Web font loading at Ithy: a 76% smaller Latin payload and three lab observations

Ithy’s actual font-file measurements and three saved mobile lab observations. Download the data and see why smaller payload, lab results, field experience and rankings are different claims.

Published 2026-10-24 · AI-assisted Ithy editorial. Primary references checked 2026-10-24. Original examples are labeled; this is not a named human expert review.

Ithy reduced the combined regular-and-bold Latin font-file payload from 367,620 to 88,140 bytes while retaining its original full fonts as fallbacks. One later mobile lab observation was faster, but this was not a controlled experiment and does not establish a search-ranking gain. This case study separates the reproducible byte comparison from the more limited page-speed observations.

What changed, and what was measured

The work concerned shared web-font loading on an article and research site. We used the author's unmodified LatoLatin 2.015 files, kept the full faces for other required characters, and retained the full fonts for standalone article downloads. The author's Lato download page describes full and Latin-only webfont packages. The licenses remain bundled with the font assets.

We also changed font discovery/loading around the existing interface. The page-speed sequence includes earlier rendering work and changing runtime conditions; it cannot isolate the effect of file size alone. The mobile observations below concern the homepage, not a controlled sample of all article URLs.

The reproducible result: font-file bytes

PairRegular plus bold bytesDifference
Original full faces367,620Reference
Author's Latin faces88,140279,480 fewer bytes, or 76.02%

These are font-file lengths, not total page weight or a promise that every language downloads the same resources. The English browser check requested the smaller regular and bold faces. A page needing additional characters, weights or styles can have a different font set.

For regular, bold and italic, the verification compared the 390 shared glyphs per style: outlines, advances and line metrics matched for those glyphs. That is a bounded compatibility check. It does not prove identical rendering on every operating system, coverage of every character or equivalence between arbitrary fonts that share the name “Lato.”

Download the original payload measurements. The public files can also be inspected directly: full regular, full bold, Latin regular and Latin bold.

Three saved mobile lab observations

ObservationLighthouse performanceFCPLCP
Earlier baseline616.0 s7.3 s
After rendering changes624.8 s7.3 s
After the font-loading change862.3 s3.8 s

The reports were saved during the October 23–24, 2026 work. Each row is one mobile lab observation. There was no randomized assignment, repeated-run distribution or controlled causal comparison. Server conditions, cache state, other source changes and test variation can affect the result. The table should not be read as “76% smaller fonts caused a particular ranking or speed increase.” Download the observation table.

Why the real-user assessment still failed

The last report also showed historical 28-day field values of LCP 4.1 seconds, INP 594 milliseconds and CLS 0.16, with a failed Core Web Vitals assessment. Those field values are different observations from the single lab load. They did not become a clean post-change cohort immediately after release.

Google's PageSpeed documentation distinguishes Lighthouse lab diagnostics from CrUX field experience over the previous 28 days. A better lab observation can identify progress worth investigating without proving that real-user experience or organic traffic has recovered.

Checks that prevented a misleading “optimization”

  1. Compare font versions. An earlier candidate was rejected when its metrics differed. A smaller file is not enough if the intended design changes.
  2. Keep necessary character coverage. Use explicit ranges and appropriate fallbacks rather than assuming an English page represents every script.
  3. Check delivery separately. Inspect which resources the browser actually requests; CSS declarations alone are not a transfer measurement.
  4. Preserve portable downloads. A self-contained HTML export needs its intended embedded resources rather than depending on the live site's subset URLs.
  5. Report what was not measured. We did not establish a ranking effect, a universal device result or a controlled font-only speed effect.

Google's font-loading guidance explains discovery, delivery and rendering tradeoffs, including the potential for font swapping to shift layout. It also cautions that preloading can consume resources and bypass character-range negotiation. Preloading every font is not a general solution.

For a related document problem, see font availability and portability in Word. For evaluating the claims in this case study, see how to check a citation against the specific claim it supports.

For a correction, send the specific claim and a supporting source through Ithy Help. Research methodology · More research guides.