two things. A before and after for each of the three measurements with the date and the device stated, and a short note of what caused each failure. Speed work you cannot explain is speed work nobody will maintain, and next quarter's new hero video will undo all of it.
If your field data is failing and you are not sure which of the three is at fault, that is the first thing to establish. Everything else is guessing with a stopwatch.
Want this run on your site?
Tell me the domain and the pages that matter. I will come back with what I would fix first and what it should move.
Start a project