A similarity score looks like a verdict. Two PDFs are "98% similar", and the number feels reassuring, as if nothing important could have changed. That reading is tempting, and it is wrong in one specific way: a similarity score measures how much text stayed the same, not how important the changed text is. One changed number in a forty-page document barely moves the percentage, yet it can change what you owe, when you must pay, or who is responsible for what. The score is a useful summary. It is never a substitute for reading the highlights.

Compare PDF reports a similarity score alongside its word-level highlights and its counts of added, removed, and unchanged words. The three pieces belong together. The score tells you the scale of change, the counts tell you its shape, and the highlights tell you its substance. Used alone, the score misleads. Used as a starting point, it helps you decide how much attention a revision deserves.

What the score actually counts

In plain terms, the similarity score reflects the share of words that are unchanged between the two versions. If almost every word in the new file also appears in the old file in the same places, the score is high. If large passages were rewritten, added, or deleted, the score falls. It is a proportion, not a judgement. A document that kept 9,800 of its 10,000 words scores very high, whatever those 200 changed words say.

That proportion explains why long documents almost always score high. A single altered sentence in a short letter moves the percentage visibly, because the sentence is a large share of the whole. The same sentence in a long report is a rounding error in percentage terms. The longer the document, the less a high score tells you about any particular passage. Length dilutes the score, and that dilution is exactly where small, consequential edits hide.

Counts help correct for this. A report of "98% similar, 34 words added, 22 removed" describes a small edit in a large document. A report of "98% similar, 400 words added, 380 removed" describes a much larger edit in a much larger document, even though the percentages match. Always read the counts beside the score, and remember that neither number ranks changes by importance.

What a high score can hide

Consider a fictional example, invented only to make the point concrete. A small events company, which we will call Northfield Catering, sends a fictional client, Daniel Okafor, a twelve-page catering proposal with prices, timings, and terms, via the invented address hello@example.org. The client asks for a minor change to the menu, and a revised proposal arrives. The comparison reports 98.6% similarity, with a handful of added and removed words. On the surface, that looks like the menu tweak that was requested.

Working through the highlights tells a fuller story. The menu change is there, as expected. But so is a changed delivery charge, from 60 to 95 in the fictional currency of the example, and a changed cancellation window, from fourteen days to seven. Both edits sit inside paragraphs that are otherwise untouched, so they contribute almost nothing to the percentage. Yet either one could matter more to the client than the menu change that prompted the revision. The score said "barely changed". The highlights said "changed in three places, and two of them cost you something". Only the second statement is useful.

This is not a flaw in the score. A proportion cannot know that a delivery charge matters more than an adjective. Importance is a human judgement, based on what the document is for. The score's job is to summarise volume; the reader's job is to weigh meaning.

What a low score does not mean

The mirror-image mistake is treating a low score as proof that something went wrong. A low score often means a document was genuinely reworked: a policy rewritten in plainer language, a report restructured so sections moved, or a draft expanded with new material. In those cases many words changed, but the revision may be exactly what was asked for. A low score says "read this carefully", not "reject this".

Reordering also affects scores in ways that surprise people. If two sections swap places, the words are all still present, but their positions changed, and a page-by-page comparison will report substantial differences on the affected pages. That is accurate: the pages did change. It does not mean content was lost. When you see a low score, check first whether material moved rather than vanished, and whether the new wording says the same thing in different words. A paraphrased paragraph scores as changed even when its meaning survived intact.

Reading the score, counts, and highlights together

A practical routine takes under a minute. First, look at the score and counts to gauge scale: is this a tweak, a revision, or a rewrite? Second, scan the highlights in page order, and mark anything involving numbers, dates, names, obligations, or deadlines. Those categories carry consequences, while reworded introductions usually do not. Third, read the full sentence around any flagged number or date, because a changed figure only makes sense in context: a price that moved because the quantity changed is a different story from a price that moved on its own.

This routine also protects you from the opposite trap, which is ignoring a revision because its score looks high. Teams sometimes wave through anything above 95% without reading the highlights. That habit converts the score from a summary into a rubber stamp. The threshold is arbitrary, and the one change that matters does not respect thresholds. A better rule is simpler: no revision is approved on its score. The score decides how long the review takes, not whether it happens.

Honest limitations

A similarity score is computed from the text layer of each PDF. Scanned, image-only pages contribute no comparable text until they have been through OCR, and layout or image changes that leave the words alone do not move the score. The score also cannot detect meaning: a sentence reworded to say the opposite thing may score as a modest change, while a long paragraph rewritten to say the same thing scores as a large one. Treat the score as a map of where text moved, and bring your own judgement to what the moved text means.

FAQs

What does a 98% similarity score between two PDFs mean?

It means almost all of the words are unchanged between the two versions. It does not mean nothing important changed: a single altered price, date, or deadline barely affects the percentage but can matter a great deal. Always read the word-level highlights beside the score.

Why do two very different revisions show the same score?

Because a score is a proportion. A small edit in a short document and a large edit in a long document can produce the same percentage. The added and removed word counts, shown alongside the score, distinguish those two cases.

Can a similarity score tell me whether a change is good or bad?

No. The score measures how much text changed, not whether the change is correct, fair, or intended. Judging a change needs context: what the document is for, what was agreed, and what the new wording actually commits each side to.

Should I approve a revision just because its score is high?

No. Use the score to gauge how much reviewing lies ahead, then review the highlights, paying special attention to numbers, dates, names, and deadlines. If you need to check claims or figures in a document against other sources, the guide to fact-checking AI content covers a careful checking routine, and Text Compare handles comparisons of plain text passages rather than PDF files. For a second look at a PDF that arrived as scans, OCR PDF can add the searchable text layer a comparison needs.

human review pending client designation