Reduce a WOFF2 font to the glyphs you need, choose font-display deliberately, and preload only a file needed above the fold. In our delayed-font fixture, swap produced CLS 0.00646, optional zero, and preload did not remove the shift.

A font can delay visible text, change line wrapping after download and compete with an LCP image for bandwidth. There is no universal “preload plus swap” recipe. The outcome depends on arrival time and how close the fallback metrics are. This guide gives a repeatable workflow for a local WOFF2 font. If the moving content is an image or banner, start with the CLS diagnostic guide.
What you need
- WOFF2 files and access to
@font-faceand the HTML head. - A list of initial font weights and language subsets used above the fold.
- Chrome DevTools Network and Performance recordings with an empty cache.
- A page where the font appears early enough that its loading strategy matters.
Step 1. Find the files the browser actually downloads
Open Network, filter by Font and reload without cache. Record each WOFF2 size, request start and initiator. If five weights download while the first screen uses one, remove unnecessary early requests first. Separate Latin and Cyrillic files can make sense on a bilingual site, but inspect unicode-range: a Russian page should not fetch unrelated subsets.
Subsetting reduces bytes, but excessive splitting creates requests and weakens cache reuse. Split by useful glyph groups, not by individual words. Keep currency marks, digits, punctuation and interface characters. A missing glyph may fall back to another face and alter geometry in an unexpected place.
Step 2. Choose font-display for the job
| Value | What visitors see | Trade-off |
|---|---|---|
| swap | Fallback text immediately, then web font | Late swap can reflow lines |
| optional | Fast fallback; late font may not apply this visit | Some users will not see the brand face |
| fallback | Short wait, then fallback | Behavior needs network testing |
| block | Text may be invisible while loading | Visible content may be delayed |
For body text, swap with a well-matched fallback is often reasonable. For a decorative face, optional may be preferable. Do not optimize only the CLS number: users must be able to read text promptly. Compare cold-cache loading in the speed tool and DevTools.
Step 3. Match fallback metrics
Fonts differ in glyph width, x-height, line spacing and baseline. A swap can move a word onto the next line even when both files load quickly. Start with a fallback of a similar type. If needed, use size-adjust, ascent-override, descent-override and line-gap-override in a dedicated fallback @font-face. Derive values from measurement rather than copying numbers from another site's CSS.
@font-face {
font-family: "Site Text";
src: url("/fonts/site-latin.woff2") format("woff2");
font-display: swap;
unicode-range: U+0000-00FF;
}
body { font-family: "Site Text", Arial, sans-serif; }
/* local fixture: swap CLS 0.0064575; optional CLS 0 */Add another @font-face for Cyrillic with its actual file and range. Do not claim that one subset contains glyphs it lacks. Test RU and EN pages with numbers, currency and unusual characters. A small width difference can add a full line in a long heading.
Step 4. Preload only a critical font
Preload can help when CSS would reveal a font late. Use exactly the same URL, format and crossorigin, including for a same-origin font because fonts are CORS resources. A mismatch can lead to two downloads. Confirm in Network that the browser issues one request for the file.
<link rel="preload" href="/en/fonts/site-latin.woff2"
as="font" type="font/woff2" crossorigin>
<!-- Only when this file is used above the fold. -->Do not preload every weight and language. Early requests compete with HTML, CSS and the LCP image. If a font is already discovered immediately, preload may provide little benefit. Our fixture delayed the file response by 700 ms; the preload variant still produced CLS 0.0064575 because the font did not arrive before initial layout. That does not establish that preload is useless on other sites.

What the delayed-font test showed
We compared swap, optional and swap with preload on one artificial page. Each variant ran three times in fresh browser contexts at 800×900; the server delayed WOFF2 by 700 ms. All three swap runs scored 0.0064575 CLS. All optional runs scored zero. Preload with swap also scored 0.0064575. The simple page's LCP records early text paint and does not tell us how a production site would perform.
This isolates one late-swap effect for one Inter WOFF2 file. It is not field data and does not prove optional wins for every site. If the brand face is essential above the fold, also evaluate the visual quality of fallback and business impact. A site whose main text is the LCP element may behave differently.
Verify after release
Record Network and Performance on a mobile viewport with cold cache and a slow connection. Inspect font request time, duplicate downloads, line wrapping and Layout shifts. Then compare field metrics for the same URL and device, remembering the CrUX time window. The field versus lab guide explains that comparison. If fonts arrive late only for some visitors, inspect CDN region and caching before randomly changing CSS.
After replacing a font, verify that the server returns font/woff2, appropriate cache headers and a versioned URL. A 404 font often looks like a visual regression rather than an outage. Add a RU and EN visual check to the launch checklist. For a slow file response, the curl timing guide can separate network phases.
How we made this
On 14 November 2026 a local HTTP fixture in Chrome 154 delivered Inter WOFF2 after a 700 ms delay. Three variants ran three times in fresh contexts; PerformanceObserver collected CLS and LCP. Results are in lab/data.json, with code in lab/server.py and lab/measure.cjs. This test contains no field sample and does not prove an LCP change on enterno.io. Recommendations were checked against web.dev and MDN.
What to check on a bilingual page
Russian and English pages use different character sets and heading lengths. Separate WOFF2 subsets help when CSS uses accurate unicode-range declarations and the browser does not fetch both unnecessarily. Test Cyrillic and Latin text, currency symbols, dashes, quotation marks and digits. A missing glyph can fall back to a different face and move part of a line.
Record a cold-cache first visit and a warm-cache return visit. If fonts live on another domain, check CORS and redirects. Replacing a font file at an unchanged URL with long-lived cache headers can leave some visitors seeing the older face. Versioned file names let you tie a measurement to the actual release.
Preload cannot repair an ill-matched fallback face. If the main file misses the first frame, the browser still lays out fallback text. Compare line lengths and block heights first, then measure whether earlier discovery helps your CSS chain. Disable preload for a control run and compare request order as well as the final score. If it makes no difference, remove the extra priority.
FAQ
Should I preload every WOFF2 file?
No. Preload only a first-screen font that would otherwise be discovered late. Too many early requests can starve critical resources.
Is font-display: optional always better than swap?
No. Optional can keep layout stable, but some visits will use fallback throughout. Choose based on brand requirements and measured results.
Why did the font load without changing CLS?
It may have arrived before first layout, matched line metrics, or stayed unused this visit. Inspect the Layout shifts track and computed font on the element.