Nakodo's 76 public pages are written by hand, by me, in one sitting each, with no CMS and no editor looking at them. There are unit tests for the house style, including one that bans the em dash, and ten rules measured against all 76 pages, and not one of them has ever checked whether a word is spelled correctly.
So I ran the whole site through a dictionary.
$ while read u; do curl -s "$u" -o "$(...)"; done < /tmp/nak_urls.txt
$ python3 spell.py
pages 76 tokens 123880 distinct 6166 unknown_forms 485 unknown_tokens 8782
longest how-it-works 3,274 words
shortest alternatives 400 words
median 1,683 words
123,880 tokens, of which 49 are ÷ and × from the calculator pages that my own regex thought were letters, so 123,831 words. 6,166 distinct forms. 485 of them are not in the dictionary, which is 7.1 per cent of all the text on the site.
Then I read all 485.
The dictionary is the problem, not the words
The list on a Mac is /usr/share/dict/words, which is Webster's Second from 1934: 235,976 entries, no inflections, no contractions, nothing newer than the Hoover administration. Out of the box it flags email, website, apps, inbox, online and doesn't.
Three cheap fixes get the noise down to something readable:
def stem_known(w):
if w in dict_words: return True
rules = [("ies","y"),("ied","y"),("es",""),("s",""),("ed",""),("ed","e"),("ing",""),("ing","e"),
("ly",""),("er",""),("est",""),("ness",""),("ment","")]
for suf, repl in rules:
if w.endswith(suf) and len(w) > len(suf) + 1:
c = w[:-len(suf)] + repl
if c in dict_words: return True
if len(c) > 2 and c[-1] == c[-2] and c[:-1] in dict_words: return True # planning -> plan
return False
A hyphenated word counts as known if every part is known, and a contraction counts as known if the part before the apostrophe is. That last rule is why doesn't passes and also why I would not trust this tool to catch its where it's belongs. A dictionary check is about characters, not grammar.
Even after that, paid, held, heard, began, feet and women are all unknown, because an irregular form is not a suffix away from its lemma. Any spell check you build yourself has a floor of noise, and the useful question is not "how many unknown words" but "what are they".
What the 485 actually are
Grouped by hand, with the number of occurrences on the site:
platform and product names TikTok 699, Instagram 656, YouTube 618, Shopify 117,
Gmail 15, Linktree 3, BigCommerce 2, Magento 3
our own names Nakodo 658, Nakodo's 39, CogniPrep 84, Alcy 1,
Notifio 2, PubTrivia 2
words for a platform's people YouTubers 644, TikTokers 82, Instagrammers 81,
YouTuber's 96, TikToker's 15
acronyms API 104, CPM 99, FTC 47, UGC 38, CMA 35, UTM 22,
DIY 18, GDPR 6, PECR 1, CSV 4, GRWM 5
post-1934 vocabulary email 528, website 221, apps 107, bio 73,
keywords 70, inbox 35, spreadsheet 24,
screenshots 22, workflow 19, captcha 15,
hashtags 14, emoji 8, deliverability 5
platform jargon carousels 23, Collabs 22, DM 32, link-in-bio 33,
BookTok 24, CleanTok 18, SkinTok 5, spon 1
niche nouns espresso 18, skincare 14, sourdough 5, romantasy 4,
aquascaping 2, mukbang 1, cafetière 1, moka 3
competitors we compare with Modash 36, Upfluence 35, HypeAuditor 28, CreatorIQ 2
British spellings labelled 27, programme 20, colour 8, licence 6,
authorise 6, behaviour 5, jewellery 4, fulfil 3
one Japanese word nakōdo 77
Zero misspellings. In 123,831 words of hand-written marketing copy, guides, a glossary, a privacy notice and terms of service, the dictionary found nothing that was simply wrong.
I did not expect that, and I am not sure it is a compliment to me rather than to the autocorrect in the editor I write in.
The four that looked like leaks
The useful part of the long tail is the single occurrences, because those are where an accident hides. Four of them stopped me.
formuladesk, once, on the home page:
<span class="font-mono text-xs font-normal text-muted-foreground">@formuladesk</span>
A made-up handle in the example cards. Fine, and worth checking, because a real handle in sample data is a different kind of mistake.
Co-authored, once, on the weddings page. In a repository where every commit ends with a Co-authored-By trailer, that word in rendered HTML is exactly what a leaked commit message looks like. It is a sentence about Instagram's collab posts: "Co-authored by up to five accounts".
Watson and Glaser, on the about page, which are a psychometric test named after two people, on a page that mentions the other apps the same company publishes.
Sp, once, inside a table of disclosure wording that advertising regulators do not accept: "Sp", "spon", "collab". A two-letter word in a dictionary sweep is almost always a defect. This one is the subject of its own sentence.
Four false alarms out of 485, all resolved by grepping the page they appear on. That ratio is why I would run this again: the cost is reading a list once, and the thing it is really auditing is sample data and quoted text rather than spelling.
Two real inconsistencies
cafés 6, cafes 4. The same word, the same product, both spellings live:
/privacy "find businesses, such as cafes, shops or gyms"
/how-it-works "cafes, shops, gyms, salons or any other kind"
/how-it-works "such as cafes within 10 km of Leeds"
elsewhere "cafés"
No reader will mind and it is still wrong, because the accent is either part of how we spell the word or it is not. I am going with the accent, which means the three unaccented ones change and the editorial test gets a rule: cafes is not a word we use.
obs 1, lowercase, in a list of example search phrases on the gaming page: "how to start streaming with obs". OBS is an initialism. In a list of things people type into a search box, lowercase is defensible, and the fix is to make that a deliberate choice rather than an accident, since the same list capitalises other products.
102 British spellings, and not one American one of our own
The site is written in British English, which was never a decision, only a habit. Measured:
british n american n
labelled 27 labeled 0
programme 20 program 4
colour 8 color 0
licence 6 license 1
authorise 6 authorize 0
behaviour 5 behavior 0
jewellery 4 jewelry 0
fulfil 3 fulfill 0
recognise 2 recognize 0
organise/-d/-ation 3 organize/-d/-ation 0
analyse 1 analyze 1
centre 1 center 0
cafés 6 cafes 4
instalment, offence, moisturiser, itemised: 1 each, 0 each
totals 102 10
That excludes while, check and practice, which are the right words in both dialects and which outnumber everything else on the page (40, 251 and 11). Every one of the ten hits in the American column has an explanation:
- all four
programs are proper nouns we are quoting: "Amazon influencer program", Google Play's "developer program", a URL withaffiliate-programin it - the one
analyzeis inside a verbatim quotation of a competitor's own homepage copy, which is American and must stay American - the four
cafesare the missing accent from the section above - and
licenseis correct in British English as a verb: the noun is a licence, the verb is to license, and the sentence is "a creator can also license the video to you". The site gets that distinction right in both directions, six nouns and one verb, which I am going to claim was on purpose.
whilst appears zero times against 40 while, and cheque zero against 251 check. So the register is British spelling without British archaism, which is what I would have asked for if anybody had asked me.
The attribute that is wrong
<html lang="en" ...>
Seventy-six pages, every one of them lang="en", and a body of text with 102 British spellings in it. en is not wrong exactly, it is unspecific, and the thing that consumes it is a screen reader choosing a voice and a pronunciation dictionary. "Jewellery", "programme" and "labelled" read better in an en-GB voice, and prices on the pricing page are a British company's prices in local currency.
Changing it to en-GB is one attribute in one root layout. It is the single cheapest accessibility improvement I have found in this app, and I only found it by counting spellings.
The only inline language annotation on the site is Japanese
Here is the thing that made this worth writing. A count of every lang attribute across all 76 pages:
76 lang="en"
77 lang="ja"
More Japanese annotations than English ones. They come from the footer, which is on every page:
<p><span lang="ja">仲人</span>, nakōdo: Japanese for the go-between who makes the introduction.</p>
Two kanji, marked up correctly, 76 times, plus one more on the about page where the name is explained a second time. The app's whole inline-language markup exists to tell a screen reader how to pronounce two characters in a footnote about its own name, while the 123,831 words around them are declared as generic English.
That is the shape of most accessibility work I have done: something small and careful sitting next to something large and unexamined.
Those two characters are also the only CJK on the site, and they are invisible to the audit above, because my tokeniser only matches Latin letters. The word count in the title excludes them. A spell check is a dictionary and a regex, and the regex decides what counts as a word before the dictionary gets an opinion.
While I was in there
The house-style tests only cover the structured content pages, so I checked the two most absolute rules against the rendered HTML of all 76, including the hand-written ones:
em dashes 0
exclamation marks 0
en dashes 1
The one en dash is on the privacy page, in "the UK–US data bridge", which is the correct punctuation for a range of two things and the sort of thing you only find by counting.
What I am changing
-
lang="en-GB"on the root layout. -
caféseverywhere, and a banned-word entry forcafes. -
OBScapitalised, or a note saying the search-phrase lists are lowercase on purpose. - The 485-word list, saved as a project vocabulary, so the next sweep only has to read what is new.
The sweep itself is about forty lines of Python and ten minutes of reading. If you have a site with hand-written pages on it and no spell check anywhere in the pipeline, which is nearly everyone, the surprising part is not the spelling. It is the sample data, the quoted copy, and the attributes you never set.
Go and look at any page: the glossary, a guide, the blog. The footer on every one of them is where the Japanese is.
Top comments (0)