Skip to content
ToolBoxGeniehome

String Similarity Checker

Text Tools · Added

"Close enough" is not a number until you pick a metric. This scores two strings three ways — edit distance, Jaro-Winkler and the Dice coefficient — and says what it means when the three disagree, which is the genuinely informative case.

What are you doing?
Try:

Result

Similarity

95.3%

Nearly identical — the difference is at most a character or two.

Levenshtein distance
1

1 single-character edit apart

Levenshtein ratio
92.3%

1 − distance ÷ the longer length

Jaro-Winkler
97.9%

Rewards a shared prefix

Dice
95.7%

Overlap of character pairs

Levenshtein
92.3%
Jaro-Winkler
97.9%
Dice
95.7%

The three measures answer different questions, which is why all three are shown. Levenshtein counts edits, so it is the one to use for typos. Jaro-Winkler was designed for matching names and gives a bonus for a shared prefix — which is right for people and wrong for identifiers, where two IDs differing only in the last digit score above 98%. Dice compares character pairs and ignores order beyond adjacency. Where they agree the answer is solid; where they diverge, the divergence is telling you which kind of similarity you have.

How to use the string similarity checker

  1. 1Type or paste a string into each box.
  2. 2Choose whether case, punctuation and repeated spaces should count.
  3. 3Read the three scores and the plain-language verdict beneath them.
  4. 4Switch to the candidate list to rank one string against many at once.

Examples

The textbook pair

Input
kitten against sitting
Result
Edit distance 3 (57.1% alike), Jaro-Winkler 0.746, Dice 0.364 — 'loosely similar', flagged because the metrics disagree.

Three single-character edits: k becomes s, e becomes i, and a g is appended.

A name typed twice

Input
Robert Smith against Robbert Smith
Result
Edit distance 1, Jaro-Winkler 0.982, Dice 0.957 — 'nearly identical'.

About the string similarity checker

Three ways to be nearly the same

Similarity is not a single quantity. Levenshtein distance asks how much work it takes to get from one string to the other and answers in keystrokes. Jaro-Winkler asks how many characters roughly line up, then adds a bonus for a shared prefix — a deliberate bias, introduced because people mistype the ends of names far more often than the beginnings.

The Dice coefficient takes a third view entirely: it chops both strings into overlapping pairs of adjacent letters and asks how many pairs the two sets have in common. That makes it indifferent to where the material sits, which is exactly right for matching 'Smith, John' against 'John Smith' and exactly wrong for catching a transposed digit.

Thresholds are a property of your data

There is no universal cutoff above which two strings count as the same thing. A threshold that reliably merges duplicate customer records will happily merge two genuinely different people in a dataset full of common surnames, and one tuned for product codes will reject every legitimate spelling of a street address.

The workable method is to score a few hundred pairs you have already judged by hand, look at where the mistakes fall, and set the cutoff from that — separately for every field. A score is evidence for a decision, never the decision itself.

Frequently asked questions

Which of the three scores should I use?
It depends what 'alike' means for your data. Edit distance is the honest choice for typing slips, because it counts actual keystrokes. Jaro-Winkler was built for personal names and deliberately rewards a matching start. Dice looks at shared adjacent letter pairs and ignores where they sit, so it survives reordering that the other two punish heavily.
Why do the metrics sometimes disagree so sharply?
Because they measure different things, and the gap is information rather than noise. A wide spread between Jaro-Winkler and Dice nearly always means the two strings open the same way and then diverge — 'international' against 'internal' being the shape of it. This page flags that pattern when it appears instead of averaging it away.
What exactly does an edit distance of 3 mean?
Three single-character operations — insert, delete or substitute — will turn one string into the other, and no shorter sequence will. It is a count rather than a proportion, so 3 is trivial across two paragraphs and severe across two four-letter words. That is precisely why the percentage shown beside it divides by the longer of the two.
Are emoji and accented letters handled properly?
Yes. The comparison walks text by Unicode code point rather than by UTF-16 unit, so an emoji outside the basic plane counts once instead of twice. Accented letters are compared as written: é and e are different characters, and if you need them treated alike, normalise the text before pasting it in.

Related tools