We had never run Lighthouse on this site. We have a contrast check, a check that no page scrolls sideways on a phone, and a sweep that every calculator still computes. Today we ran Google's Lighthouse 12.8.2, mobile mode, on the 16 pages a first-time visitor is most likely to open: the home page, the research and writing indexes, the stories page, one blog post and the 11 flagship tools.
The tools did fine. All 11 flagships scored 100 for accessibility and 100 for SEO. The weak pages were the indexes:
| page | accessibility |
|---|---|
| /research/ | 88 |
| /writing/ | 92 |
| a blog post | 95 |
What failed:
- Coloured labels too faint to read. Each research theme has its own colour, and the small theme label was drawn in it. 18 of the 31 light-mode colours were below the 4.5:1 contrast that WCAG asks of small text. The lightest green was 2.59:1.
- A skipped heading level. The "Start here" cards were
h3headings directly under the page'sh1. A screen reader user moving by heading would find no section heading above them. The small grey label above the cards is now anh2and looks the same as before. - Links shown by colour alone. "Topic page →" sat inside grey text and only its blue told it apart, at 1.39:1 against that grey. It is underlined now.
- Tiny tap targets. The writing index lists 67 page numbers as links 9 pixels wide. Each is now at least 24 by 24, the WCAG 2.2 minimum, and the row reads as an even grid. The tag pills on posts went from 21.8 to 25.8 pixels tall.
The contrast fixes came from another session working the same morning (PR #14667). This one fixed the heading, the link and the tap targets. With both, all 16 pages score 100 for accessibility.
Why did the tools pass while the indexes failed? Every tool page goes through the same shared build steps and checks. The index pages are written by hand, one generator each, and the browser checks we run were mostly aimed at tools. The script that ran these numbers is now in the repository (platform/pages/lighthouse.mjs), so the next check covers the indexes too.
Two limits on these numbers. They come from a local build, not the live site. Performance under simulated phone throttling moves a point or two between runs. The slowest pages were the wall-framing takeoff and the AI bookkeeper, at 83 and 84, with a first paint of 3.3 seconds. They are next.