Compare two sets of normalized unique tokens. The percentage measures shared vocabulary, not meaning, plagiarism or whether either passage is original.
The text similarity checker compares the distinct tokens in two inputs after a basic cleanup. Paste a passage in each box and select Compare Texts. It reports the number of shared unique tokens relative to the larger token set as a percentage. This is useful for a narrow vocabulary comparison, such as inspecting whether two drafts use many of the same terms. It does not search other documents, understand ideas or determine who wrote either passage.
Choose samples that fit the question you are asking. Comparing an entire guide with a short caption is different from comparing two revisions of the same paragraph. Navigation text, repeated notices and unrelated appendices can change the vocabulary sets even when the relevant prose is unchanged. Keep the source selections with the result so another reviewer can understand which material was compared rather than treating the displayed number as a property of the whole website.
Each input is lowercased, characters outside basic word characters and whitespace are removed, and the result is split at whitespace. Empty pieces are discarded. The recognized alphabet is English-oriented: letters A–Z, digits and underscores survive, while many accented and non-Latin characters do not. This is not a language-aware tokenizer or a general representation of every word a reader sees. Check a representative sample before relying on the score for multilingual material.
Punctuation is deleted rather than replaced with a separating space. That means “red-blue” becomes the token “redblue”, whereas “red blue” becomes two tokens. Apostrophes and slashes can create similar differences between visible wording and the normalized set. A changed punctuation style can therefore alter the score without a major change in subject. Review those rules when another comparison tool disagrees, because not every method uses the same preprocessing or definition of a token.
For sets A and B, the numerator is the count of distinct tokens found in both sets. The denominator is the larger of the two set sizes, not their union size. The displayed value is that fraction multiplied by one hundred and rounded to two decimal places. This is not the Jaccard union formula or a semantic similarity model. Use the formula implemented here when explaining the result rather than assigning it a familiar algorithm name that calculates something else.
For “red blue” and “blue green”, each normalized set has two members and the shared set contains only blue. One divided by two gives fifty percent. If one input contains ten unique tokens and the other contains those ten plus another ten, the shared count is ten and the larger set has twenty members, again producing fifty percent. The result describes overlap relative to the larger vocabulary, not the percentage of sentences that are copied.
Repeated occurrences of the same token contribute only one set member. “cat cat dog” and “cat dog” have identical normalized vocabulary and produce one hundred percent under this method. Their lengths and repetition patterns differ, but the set comparison no longer records those differences. This makes the tool unsuitable for measuring how many words were copied in sequence or whether a paragraph repeats a term excessively. Use an occurrence counter for those separate questions.
Word order is also discarded. “Dog bites man” and “Man bites dog” contain the same unique vocabulary and can score one hundred percent despite describing different events. Negation, context and relations between words are not modeled. A strong vocabulary overlap can be an expected feature of two texts about the same subject rather than evidence of duplicate meaning. Conversely, a paraphrase can share few exact tokens while conveying almost the same idea.
The current output labels values above seventy percent as very high, above forty through seventy as medium and lower values as low. These are fixed interface branches using the rounded percentage, not a validated plagiarism decision. The low category includes the phrase “original content”, but this calculation cannot establish originality. A passage copied from an outside source can receive a low score against an unrelated second input because the relevant source was never supplied or searched.
Do not use the badge as proof in an academic, editorial or ownership dispute. Review actual matching language, attribution and source context with an appropriate process when those questions matter. Likewise, a high score is not automatically a content problem if both drafts need the same technical vocabulary. The useful conclusion is limited to the documented overlap of the supplied normalized sets, not a broader claim that one text is acceptable, copied or better written.
When comparing revisions, keep the same section boundaries and note whether a glossary, list or heading has been added. A longer vocabulary set can lower the fraction even when every original term remains. Record both the nature of the edits and the result so a later reader can distinguish changed subject coverage from a simple change in sample size. A smaller percentage is not itself an improvement and should not become an editing target.
For an editorial review, read the actual passages side by side after using the checker. Identify repeated explanations, missing definitions and changed claims. The score can suggest that the vocabulary differs, but it cannot tell which version answers the reader's question more clearly. If your task concerns exact phrases, use the Keyword Frequency Checker or inspect the source directly. Keep semantic review separate from this particular set-based statistic so the result is neither understated nor overstated.
An input containing no recognized tokens causes the tool to report that one text has no valid words. That message reflects its restricted normalization, not a judgment that a reader's language lacks words. Hindi-only text, punctuation-only input or certain symbol-heavy samples can fall outside the rule. Select a suitable language-aware comparison method when multilingual accuracy is required instead of treating a failure or zero overlap as evidence that the texts are unrelated.
Digits and underscores can enter the vocabulary sets just like basic letter strings. A table of product codes is therefore measurable, but the result does not validate those codes or describe their business meaning. Spaces and punctuation in an identifier can affect token boundaries. Test a small known example from the actual data type before using a large result, and retain the original spelling when normalization might remove meaningful distinctions from names, versions or identifiers.
Select Compare Texts again after changing either input because the output represents the last explicit run. The function does not maintain a revision history or highlight matching tokens in context. Save the samples and the calculation definition in your normal notes if you need a reproducible comparison. A short expected example such as the two-color sets above is a useful first check before interpreting a larger and less transparent number.
Finish by describing the result precisely: shared unique normalized vocabulary relative to the larger supplied set. Avoid translating it into a claim about copied sentences, percent originality or search penalties. The checker can be one focused inspection aid within a broader content review, while accuracy, attribution, organization and reader usefulness require their own evidence. Clear reporting of those limits is what makes the measurement useful rather than just a reassuring or alarming percentage.
It compares sets of unique normalized tokens from two supplied inputs, not sentence meanings or outside documents.
The shared unique-token count is divided by the larger set size. The denominator is not the union of both sets.
No. Each normalized token appears only once in its set, regardless of its occurrence count.
No. Token order is discarded when the sets are created.
Not necessarily. Different sentences can have the same vocabulary while changing the relationships between the words.
No. The tool does not search sources or establish authorship. The interface label should not be treated as proof.
It is removed rather than consistently replaced with spaces. A hyphenated pair can therefore become one joined token.
Its normalization is not language-aware and can remove Hindi characters. Use a suitable method for that requirement.
No. They are fixed local labels and cannot certify copying or originality.
Select Compare Texts again after changing either passage to generate a current result.