Finding typos across an entire website

A customer reports a typo on the pricing page. You fix it. Two weeks later the next one arrives, this time from a blog post from 2022, and between those two reports there is, in experience, a whole stock of errors nobody knows about, because nobody proofreads 600 pages in one go. It gets expensive in two places: in the title Google puts into the snippet, and in the product name your customers search for.

How to go about it in JMX

1. Load a crawl. Enter the address at the top, Start. The check works on the loaded crawl. Nothing is fetched afterwards. Without a license every analysis stops at 25 pages — enough for a trial run, not for the whole stock.
2. Provide a dictionary. In the Spelling tab the Default dictionary sits at the top. JMX ships none but looks for the ones already on the machine — every LibreOffice installation brings some, under Linux /usr/share/hunspell, under macOS /Library/Spelling. If nothing turns up, pick a .dic file via Choose dictionary …; the .aff file of the same name has to sit beside it, otherwise the rules for inflection and compounds are missing.
3. Assign languages. Under Languages on this site stand all the languages of the crawl with their page counts, each with its own dictionary. Set anything you have no dictionary for to — do not check —. Pages without a lang attribute form a row of their own and run against the default dictionary.
4. Set the options. Under What gets checked, Check foreign words against a second dictionary hides the English inside German texts. Compute corrections is off by default, because a suggestion costs a multiple of the check.
5. Check. The list is not sorted by frequency but by how likely an entry is to be an error. The Classification column says so: up to two pages typo, from three pages vocabulary. A word on one page is a slip, the same word on forty pages is the client’s technical term.
6. Work from the top. The Context column shows the sentence fragment that lets you decide. Take brand and technical terms out with Accept word; the list applies across projects.
7. Pass it on. Export writes CSV or Excel for the editorial team. Only the words that look like typos appear in the client report.

What to watch out for

The classification vocabulary is not absolution but a probability. A misspelled word that a template spreads across the whole site sits on hundreds of pages and therefore gets the same classification as the correctly spelled brand name — so exactly the case that costs the most slips downward. Look through the top vocabulary rows once before you accept anything. A word once accepted is never queried again, in any project.

Hunspell checks whether a word is in the dictionary. Nothing more. “its” for “it’s”, “their” for “there”, a wrong inflection, a swapped pair of words: all valid words, none of them a finding. The foreign-word option sharpens that further, because it hides every German typo that happens to produce an English word — “Mist” produces no finding that way. It is off by default.

One limit belongs in the planning: the check runs in the interface. jmx-cli writes a report, a sitemap, a task list and the comparison with the previous run, but knows no switch for spelling. A nightly run reports a new noindex to you, not a new typo.

After the first pass the list of your own words is filled, and the second pass takes minutes instead of half an hour — whatever still shows up then was written since the last time.