Word Frequency Counter
Every repeated word in your draft, ranked by count and updated live — the mirror that shows you the words you lean on without noticing.
Keywords appear once you type.
How the table is built
The table is the output of a short pipeline, and every row can be traced through it. First the text is split into words by the site's tokenization rule — a word is a run of letters and digits in which internal hyphens and apostrophes do not break the run, so "well-known" and "editor's" are one token each; the what counts as a word guide walks through the rule and its edge cases. Second, every token is lower-cased, so "Draft" at the start of a sentence and "draft" in the middle of one land in the same row. Third, two kinds of token are set aside: a fixed list of 74 common English function words ("the", "and", "of", "was"…) and any single-character token. Fourth, what remains is counted. Fifth, the rows are sorted by count, highest first, and rows with equal counts are ordered alphabetically. The live table shows the ten highest rows; the worked examples below use the full ranking wherever the cutoff would hide something worth seeing.
The density column is the same count expressed as a share of all words in the text — stop words and single characters included in the base, even though they never appear as rows. That is the convention the keyword density checker uses, and the keyword density guide explains why the base matters whenever two tools disagree. On this page the counts are the point; the percentage is there so you can tell whether a row is heavy for the length of the text or merely heavy because the text is long.
A worked example: case, ties, and the words you never see
Paste this passage: "Draft the plan today. Review the draft tomorrow, and revise the plan after the review. A good plan survives its first review." The counter reports 22 words and 14 unique, and the ranking begins "plan" 3 (13.64%), "review" 3 (13.64%), "draft" 2 (9.09%). Three things in that list are worth noticing. "Draft" opens the passage with a capital and comes back lower-cased; both spellings sit in one row because ranking is case-insensitive. "plan" and "review" have identical counts, and "plan" is listed first only because it comes first alphabetically — the order inside a tie carries no information about which word is the bigger habit. And the passage's most visible repetition, "the", is absent: of the 22 words, 15 appear in the ranking and the other 7 are stop words or single characters ("the", "and", "its", "a"), counted in the totals but never listed.
The density figures show why the base matters. "plan" is 3 ÷ 22 = 13.64% because the base is the whole passage. A tool that removed stop words before dividing would put the same word at 20% of the 15 words it kept. Neither number is wrong, but they are not comparable, and a percentage quoted without its base cannot be checked against this table.
No stemming, on purpose
Nothing beyond case is merged. There is no stemming and no lemmatization, so "run", "runs", "running", and "ran" occupy separate rows. Take this passage: "She runs every morning. Running clears her head, and a long run on Sunday resets the week. Last year she ran her first marathon, and running still feels new to her. A short run before work beats none." It has 38 words, and the ranking begins "run" 2 and "running" 2 — a tie, alphabetically ordered — with "runs" and "ran" at 1 and 1 further down. Add the four rows and the verb family accounts for 6 of 38 words, 15.79% of the passage, which no single row announces.
That is the trade the counter makes deliberately. A merged row would report the family total but hide its shape — whether the habit is the gerund, the past tense, or the bare verb — and it would have to decide silently that "ran" belongs with "run" while "runner" does not. Separate rows put those decisions in front of you: when several forms of one verb chart at once, that verb is carrying too much of the prose, and the form at the top tells you which construction to vary. The cost is a little arithmetic when you want the family total, which is why the examples here add the rows for you.
The same passage shows a limit of the live table on short texts. The full ranking has 24 rows, but the table shows ten, and after the two tied leaders every remaining word appears once — so the alphabetical tie-break, not repetition, decides which words fill the visible rows. Only 2 of the four verb forms make the cut. On a draft of ordinary length the words that appear once fall well below the cutoff and the visible rows are all genuine repeats; on a single paragraph, read the order of the one-count rows as alphabetical, because that is all it is.
Editing with the table: a draft, before and after
The workflow that earns its keep: paste a finished draft, read the visible rows, and ask of each one whether the word is doing necessary work every time it appears or is a habit. Here is a short draft with a habit in it: "The results were really encouraging. The team was really pleased, and the client was very happy with the very first version. It was just a really good week." The counter reports 28 words and 20 unique, and the ranking leads with "really" 3 (10.71%) and "very" 2 (7.14%); everything after those two appears once. The full ranking has 12 rows, so 2 of them sit below the table's cutoff — harmless here, since they are all single occurrences.
Now the same paragraph with the intensifiers removed: "The results were encouraging. The team was pleased, and the client was happy with the first version. It was a good week." The counter reports 22 words and 17 unique; "really" and "very" no longer chart, and every remaining row shows a count of 1. Two numbers moved in the direction you want. The word count fell by 6 with no loss of meaning, and the unique-to-total ratio rose from 0.71 to 0.77, because the words removed were the repeated ones. Re-pasting after each pass and watching the top rows flatten is the whole method; the table cannot tell you which repetitions are deliberate, but it reliably shows you which ones exist.
One word in the draft never appeared in the table at all. "just" is on the stop list, along with "like", "only", "also", "more", and "most", so the counter will not surface it however often it recurs. The list is fixed and was chosen to remove function words, but a few of the words it removes are also common verbal habits. If one of those is yours, use your editor's find function for it; the table cannot help.
Reading the unique-word number
"Unique words" is the number of distinct words after case-folding, and unlike the ranking it includes stop words and single characters — which is why the first passage above shows 14 unique words but only 10 rows. Divide unique by total and you get a rough lexical-variety gauge. This paragraph: "A lighthouse keeper logs the weather twice a day. Wind, cloud, and swell go into the book before breakfast, and again before dark. The entries are short, but the keeper never skips one, because a missing line is the first thing an inspector notices." has 44 words and 35 unique, a ratio of 0.80.
The caveat is the important part: the ratio falls mechanically as texts grow, because common words must repeat. Paste the same paragraph twice and the counter reports 88 words and still 35 unique — the ratio halves to 0.40 although not one new word was written. Combine the four unrelated passages on this page into one text and you get 132 words and 84 unique, a ratio of 0.64 — below the lighthouse paragraph alone — and the top row of the combined ranking is "first" at 4, a word no single passage uses more than once. On long texts the top of the table can be filled by incidental words that accumulated, not words anyone chose.
So never compare the ratio across texts of very different lengths, and do not treat any value as a target. Do compare it across drafts of the same piece, where a rising ratio after an editing pass is direct evidence the vocabulary loosened, as the before-and-after example shows.
Where the table misleads
Several habits of the tokenizer are worth knowing before you act on a row. Possessives and contractions are separate rows from their base words: "keeper's" and "keeper" do not merge, so a noun used partly in the possessive is under-reported by the row you would look at. Hyphenated compounds rank whole, so "well-known" never contributes to "known". Numbers chart like any other token, while a single digit does not: "2,500" tokenizes as "2" and "500", the first excluded as a single character and the second listed as a word. Case-folding merges proper nouns with common nouns — "Bill" and "bill" share a row — which is invisible in the table and only matters when a name coincides with an ordinary word.
Structural limits matter as much as tokenization ones. The table shows single words, so a two-word phrase repeated throughout a draft appears only as its halves ranking near each other. The stop list is English; the tokenizer handles accented and non-Latin scripts, so counting works in other languages, but their function words are not filtered and will fill the top rows. And the density column is a description, not a target: the keyword density guide covers what the percentage can and cannot tell you. When a row surprises you, open the draft and read the sentences the word appears in — the table finds the repetition; only reading decides whether it is a fault.
Frequently asked questions
What should I look for in the frequency list?
Your habits. Intensifiers ("really", "very"), hedges ("perhaps", "somewhat"), and pet adjectives repeat unnoticed while you write and stand out immediately in a ranked list. In the worked draft above, "really" charts 3 times in 28 words. A word you expected to see near the top is your subject; a word you did not expect is worth a second look.
How is this different from keyword density?
Same engine, same ranking, different emphasis. The keyword density checker leads with each word's share of all words, which is the convention for comparing a draft against a brief; this page leads with raw counts and the unique-word total, which is what an editing pass needs. Both tables exclude the same 74 stop words and single characters, and both divide density by all words.
Are word forms combined?
No. "run", "runs", "running", and "ran" count separately; only case is folded. In the running example above the four forms add up to 6 of 38 words, but they appear as four rows, because merging them would hide which form is the habit.
What does the unique-word number include?
Every distinct word after case-folding, stop words and single characters included, so it can exceed the number of rows in the table. Divide it by total words for a rough variety gauge, then compare that ratio only across drafts of the same piece: it falls automatically as a text gets longer, so a paragraph and a chapter cannot be compared by it.
Why is a word I know I overuse missing from the table?
It is probably on the stop list. The 74 excluded function words include a few habitual ones — "just", "like", "only", "also", "more", "most" — and single-letter tokens such as "I" and "a" are excluded too. They still count toward the word total; they are never listed. Search the draft for them directly.
Two words have the same count. Why is one listed first?
Ties are broken alphabetically. Rows are sorted by count, highest first, and rows with equal counts are ordered by the word itself, so the order within a tie says nothing about which word matters more. On short texts most rows tie at one occurrence, and the visible order is simply alphabetical.
Why do numbers and hyphenated words show up?
Because they are words to the tokenizer: a run of letters and digits, with internal hyphens and apostrophes keeping the run whole. "2024" ranks like any word, "state-of-the-art" ranks as one row, and "editor's" is a different row from "editor". The what-counts-as-a-word guide lists the rule and its edge cases.
Does it work in languages other than English?
The counting does — the tokenizer is Unicode-aware, so accented and non-Latin words are counted correctly — but the stop list is English, so in another language the most common function words are not filtered and will rank at the top.
How much text can I paste?
Up to 2,000,000 characters, processed entirely in your browser; nothing is uploaded.
Analysis happens locally in your browser — drafts are never transmitted or stored. See the methodology page.