When I audited all 38 pages of our site for accessibility a few weeks ago, the contrast section counted elements: 546 links at 2.53:1, a focus ring at 1.82:1. That is how every tool I know of reports it, including the ones built into the browser. An element is a row, and a row is a row whether it holds one word or two hundred.
That bothered me, because a reader does not read elements. So I ran it again weighted by character.
Method
Playwright, the 38 URLs in our sitemap, and a walk over every text node in the body. For each node I took the computed color of its parent, then climbed the tree until I found an ancestor with a background whose alpha is over 0.5, which is the colour the text is actually drawn on rather than the one I assumed it was on. Then I weighted each colour pair by how many characters sit in it.
text nodes measured 4,207 characters 214,160 distinct colour / background pairs 36 distinct backgrounds 10 characters in text large enough to qualify for the 3:1 threshold 9,026 (4.2%) Ten backgrounds, not one, which is the part you cannot get from a stylesheet: 197,293 characters sit on #0a0a0a, 14,886 on the #111111 of raised cards, and small amounts on #0d0d0d, #141414, #1e1e1e and, on the inverted buttons, 618 characters of near-black on near-white.
The result
characters failing their WCAG AA threshold 84,508 (39.5%) characters under 3:1 at any size 29,622 (13.8%) Here are the pairs that carry the text, which is nine tenths of the whole site:
colour on background characters ratio AA #a3a3a3 on #0a0a0a 92,617 7.85 pass #737373 on #0a0a0a 52,766 4.18 fail #525252 on #0a0a0a 23,332 2.53 fail #fafafa on #0a0a0a 21,159 18.97 pass #a3a3a3 on #111111 6,420 7.49 pass #d4d4d4 on #111111 3,683 12.74 pass #e5e5e5 on #0a0a0a 2,915 15.72 pass #404040 on #0a0a0a 2,335 1.91 fail #737373 on #111111 1,881 3.98 fail #3a3a3a on #0a0a0a 1,298 1.74 fail Two greys are 90 per cent of the failure: #737373 and #525252 together carry 76,098 of the 84,508 failing characters.
The ranking changes when you count characters
This is the bit I actually wanted. Same measurement, two ways of totalling it:
colour text nodes characters characters per node #a3a3a3 685 99,405 145.1 #737373 510 54,882 107.6 #525252 1,307 24,728 18.9 #fafafa 896 22,429 25.0 #404040 346 2,415 7.0 Count elements and #525252 is the headline: 1,307 nodes, more than twice as many as any other failing colour, and it is the one my earlier audit led with. Count characters and it drops to third, because at 18.9 characters per node it is almost entirely labels, eyebrows, table captions and footer links. Two or three words at a time.
#737373 is the opposite shape. 510 nodes at 107.6 characters each, which is paragraphs. And its font sizes are:
13px 29,607 characters 14px 17,116 13.5px 6,932 12px 948 16px 205 None of that is large text under WCAG, so all of it needs 4.5:1 and all of it has 4.18:1. Fifty-two thousand characters of prose, in 13 and 14 pixel type, 7 per cent short of the threshold. An element-weighted audit put that fourth. A reader-weighted one puts it first, and the reader is right.
It misses by six values
The colours on this site are hand-typed hex in arbitrary-value Tailwind classes, which I counted once before: 1,216 hex literals across the site, the desktop UI and the email templates. So the fix is not a token change, it is a find and replace, and the question is what to replace it with.
on #0a0a0a, the lightest grey that fails 4.5:1 #787878 the darkest grey that passes #797979 what we use #737373 on #111111, the darkest grey that passes #7c7c7c Six values. #737373 to #797979 is a 2.4 per cent change in brightness and it is the difference between 39.5 per cent of our characters failing and 13.9 per cent of them failing. Because 14,886 characters sit on the #111111 cards, the value that works everywhere is #7c7c7c, which is nine steps up and still reads as muted next to #a3a3a3.
#525252 is the harder one. At 2.53:1 it is not a near miss, it is a decision: that text is deliberately quiet. 1,307 nodes of it are links in the footer and small print, and the honest answer is that most of it should be #737373's replacement too, with the genuinely decorative remainder, 361 nodes of #3a3a3a and 126 of #2e2e2e, kept for borders and separators where no characters are involved at all.
What this does not measure
Three caveats, because a number like 39.5 per cent deserves them:
- Opacity. I resolved backgrounds through alpha but I did not multiply text colour by a parent's
opacity, so any faded text is measured brighter than it renders. There is not much of it on this site, and every instance would make the result worse rather than better. - Backgrounds with gradients or images behind the text. We have none behind prose, which is lucky rather than deliberate.
- Characters are not words, and words are not reading time. A character count still over-weights long URLs and under-weights a four-word heading that everybody reads. It is a better proxy than an element count, not a perfect one.
The pages are public if you want to check any of it: a guide and a comparison page are two of the heaviest in prose at 6,267 and 6,301 characters, the download page is the lightest at 851, and the text you are squinting at on any of them is almost certainly #737373.

