"What changed between these two versions?" is the question behind most PDF anxiety — the vendor who "just updated the dates," the policy revision nobody summarized, the contract that came back "clean." The good news: when the words are what matter, a text diff answers it precisely. The catch: knowing what a diff can't see.
Every text-based PDF (one produced by Word, LaTeX, Google Docs — anything that isn't a scan) carries a text layer: the actual characters, in reading order, underneath the visual page. A PDF diff tool extracts that layer page by page, pairs the pages of the two versions, and runs the same word-level comparison developers use on source code. Deletions get struck out on one side, insertions highlighted on the other, and a summary counts what moved.
The pairing deserves a word, because it shapes everything you see. Page 3 of the old file compares against page 3 of the new one. That's the only sane default for text-only comparison — there's no anchor beyond position — and it has a consequence you should recognize on sight: if the new version inserts a page near the front, every later pair shifts by one, and the diff shows huge delete-plus-add blocks. That's not hundreds of edits; it's one structural change. The tell is the page-count summary: "pages only in one file" greater than zero means pairing shifted.
| Change | Text diff catches it? | Notes |
|---|---|---|
| Wording, numbers, dates | Yes | The core case — precise, word-level |
| Inserted or deleted paragraphs | Yes | Shown as add/remove blocks |
| Inserted or dropped pages | Yes | Counted in the summary; shifts later pairs |
| Font, color, spacing changes | No | Same words, diff shows nothing |
| Image or chart edits | No | Images have no text layer |
| Scanned pages | No | No text layer to extract — OCR first |
So: contracts, terms of service, policies, financial statements, documentation — diff the text. Brand guidelines, design proofs, layout QA — you need eyes or a visual comparison, because what changed isn't the words.
You mostly don't, not directly. A scan is a picture of text; unless someone ran OCR on it, the PDF has no characters to extract and a text diff correctly reports nothing to compare. The workflow is OCR both files first (Adobe's "Recognize Text," an OCR tool, or your scanner's software), then diff the results — and expect a little noise, because OCR introduces its own small errors that show up as spurious one-character changes. When precision matters, a human still reads the flagged regions.
Text diff is exactly the right instrument, with a workflow around it. Diff the version you negotiated against the version you received. Read the removals first — deleted clauses are where risk hides, and strikethrough makes them impossible to miss. Then read the additions, especially numbers: rates, dates, amounts, index references. A good diff turns "they said nothing changed" into "three words changed on page 4," which is precisely the conversation you want to have. What you do about a change is a legal question and yours to pursue — the diff is an inspection instrument, not advice.
It can be — check where the work happens. A diff that runs entirely in the page (extracting text with pdf.js, comparing with a JavaScript diff library, rendering highlights locally) never transmits either file, works offline once loaded, and leaves no copy on anyone's server. "Upload both documents to compare" services do exactly what the name says. For contracts, medical records, or anything under an NDA, the in-browser kind is the only category worth considering. Our Compare PDFs tool is the in-browser kind: word- or line-level granularity, side-by-side page pairs, and a change summary, with nothing uploaded.
If you already have both texts out of their PDFs, the lighter-weight Diff Checker compares pasted text directly, and developers comparing structured data have the JSON Diff for the same job on documents.
Three habits separate a useful review from noise. Start with the summary numbers, not the panes — words removed and added per page tell you where to read closely, and page-count mismatches tell you where pairing shifted. Use word granularity for dense legal text and switch to line granularity when reflow makes the word diff sparkle across whole paragraphs. And treat "no differences found" as a claim to verify once — re-run it on extracted text or spot-check by hand — rather than a guarantee that nothing moved, since the blind spots (formatting, images) are silent by nature.
Load the original and the revision, get every changed word highlighted side by side with a change summary. Free, private, in-browser.
Compare PDFs →Diff the words when the words are the risk: extract, pair, compare, and read the deletions first. Know the two blind spots — formatting-only changes and scans — and pick OCR or human eyes when those are in play. And keep the whole operation in the browser when the documents are confidential; comparison is a read-only job, and read-only jobs shouldn't require an upload.