Free text compare & diff

A diff and a similarity score answer two different questions

People reach for a text comparison tool when they remember that something changed but can't point to where. The diff answers that exactly: it walks both versions line by line and shows you the surviving lines for context, the lines that vanished, and the lines that are new. Reading it is less about staring at two walls of text and more about scanning a single coloured map of the edit — green for what was added, red for what was removed, plain for what stayed put.

The similarity percentage answers a coarser question: roughly how much of the wording overlaps at all. A pair of paragraphs that share most of their vocabulary will sit high; two genuinely different drafts on the same topic will sit lower than people expect, because shared subject matter doesn't mean shared phrasing. Use the score for a gut-check on how far apart two versions have drifted, and use the line diff when you need the actual list of changes.

How the diff is built — and why nothing leaves the tab

The line diff uses a longest-common-subsequence pass: it finds the longest ordered run of lines the two versions agree on, treats those as the unchanged backbone, and labels everything between them as added or removed. That's the same idea behind git diff and the review panes in most editors, which is why the output reads the way developers already expect. The similarity figure is a Jaccard index over the set of words — the count of words both texts share divided by the size of their combined vocabulary — so it ignores ordering and rewards genuine overlap rather than coincidental line breaks.

Both of those run as JavaScript in your own browser the moment you press Compare. Unlike our AI detector or humanizer, which send text to TextSight's API to score it, the compare tool never makes a network request — your two drafts stay in the page and are gone when you close the tab. That's a deliberate fit for the job: people most often compare things they'd rather not upload, like a contract redline, an NDA, a config file or an unpublished manuscript.

When pasting two versions beats hunting by eye

The tool is line-based and language-agnostic, so it doesn't care whether you feed it prose, code, JSON or a CSV dump. A few situations come up again and again:

• Catching silent edits in a contract or brief — paste the version you signed off on against the one that came back, and the added or removed clauses surface instead of hiding mid-paragraph.
• Comparing two drafts of an essay or article — see exactly what an editor (or you, at 2 a.m.) changed between revisions, rather than re-reading the whole thing.
• Spotting near-duplicate copy — run two product descriptions or landing pages and let the similarity score tell you how much they actually overlap before you publish both.
• Reviewing a snippet of code or config — drop in two versions of a file when you don't have version control handy and want a quick added/removed view.
• Checking what a rewrite really touched — compare your original against a paraphrased or AI-assisted version to see whether the meaning moved or just the wording.

What the score won't tell you

A couple of things are worth knowing so you don't over-read the result. Because the diff aligns on whole lines, reflowing a paragraph — say, joining two short lines into one long one — can show up as a removal plus an addition even though almost every word survived. If a comparison looks noisier than the change felt, that's usually why; the similarity percentage, which works on words rather than lines, is the better signal in that case.

It's also worth remembering what Jaccard does and doesn't measure: it's pure vocabulary overlap, blind to meaning. Two passages can say opposite things with nearly identical words and score high, while a faithful paraphrase that swaps in synonyms can score lower than you'd guess. Treat the number as a quick distance gauge between versions, not a verdict on whether the meaning held — and when the real question is "does this read as AI-written?", that's a job for the AI detector, not the diff.