How this similarity check actually works
A plagiarism checker is a similarity tool: it samples sentences from your document, searches the public web for close matches, and reports how much of your sampled text overlaps with content that already exists online.
No web-based checker can prove that a document is original. What it can do is tell you which sentences already appear somewhere public — which is enough to catch an unattributed paste, a paraphrase that stayed too close to its source, or a boilerplate paragraph a subcontractor reused across three clients.
Because the scores these tools produce are often misread, this page documents the exact pipeline behind this checker: how sentences are selected, what is searched, how the percentage is calculated, and what the verdict does and does not mean.
The four stages of a check
Your text is not compared against a private database of student papers. It is compared against the live public web, one sampled sentence at a time.
- 1
Extraction
Pasted text is used as-is; an uploaded PDF is parsed for its embedded text layer, lowercased and collapsed to single spaces. A document must contain at least 50 characters to be accepted.
- 2
Fingerprinting
The text is broken at sentence punctuation, and only fragments of 8 words or more become candidates — short fragments match everything and prove nothing. Between 10 and 40 candidates are then sampled at even intervals across the document, and each sample is truncated to 300 characters.
- 3
Web search
Each fingerprint is sent as a search query to a third-party web search API, and the single best-matching public result is kept along with its URL and snippet. Nothing is added to a permanent database of submissions.
- 4
Similarity scoring
Your sample and the retrieved snippet are reduced to sets of lowercase words with common stop words such as "the", "and" and "of" removed. The score is the share of your sample’s distinct words that also appear in the snippet.
Because sampling is interval-based, a 6,000-word document is judged from a spread of roughly 40 sentences — not from every sentence in the file.
What the numbers on the report mean
The percentage is a share of sampled sentences, not a share of your word count. This is the single most misunderstood part of every similarity report on the internet.
- Exact Match
- A sample whose score is 60% or higher against the retrieved snippet. Read these first — they are the ones most likely to be genuine copying.
- Partial Match
- A score between 40% and 59.9%. Often a close paraphrase, a shared quotation or standard phrasing in a technical field.
- Original
- A score below 40%, or no web result found for that sentence at all.
- Overall percentage
- Matched samples divided by total samples. If 8 of 40 sampled sentences matched, the report reads 20% — regardless of how long those sentences were.
- Verdict line
- A blunt threshold: above 20% of samples matched, the report says plagiarism detected. It is a prompt to review, not a finding of misconduct.
- Stop words are ignored
- Scoring uses distinct non-stop-words, so rearranging a sentence without changing its vocabulary will not lower the score much. Rewriting the ideas will.
Reading a result correctly
Two sampled sentences, both flagged, with very different conclusions. The score tells you where to look; you decide what it means.
Example 1: A high score that is not misconduct
Photosynthesis converts light energy into chemical energy stored in glucose molecules.Score: 87%
Status: Exact Match
Source: an encyclopaedia pageTextbook definitions have one conventional phrasing. A high score here means the sentence is generic, not stolen. Either cite the source or rewrite it in your own framing — but do not panic.
Example 2: A moderate score that deserves attention
Our onboarding process reduced first-week churn by more than a third across two product lines.Score: 52%
Status: Partial Match
Source: a competitor’s case-study pageA specific, non-generic claim matching a competitor page is the pattern worth investigating. Open the source, compare the paragraphs, and confirm whether a writer paraphrased it.
Example 3: A clean report on a short document
340 words, 11 candidate sentences, 10 sampledOverall: 0.0%
Verdict: DOCUMENT IS CLEANA 0% result on a short document is weak evidence. Ten samples is the floor, and short documents produce few candidates — the smaller the sample, the less a clean verdict tells you.
Honest limitations you should know before relying on this
Direct answer: this tool is a review aid. It cannot certify originality, and there are specific categories of copying it will miss.
It only sees the public, indexed web
Paywalled journals, private course repositories, unindexed PDFs and other students’ unpublished submissions are invisible to it. Institutions that need repository matching use commercial systems such as Turnitin for exactly this reason.
Sampling means partial coverage
At most 40 sentences are checked. A single copied paragraph in a 5,000-word report can fall between samples and go unreported. If a specific section worries you, paste that section on its own and re-run the check.
Scanned PDFs return nothing useful
Text is read from the PDF’s text layer. A scan or photo of a page has no text layer and no OCR is performed, so the upload will fail or return an empty document. Run OCR first, or paste the text directly.
Word-order changes are not detected
Scoring compares word sets, so a shuffled sentence keeps a high score, while a genuine rewrite with different vocabulary scores low even if the idea was borrowed. Attribution is an editorial judgement no automated score can make for you.
Your text is sent to a third-party search service
Sampled sentences leave your browser and are sent to our server, then to an external web search API to look for matches. Do not upload confidential legal, medical, client or pre-publication material unless your organisation permits third-party processing. Details are in our privacy policy.
When to use this and when to use something else
Choose the checker that matches the consequence. A blog post and a dissertation carry very different risk.
| Situation | Better choice | Why |
|---|---|---|
| Blog, marketing or web copy before publishing | This tool | The risk is duplicate public content, which is exactly what a live web search surfaces. |
| A freelance deliverable from a subcontractor | This tool, run per section | Reused boilerplate usually clusters. Checking sections separately beats one whole-document average. |
| Graded coursework or a thesis | Your institution’s official checker | Only institutional systems compare against student-paper repositories, and only their report counts for an academic integrity process. |
| Confidential or unpublished manuscripts | An offline comparison | Any online checker transmits your text to at least one external service. Do not send material you cannot share. |
What to do with a flagged passage
- Open the source URL and read the surrounding paragraph — snippets are easy to misread out of context.
- If the idea is genuinely borrowed, cite it. Attribution resolves most matches without any rewriting.
- If the wording is conventional (definitions, legal phrasing, standard technical descriptions), leave it and note why.
- If it is a paraphrase that stayed too close, rewrite from your own notes rather than editing the original sentence word by word.
- Re-run the check on the rewritten section alone to confirm the score dropped.
- Confirm the rewrite still fits the brief with the word counter, then package the final file with Word to PDF.
How this page is maintained
The pipeline, the 8-word candidate rule, the 10–40 sample range, the 300-character truncation and the 60% / 40% / 20% thresholds on this page were read directly from this tool’s server-side implementation rather than paraphrased from marketing copy. Scores are heuristic by design and are described here as such.
Written and reviewed by the ToolsRacks editorial team. Last reviewed: .