How the text is cut into words
Anything that is not a letter, digit, apostrophe, hyphen or underscore is a boundary. Apostrophes and hyphens are kept when they sit inside a word and stripped from the ends, so don't and e-mail stay whole while a quoted 'word' loses its quotes. Case folding is on by default, which merges The at the start of a sentence with the everywhere else — worth turning off if you are looking at proper nouns and want Apple separate from apple.
There is no stemming and no lemmatisation. run, runs, running and ran are four different entries, and so are company and companies. That is a real limitation for topic analysis and an advantage for editing, because when you are hunting for a repeated word you usually want to know that you wrote running six times specifically, not that the lemma run occurs eleven times across four forms.
The stop-word list, and why it is on by default
Without filtering, the top of any English frequency list is the same list every time: the, of, and, to, a, in, is. Those words carry grammar rather than content, they occupy roughly the top thirty ranks of every English text ever written, and they tell you nothing about this particular one. The built-in list holds about 130 of them — articles, pronouns, auxiliaries, prepositions and conjunctions — and it is deliberately conservative, so no ordinary noun or verb is in it.
Turn it off when the function words are the subject: comparing writing styles, checking whether a draft leans on passive constructions, or working with text where not and no matter. The minimum length filter is the blunter alternative and works on any language — set it to four and most English function words disappear without a list.
Frequency is not importance
The word at rank one is the word you used most, which is not the same as the word your text is about. English word frequencies follow a rough Zipf distribution: the most common word appears about twice as often as the second, three times as often as the third, and so on. That means the top few entries are dominated by whatever is structurally frequent rather than by what is significant, and it is entries five through twenty, after filtering, that usually describe the content.
For editing, the interesting reading is different again. Look for a distinctive word — an unusual adjective, a piece of jargon, a metaphor — with a count above three or four in a short piece. That is the crutch, and it will be invisible on a read-through because each individual use felt justified at the time.
The type-token ratio and its trap
Distinct words divided by counted words. Closer to one means less repetition. It is a genuinely useful number with one property that trips people up constantly: it falls as text gets longer, no matter who wrote it. A vocabulary is finite, so a 5,000-word essay reuses words a 500-word one never gets the chance to. A ratio of 0.62 on a short paragraph and 0.31 on a long chapter says nothing about the two authors. Compare it only between texts of similar length — two drafts of the same document, or two candidates answering the same question — and treat it as a comparison, never as a score.
Scripts without spaces
Chinese, Japanese and Thai do not separate words with spaces, and this tool has no word segmenter. Those characters are treated as word content, so a run of CJK text between two punctuation marks becomes one enormous token instead of the several words it contains, and the frequency table is meaningless. Korean and other space-separated scripts tokenise into whitespace-separated chunks, which is closer to correct but still counts inflected forms separately — an agglutinative language produces a distinct token for every particle attached to the same stem. For any of those, a proper morphological analyser is the right tool, and this one will only mislead you. English and other space-separated Latin-script languages are what it is built for. Nothing is uploaded either way; the counting all happens in the page.
Questions people ask
Are plurals and verb forms counted together?
No. There is no stemming or lemmatisation, so cat and cats are two entries, and run, runs, running and ran are four. Only case folding merges anything, and only when the ignore-case option is on. For topic modelling that is a real weakness and you would want a tool with a lemmatiser. For catching a repeated word in a draft it is usually what you want, since knowing you wrote leveraging five times is more actionable than knowing the lemma leverage occurs eight times spread across four forms.
Can I change the stop-word list?
Not directly — the list is fixed at about 130 common English function words. The workaround is the minimum length filter, which drops short words regardless of what they are and removes most function words at a setting of four or more. If you need a specific custom list, export the tab-separated table into a spreadsheet and filter it there; the export includes every word in the top N with its count and share.
How do I get the results into Excel or Sheets?
The tab-separated block near the bottom is built for exactly that. Copy it, paste into any spreadsheet, and the three columns — word, count, share — split into separate columns automatically because the separator is a real tab character. It includes as many rows as you set in "how many to list", up to 500. From there you can sort, filter, or diff it against a frequency list from another document.
Does it work on Chinese or Japanese text?
Not usefully. Neither language separates words with spaces and this tool has no segmenter, so an entire run of characters between two punctuation marks is counted as a single token. The result looks like a frequency table but is not one. Korean and other space-separated languages tokenise into chunks that are closer to words but still count each inflected form separately. A morphological analyser is the correct tool for any of those; this one is built for English and other space-separated Latin-script text.
What is the size limit?
It stops at 500,000 characters, roughly an 80,000-word manuscript, and tells you rather than freezing the tab. Below that the count is effectively instant since it is a single pass with a hash map. Everything runs in your browser and nothing is transmitted, so an unpublished draft or an internal document is safe to paste in.