Our house style for written copy bans the em dash and the exclamation mark, and there is a test in the repository that enforces it over the content source. A test over source is not the same claim as a fact about production, though. Source goes through a markup layer, a React tree, HTML entity encoding and a CDN before anyone reads it.
So I counted the characters that actually arrive. Not bytes, not words: code points, one census over every public page.
The subject is Nakodo, 76 public pages, server rendered.
How it was measured
Each of the 76 sitemap URLs is fetched with curl, not read out of a browser, because my browser profile for this work is signed in. Each document goes through Python's html.parser with convert_charrefs=True, skipping the contents of script, style, head and title, and the remaining text is whitespace-normalised and fed to a Counter.
Three details decide whether the numbers mean anything:
-
Entities have to be decoded. The raw HTML contains
'3,445 times. Those are apostrophes, not three-character strings, and a census that counts the source literally reports a punctuation mark that nobody typed and nobody sees. -
The script contents have to go. A React server component payload carries an escaped copy of the page's text. Leave
scriptin and every visible character is counted roughly twice, which happens to preserve the ratios and destroy the totals. - Code points, not bytes. The pound sign is one character and two bytes. The two Japanese characters in our footer are one each and three bytes each. A byte census of a mostly-English site tells you about UTF-8, not about the writing.
748,219 characters, 97 distinct code points
83 of them are ASCII. 14 are not. The non-ASCII characters appear 453 times in total, which is 0.06% of the text:
| Code point | Character | Name | Occurrences | Pages |
|---|---|---|---|---|
| U+00A9 | © |
COPYRIGHT SIGN | 78 | 76 |
| U+4EF2 | 仲 |
CJK IDEOGRAPH | 77 | 76 |
| U+4EBA | 人 |
CJK IDEOGRAPH | 77 | 76 |
| U+014D | ō |
LATIN SMALL O WITH MACRON | 77 | 76 |
| U+00A3 | £ |
POUND SIGN | 46 | 5 |
| U+00F7 | ÷ |
DIVISION SIGN | 33 | 11 |
| U+00E9 | é |
LATIN SMALL E WITH ACUTE | 21 | 5 |
| U+00D7 | × |
MULTIPLICATION SIGN | 16 | 2 |
| U+00B7 | · |
MIDDLE DOT | 9 | 1 |
| U+20AC | € |
EURO SIGN | 8 | 3 |
| U+201D | ” |
RIGHT DOUBLE QUOTATION MARK | 4 | 2 |
| U+201C | “ |
LEFT DOUBLE QUOTATION MARK | 4 | 2 |
| U+00E8 | è |
LATIN SMALL E WITH GRAVE | 2 | 1 |
| U+2013 | – |
EN DASH | 1 | 1 |
Four of those fourteen are on every single page, and they are the same two strings: the copyright line in the footer, and a one-line footnote next to it explaining the brand name, which is Japanese and therefore carries the only ideographs and the only macron on the site. That footnote is also why no page on this site has a purely ASCII alphabet, which was not a result I predicted.
The division and multiplication signs are arithmetic in the guides, where a worked example reads fee = median views ÷ 1,000 × CPM. Someone chose the real operators over / and x, and the choice reaches production 49 times. The accented letters are two words: cafés and cafetière, in example content about coffee. The middle dot, nine times on one page, separates stats in the example cards on the home page.
The currency signs are not constants
The pound sign is on five pages, the euro sign on three. That is not a typographic choice, it is a feature: prices are shown in the visitor's currency, so the pricing page came back in euros for the machine that ran this fetch and will come back in pounds for a visitor in the UK. A character census of a page with localised prices measures the fetch, not the page, and the honest way to report it is as two numbers rather than one.
The dollar sign, 34 occurrences on four pages, is different: those are worked CPM examples in the guides, written in dollars because that is the unit the industry quotes in, and they are the same for everyone.
Twelve printable ASCII characters that never appear
Of the 95 printable ASCII characters, 83 are used and these 12 are not:
! * < > Z \ ^ ` { | } ~
The first one is the point of the exercise. The exclamation mark is banned by our editorial test, and 748,219 characters of production text contain none. The test and the deployment agree, which is the whole reason to do a census over output.
The rest split into three stories. Z is missing by accident of vocabulary: there are 238 lowercase z characters and not one capital, because no sentence on the site starts with a word like Zero and no proper noun in our copy begins with Z. *, `, {, }, |, \, ^ and ~ are missing because there is no code on the public site: as the element census found, there is no code or pre element anywhere, so there is nowhere for a brace or a backtick to live. < and > are absent for the dull and reassuring reason that nothing is double-escaped.
Square brackets, by contrast, are used 349 times across 25 pages, because the guides teach search phrasings with a placeholder in them, written as [problem] for the reader to replace.
The quotes tell you who wrote the page
| Character | Count |
|---|---|
' straight apostrophe |
2,364 |
" straight double quote |
952 |
“ ” curly double quotes |
8 |
’ curly apostrophe |
0 |
| U+00A0 no-break space | 0 |
| U+2026 horizontal ellipsis | 0 |
| U+2014 em dash | 0 |
| U+2212 minus sign | 0 |
Zero curly apostrophes in 748,219 characters is a strong signal: nothing on this site was pasted out of a word processor. The eight curly double quotes are the exception, and they are all on two pages, the privacy notice and the terms, in phrases like reply “stop” and “we” and “us” means CogniPrep Ltd. Those two documents were written in a different register from the rest of the site, and the typography records it.
The same two pages hold the other anomaly. There is exactly one en dash on the entire website, and it is in the privacy notice, in the phrase the UK–US data bridge. One character out of 748,219, in a compound that is conventionally set with an en dash in formal writing, sitting on a site whose style guide bans dashes of that family. I only know it is there because I counted.
Character classes
| Class | Count | Share |
|---|---|---|
| Letters | 593,726 | 79.4% |
| Spaces | 127,344 | 17.0% |
| Digits | 4,579 | 0.6% |
| Commas | 7,932 | 1.1% |
| Full stops | 7,208 | 1.0% |
| Hyphens | 900 | 0.1% |
| Colons | 868 | 0.1% |
| Question marks | 419 | 0.06% |
| Semicolons | 177 | 0.02% |
0.6% digits on a site that is largely about numbers surprised me, until I remembered that the numbers live in tables and in the four calculators, where they are inputs rather than copy. The 419 question marks are almost entirely the FAQ blocks. The 177 semicolons are spread over 65 of the 76 pages, which is about one semicolon every 4,200 characters: not absent, not a habit.
Per page, the median alphabet is 74 code points. The widest is how to measure YouTube sponsorships at 81, because it contains arithmetic, currency and a table. The narrowest is the Instagram niche index at 59, which is a page of links.
What I take from this
- Move the character rules from the source test to the output. The existing test reads content source. It cannot see a dash that a markup helper inserts, a quote a component hardcodes, or a space that a library turns into a no-break space. A test that fetches the built pages and asserts that the visible text uses nothing outside a 97-character allowlist would have caught the en dash on the day it shipped, and would fail loudly the first time a typographic apostrophe arrives.
- Decide about the legal pages rather than discover them. Eight curly quotes and one en dash in two documents is either an acceptable register for legal copy or a bug. It should not be both, which is what it is today.
-
The absences are documentation. No
!, no*, no backtick and no curly apostrophe is a more precise description of a writing style than any style guide paragraph, and it is a one-command measurement.
Reproducing it on your own site takes about twenty lines: fetch the sitemap, parse out the visible text, and print a sorted Counter. The two numbers I would look at first are the count of distinct code points, which should be small and boring, and the list of printable ASCII you never use, which is where your own house style turns out to be hiding.
Top comments (0)