Fix the text and test conditions
Choose a heading or component. Record text, version, viewport, zoom and network profile. Comparing a short English heading with longer Spanish copy does not establish a font regression.
In a fictional enquiry flow, a button moves down as a heading becomes three lines. Note the timing against the font request. Coinciding events provide a lead to investigate rather than a completed causal explanation.
Compare three loading cases
Load with a cold cache, reload with a warm cache, and block the relevant font file in a separate run. Record what actually renders. A quick reload can conceal a first visitor's transition.
Also check the family and weight actually used by the browser. A missing weight or character can trigger an unexpected substitution. Readable fallback text needs to work even when the intended file never arrives.
Isolate swapping from other changes
Keep other CSS and text unchanged while comparing font cases. Inspect wrapping, text dimensions and component space. Images without reserved space or late inserted content can move the same button and need separate tests.
font-display controls text presentation and swap periods. Choosing swap does not itself eliminate movement. Do not hide the problem with a fixed component height that clips longer text.
| Observation | Next test | Avoid concluding |
|---|---|---|
| Movement on first load | Compare timing and actual font | All movement comes from fonts |
| Stable warm cache | Repeat with cold resources | First visits are proven stable |
| Only some characters fail | Inspect glyph coverage and fallback | The whole family failed to load |
| Blocked file clips text | Test flow and component space | Hide text until the file arrives |
| Movement without swapping | Isolate CSS, images and inserted content | Change font-display at random |
Keep the correction and a working failure case
Repeat the same test after one bounded change. Include longer headings, relevant languages, mobile width and text zoom. Record when the file is used and which components remain readable without it.
A stable test page demonstrates the tested situation. It does not prove an improvement to the entire site's field CLS. Follow real data where available and keep the case for future font and template regressions.