Understanding Similarity Reports: What a Plagiarism Score Actually Means

Details vary by service. Research papers and university guides describe five broad steps:

1. Database. Your text is compared with an index of web pages and published articles. Depending on the service, the index also holds earlier submissions. The University of Kansas teaching center notes that a paper submitted to such a database becomes part of the database. That is how your own earlier coursework can show up as a source.
2. Fingerprinting. The text is split into short, overlapping sequences of words or characters, often called n-grams. In three-word sequences, "the study shows a clear effect" becomes "the study shows," "study shows a," and so on. Each sequence is converted into a short number called a hash. A selection of these hashes forms the document's fingerprints. A 2003 paper on document fingerprinting presents this method as a way to find partial copies within large sets of documents.
3. Matching and thresholds. Sequences that share a fingerprint with an indexed document count as matches. Overlaps below a minimum length, such as a single shared word, are ignored. The 2003 paper argues that a match must be long enough to imply copying, not just a common word or idiom.
4. Filters. A University of Reading staff guide describes filters that can exclude quotations, the bibliography, or small matches. The same draft can therefore score differently under different settings.
5. Report. Matched passages are highlighted and linked to their sources. The percentage shows the total amount of matched text as a proportion of the document.

What the comparison can miss

A checker finds shared wording, not borrowed ideas. A 2025 systematic survey of detection methods describes string matching as effective for verbatim copying. The same survey says string-matching and n-gram methods struggle with paraphrased or obfuscated text. It notes that translation without credit makes detection more complex. Catching a borrowed idea, it adds, requires models that understand meaning at the level of discourse. Images are another gap. A primer for researchers says that, compared with detecting textual similarity, image plagiarism detection is "a more challenging task." Anything the checker misses can still be plagiarism. Our guide on paraphrasing plagiarism shows how an uncited idea in new words still counts.

The percentage says how much of your text matches indexed content, not why it matches. The Reading staff guide puts it simply: "a similarity match does not always mean plagiarism." Written for teaching staff, it adds that the decision on plagiarism "remains your academic judgment." The software doesn't make that call.

The check is purely textual. Unless a filter removes quotations, a correctly quoted and cited passage counts toward the percentage just like copied text with no attribution. Yet that difference is what defines plagiarism: using someone else's words or ideas without credit. Our guide on what counts as plagiarism gives the fuller definition. The report informs that judgment but doesn't replace it.

No single number separates a safe score from a problem one. A library guide at the University of Dundee says outright that there is "no fixed 'acceptable' number for a similarity report." The Reading staff guide agrees that there is "no ideal percentage to look for." It says the figure "will vary depending on the type and length of assignment and the requirements of the work involved."

A literature review that quotes heavily and cites every source correctly can score higher than a short argumentative essay with few citations. That is expected, not suspicious. Because what is normal depends on the assignment, Certentis doesn't publish a universal threshold. Its benchmarks page applies the same rule to accuracy. It gives no accuracy figure for the plagiarism checker, because no suitable labeled reference corpus exists to test one against.

The table lists four sources of harmless overlap and one that needs your attention.

A research-library FAQ at MD Anderson Cancer Center adds that a single match under 2% is most likely a proper name or a few words of text. It still advises skimming those matches.

Look at where the overlap sits, not at the overall percentage. The MD Anderson FAQ tells authors to focus on one or more sentences that exactly or closely repeat another source, and on matching paragraphs. If such a passage has no quotation marks and no citation, that is what unattributed copying looks like in a report. Check it however low the total is. For each highlighted match, ask:

• Is the matched text inside quotation marks, and is the source cited?
• Does the match sit in your reference list rather than your argument?
• Is it a short, generic phrase or technical term rather than a distinctive sentence?
• Is a single source responsible for a long, continuous block of your own prose?

If a long block of your prose matches one source and has no quotation marks or citation, fix that passage before you submit. The goal is correct attribution, not a lower percentage.

Run your draft through a checker while there's still time to act on the result. Then open every flagged passage instead of reacting to the overall percentage. Confirm that each quotation has both quotation marks and a citation and that paraphrased material names its source. Any long match should sit in your reference list or a cited quotation, not in unattributed text. The Certentis plagiarism checker produces the kind of report described here, so you can work through it before a grader does.

If a match turns out to be your own earlier coursework, don't just ignore it. Check it against your institution's self-plagiarism rules, and see our guide on self-plagiarism. If the fix is to attribute a paraphrase properly rather than delete it, our guide to citing a paraphrase covers the mechanics.