Text Similarity Checker

Three measures, because one number lies

Any single similarity percentage hides the thing you wanted to know. These three disagree with each other in useful ways, and the disagreement is the output worth reading.

• Cosine similarity, over term frequency vectors. Each text becomes a vector of word counts; the score is the cosine of the angle between them. It is about proportions — two texts that use the same words in the same ratios score high even if one is three times longer. Length-insensitive by construction, which is usually what you want.
• Jaccard index, over token sets. Distinct words in both, divided by distinct words in either. It ignores how often anything appears and asks only whether the vocabulary overlaps. It punishes a length difference hard: a long text has more distinct words, which inflates the denominator whatever it says.
• Longest shared phrase. The longest unbroken run of identical words. Not a percentage and not affected by length at all — it is the one that answers "was any of this copied rather than rewritten", and it is the one to look at first when a rewrite came back suspiciously fast.

Reading the disagreement

The interesting cases are the ones where the numbers pull apart, and the tool names which case you are in.

Cosine high, Jaccard low. The words they share carry most of the weight in both texts, but one of them has a lot of extra vocabulary the other does not. Usually a length difference: a short passage and the long article it was expanded into.

Jaccard high, cosine low. Nearly the same word list, used in very different proportions. Typically one text hammers a term the other mentions once. This is the pattern you get when someone "optimised" a draft by repeating the keyword.

Both high, long shared phrase. Passages were reused verbatim. For two pages on one site, go to the duplicate content checker, which will show you exactly which passages and highlight them in both.

Both high, short shared phrase. Same subject, same vocabulary, different sentences. This is what a genuine rewrite looks like, and it is the result you want when you asked someone to rewrite something rather than to spin it.

This tool or the duplicate checker?

Use this one when the question is "how far did this draft move" — reviewing a rewrite, comparing two versions of a page over time, checking whether an edit actually changed anything. It scores the whole of both texts.

Use the duplicate content checker when the question is "will these two pages compete in search". It hunts for shared passages rather than scoring the whole, highlights them in both texts so you know what to rewrite, and the copy on it explains what duplicate content does and does not cost you.

One thing neither of them does: judge quality. A rewrite can score exactly as you hoped on all three measures and still be worse than what it replaced.