What SnapTextCompare Compares
SnapTextCompare is a text comparison tool: it takes two versions of the same content and highlights every addition, deletion and modification. Paste text, or upload PDF, DOCX and TXT files — the extraction and the comparison both happen locally, so your content is never sent to our servers. That makes it usable for contracts, proprietary code and internal documentation.
Use it to compare document drafts, review contract revisions, check code changes before merging, or verify what an AI model rewrote.
Comparing AI Output (ChatGPT, Claude, Gemini)
One of the fastest-growing reasons people compare text today has nothing to do with contracts or code: it's checking what a language model actually changed. If you ask ChatGPT to "tighten this paragraph" or Claude to "rewrite this in plain English", the response looks plausible — but plausible is not the same as faithful.
Three concrete workflows people run here every day:
- Prompt A vs. prompt B — Run the same source text through two different prompts and diff the results to see which instruction actually moved the output.
- Model vs. model — Paste a ChatGPT answer on the left and a Claude or Gemini answer on the right. The diff shows exactly where the outputs diverge, so you know which claims to verify yourself. Agreement between two models is not evidence that either is correct.
- Original vs. AI edit — The most important one. Diff your own draft against the AI's revision before you accept it. Changes to numbers, dates, names, and citations are usually the ones worth checking most carefully, and our importance scoring surfaces those categories first.
Because everything runs locally, you can do this with unpublished drafts, client material, or internal documentation without sending a single byte to another AI service.
How to Read the Similarity Score
The similarity percentage comes from a normalised Levenshtein edit distance: the minimum number of character insertions, deletions, and substitutions needed to turn the first text into the second, divided by the length of the longer text. It is a character-level measure, not a meaning-level one — which is why we always pair it with a plain-language explanation.
- 99–100% — Effectively identical. Usually whitespace, a stray character, or an invisible encoding difference.
- 95–99% — Almost identical. Minor wording or punctuation edits; skim the highlights and move on.
- 80–95% — Noticeable edits with the structure intact. Typical of a real editing pass.
- 50–80% — Substantial rewriting. Read the whole diff rather than trusting the score.
- Below 50% — Two largely different texts. The percentage stops being useful; use the change breakdown instead.
A high similarity score never means "safe to approve". A contract can be 99.4% identical and still have had its payment terms flipped from 30 to 90 days. That single-digit difference is why we score importance separately from similarity, and why numbers, dates, currency amounts, and named parties are weighted far above formatting noise.
Word, Line, or Character — Which Mode to Use
Picking the wrong granularity is the most common reason a diff looks noisy. As a rule of thumb:
- Word mode for prose, marketing copy, contracts, and anything a human wrote in paragraphs. Reflowed line breaks won't create false positives.
- Line mode for source code, CSV rows, logs, configuration files, and JSON — anywhere a line is a meaningful unit and order matters.
- Character highlighting on top of either mode when you need to catch a single transposed digit, a changed currency symbol, or an altered URL slug.
Every mode runs through Google's diff-match-patch library in your browser. Word mode tokenises the text and diffs the tokens, line mode encodes whole lines, and character mode diffs raw characters — the same underlying shortest-edit-script diff in all three cases, so moved sentences and reflowed paragraphs stay readable instead of collapsing into one giant red-and-green block.