Two drafts, and one paragraph you cannot place
You have merged two writers' sections into a single chapter and one paragraph reads like a sentence you have seen before. Or you have rewritten a supplier's product description in your own words and want to know how much of their phrasing is still sitting in yours. Neither question has a yes or no answer, and a tool that returned one would be lying to you.
This one returns four numbers and the list of three-word runs the two texts share. It measures overlap in vocabulary and phrasing. It cannot see who wrote what first, whether anything was copied, or what either writer intended, and the sooner you accept that framing the more useful the output becomes.
What each of the four measures actually compares
Paste one text on each side and you get four percentages. They are not four opinions about the same thing; each is blind to something the others can see.
| Measure | What it compares | Blind to |
|---|---|---|
| Word-set Jaccard | Unique words present in both, over all unique words present in either | Word order, repetition |
| Frequency cosine | How often each shared word appears in each text | Word order |
| Word-set Dice | The same overlap, weighted toward shorter texts | Word order, repetition |
| Three-word phrase Jaccard | Runs of three consecutive words found in both | Very little |
The first three are bag-of-words measures: they tip the text into a container, shake it, and count. Only the phrase measure cares about sequence. That is the most useful thing to understand on this page, because it explains nearly every surprising number you will ever see here.
Dice always returns a higher figure than Jaccard on the same pair, since it doubles the shared count before dividing. Both are legitimate, simply tuned differently, and a 40% Jaccard commonly arrives as a 57% Dice.
Why the phrase measure is the one to watch
Every text is broken into words first, ignoring case and punctuation, then into every run of three consecutive words. Two texts that both contain "morning practice improves" share that run, and every shared run is listed under the scores.
The measure is strict about order. Take these two texts:
The sentences are identical, merely swapped. Word-set Jaccard, cosine, and Dice all report 100%, because the same six words appear in both. Phrase Jaccard reports 33.3%, because only two of the six three-word runs survive the reordering. Same vocabulary, different text, and only one of the four numbers noticed.
A smaller pair makes the same point from the other direction. Hello world against Hello world scores 100% on the three word measures and 0.0% on phrases, because two words contain no run of three. Verbatim copying of a short string is invisible to the sharpest measure in the set.
What a high score does and does not mean
Two writers reaching for the same stock phrase is ordinary. "In the current economic climate" belongs to nobody. Function words such as the, and, of, and to appear in nearly every English sentence and count as shared vocabulary like any other word, which inflates the bag-of-words figures on texts that share nothing meaningful.
The reverse error is just as common. A genuine paraphrase that swaps every content word and rebuilds the sentences drops to near zero on all four measures, and writers sometimes read that as proof of originality. It is not. Restructured text carrying a borrowed argument passes cleanly and always will, because these measures only ever see surface vocabulary. What a high score establishes is shared phrasing, and shared phrasing is a question to investigate rather than a finding to report.
Where the numbers earn their keep
Three jobs make this tool worth opening.
Auditing your own paraphrase. Put the source on one side and your rewrite on the other, then read the shared phrase list. Any run of three or more words appearing in both is a phrase you did not actually rewrite, and the list names exactly which ones to attack. This is the practical companion to a proper method for paraphrasing without plagiarizing.
Comparing two versions of a draft. Paste the version your editor returned beside your original. Word-set Jaccard at 90% with phrase Jaccard at 60% says the edit was light and largely cosmetic. A large drop in phrase score with a modest drop in word overlap means the sentences were rebuilt rather than polished.
Finding accidental repetition in a long document. Split one document at its midpoint and compare the halves. A suspicious cluster of shared phrases usually means a section was pasted twice, or a point was made twice in nearly the same words.
When the question is whether a specific sentence was lifted from a source you already have, the plagiarism checker answers it sentence by sentence rather than as a vocabulary average.
Limits that decide your results
Tokenization is stricter than you might expect. Hyphens and apostrophes are part of a word, so well-known and well known are different tokens, and so are don't, dont, and do not. The effect is larger than it sounds. Take "The well-known author wrote a much-loved book that readers still enjoy" and drop only its two hyphens. The two sentences read identically, and the first is eleven tokens against the second's thirteen, but the word-set Jaccard between them is 60%, because six of the twenty-four slots no longer line up. Nothing about the meaning changed; only the tokenizer's view of it did.
Short inputs behave oddly. Under three words there are no phrases to share, so the phrase score reads 0.0% no matter how identical the texts are. An empty field returns 0.0% across the board, which looks like a clean result and is really an absent comparison.
The tool compares vocabulary, not argument, not structure, and not authorship. A translation of a passage scores near zero against its original, because the words are different even when every idea came across intact.
Frequently Asked Questions
Does a high similarity score prove one text was derived from the other?
No. It proves the two share words and, if the phrase score is high, phrasing. Independent writers reach for the same stock sentences, and quoted material is legitimately shared. Read the phrase list to see what actually matched before drawing any conclusion.
Why do two texts about the same topic score so high when they share no content words?
Function words. "The," "and," "of," "to," and "in" appear in nearly every sentence, count as shared vocabulary, and carry real weight in the frequency-weighted measures. That is why the word measures run high on unrelated texts, and why the phrase list deserves more of your attention than the percentages do.
Why is my phrase score 0.0% when the two texts are identical?
Almost certainly because they are too short. Phrase runs are three words long, so a text of fewer than three words produces no runs at all and the phrase score is 0.0%. Note the edge just above that line: a three-word text produces exactly one run, so two identical three-word texts score 100% on phrases rather than zero. The three word measures show 100% either way.
Why did my score drop when I removed the hyphens?
Because hyphens and apostrophes are treated as part of the word. well-known is one token and well known is two, so the sets stop matching even though a reader sees no difference. Make the hyphenation consistent across both texts before you compare them.
Can I compare a translation against its original?
Not meaningfully. The comparison works on shared words, and a translation shares almost none with its source. A near-zero score says nothing about whether the translation is faithful, borrowed, or excellent.
Compare one paragraph, then edit
The useful habit is narrow: put one paragraph against one source passage, note the phrase score, rewrite the borrowed runs, then compare again. Watching a specific number fall as you work tells you when a paraphrase has genuinely changed rather than merely shuffled, and that is a signal no single headline percentage can give you.