You can treat cache http as a checkbox, or as a design decision. The first path is fast and ineffective; the second takes an hour of thought and lasts for years.
Weight is not the only culprit
Cutting kilobytes helps, but load order often matters more. A lightweight render-blocking resource costs more than a heavy lazily-loaded image.
Start by identifying what blocks, then what weighs — in that order. The reverse leads to lengthy optimisation of elements with no perceptible effect.
A budget, or nothing
Without an explicit ceiling, a page grows with every change. Setting a budget — in weight, requests and render time — turns performance into a design constraint rather than an occasional project.
The budget must be checked automatically, otherwise it becomes a polite intention nobody consults.
How to measure that it works
Set the measurement before acting, not after. Record the starting value, the date, and the exact metric you expect to move. Without that initial snapshot, any conclusion is just an impression.
Prefer one metric per project. Following twelve curves at once feels like control, but makes it impossible to attribute a result to a cause.
Check before moving on
- Automate every repetitive check
- Check the rendered source, not just the display
- Schedule a review in six months
- Assign each point to a named person
- Measure before, not after
Measure in the field, not the lab
A score obtained on a fibre-connected development machine says nothing about real experience. Field data, collected from actual visits, often reveals a considerable gap.
This is especially true on mobile, where network and CPU variability make averages misleading: look at the 75th percentile, not the mean.
Visual stability, the detail that irritates
Content jumping during load causes frustration out of all proportion to its technical cause, often trivial: an undeclared image dimension, a badly calibrated fallback font.
These are the cheapest fixes in all of web performance, and among the most perceptible to users.
Read next
- Optimiser le LCP sans tout réécrire
- Auto-héberger ses polices : vitesse et conformité
- WebP, AVIF et srcset : diviser le poids par trois
Going further
This deserves a review every six months: engines change, your content ages, and a setup that was correct last year may now be counterproductive. Schedule the review rather than enduring it.
