What is the difference between Jaro and Jaro–Winkler?

Jaro similarity scores two strings on their matching characters and transpositions alone. Jaro–Winkler adds a bonus for a shared leading prefix: JW = Jaro + ℓ·p·(1 − Jaro), where ℓ is the common prefix length (capped at 4) and p is a scaling factor (usually 0.1). So Jaro–Winkler is always greater than or equal to Jaro, and it rewards strings that agree at the start — useful because typos more often fall later in a word.

How is the Jaro–Winkler prefix bonus calculated?

Count the identical leading characters of the two strings, up to a maximum of 4. Call that ℓ. Multiply ℓ by the scaling factor p (default 0.1) and by (1 − Jaro), then add the result to the Jaro score. For MARTHA and MARHTA the prefix MAR gives ℓ = 3, so the bonus is 3 × 0.1 × (1 − 0.9444) = 0.0167, lifting 0.9444 to 0.9611.

Is Jaro–Winkler a distance or a similarity?

Both terms are used, which causes confusion. The value this tool reports first — 1 for identical strings, 0 for nothing in common — is the Jaro–Winkler similarity. Some libraries report the Jaro–Winkler distance instead, which is 1 − similarity. If a library gives you 0.04 for two near-identical strings, it is reporting distance; subtract from 1 to get the 0.96 similarity shown here.

What is a good Jaro–Winkler score for a match?

Winkler's record-linkage work commonly treats scores above about 0.90 as strong matches and below roughly 0.70 as non-matches, with a review band in between. But the right cut-off depends on your data — names, product codes and addresses behave differently. Calibrate the threshold on a sample you have already labelled by hand rather than trusting a fixed number.

When should I use Jaro–Winkler instead of Levenshtein distance?

Reach for Jaro–Winkler on short strings where the start matters and typos are transpositions or dropped letters — personal names, especially. It returns a 0–1 similarity directly and gives extra weight to a shared prefix. Levenshtein counts raw edit operations and suits longer text or when every edit should weigh the same. Many deduplication pipelines try both and keep whichever separates their matches from non-matches more cleanly.

How are transpositions counted in the Jaro formula?

Take the matched characters of string A in order and the matched characters of string B in order, then count the positions where the two sequences disagree. That count is always even, because a swap affects two positions; halve it to get t. In MARTHA vs MARHTA the matched sequences are MARTHA and MARHTA, which differ at the T and H positions — two disagreements, so t = 1.

Is this calculator case-sensitive?

By default it is case-insensitive, so "Smith" and "SMITH" score 1.0. Toggle "Case-sensitive" to compare the exact characters. The tool compares Unicode code points and does not fold accents or normalise forms, so "é" written as one code point differs from "e" plus a combining accent — normalise your text first if that matters for your data.

Does this tool send my strings anywhere?

No. The matching, transposition count, prefix bonus and every intermediate value are computed in your browser with plain JavaScript. Nothing is uploaded, logged or stored, so you can paste names, IDs or private records safely, and the page keeps working offline once it has loaded.

AI · String metrics

Jaro–Winkler Similarity Calculator

Compare two strings and get the Jaro similarity, the Jaro distance and the Jaro–Winkler score, with the matching characters, transpositions, common prefix and substituted formula laid out. No signup, nothing uploaded — it all runs in your browser.

By Induwara Ashinsana— Executive Director, Ryzera TechnologiesUpdated Jul 9, 2026

Jaro–Winkler similarity

String A6/128

Any text, 1–128 characters.

String B6/128

Compared character by character against A.

Case-sensitive (treat “A” and “a” as different)

Prefix factor p

Max prefix ℓ

Examples

Jaro–Winkler similarity

0.9611

96.11% · range 0–1

Jaro similarity

0.9444

94.44% · before prefix bonus

Jaro–Winkler distance

0.0389

1 − JW · Jaro dist 0.0556

Interpretation

Very likely a match

Jaro = (1/3)( 6/6 + 6/6 + (6−1)/6 ) = 0.9444

JW = 0.9444 + 3·0.1·(1 − 0.9444) = 0.9611

Jaro distance = 1 − 0.9444 = 0.0556 · JW distance = 0.0389

Cross-check. The primary Jaro score is 0.9444; recomputing the transpositions from the matched-character substrings independently gives 0.9444. They reconcile, as they must.

Character alignment

martha

marhta

Derivation

Term	Value	How
Lengths \|A\|, \|B\|	6, 6	Code points compared
Matching window	2	max(0, ⌊max(\|A\|,\|B\|)/2⌋ − 1)
Matching characters m	6	Equal chars within the window
Transpositions t	1	2 out-of-order matched positions ÷ 2
Common prefix ℓ	3	Capped at 4
Scaling factor p	0.1	Prefix bonus weight

Method: Jaro (1989) matching window + transpositions, then Winkler (1990) prefix bonus JW = Jaro + ℓ·p·(1 − Jaro). Comparison is over Unicode code points; no diacritic folding. Nothing leaves this page.

How it works

Jaro–Winkler similarity scores how alike two strings are on a 0-to-1 scale, tuned for short strings such as names. It builds on the Jaro similarityfrom M. A. Jaro's 1989 record-linkage work and adds W. E. Winkler's 1990 bonus for a shared leading prefix. Identical strings score 1; strings with no characters in common score 0.

Jaro similarity rewards common characters and penalises those that are out of order. Two characters count as a match only when they are equal and lie within a matching window of each other:

window = max(0, ⌊max(|s1|, |s2|) / 2⌋ − 1)

With the matching characters m and transpositions t counted, the Jaro score is:

Jaro = (1/3) · ( m/|s1| + m/|s2| + (m − t)/m )

A transposition is a pair of matched characters that appear in a different order in the two strings. You count them by taking the matched characters of each string in order, tallying the positions where the sequences disagree, and halving that tally — the count is always even. If m = 0, the Jaro score is defined as 0 rather than a divide-by-zero.

Winkler then boosts strings that agree at the start, since typos tend to fall later in a word. Let ℓ be the length of the common prefix, capped at 4, and p the scaling factor (default 0.1):

JW = Jaro + ℓ · p · (1 − Jaro)

The ceiling p ≤ 0.25 keeps the score at or below 1, because the largest prefix bonus is 4 × 0.25 = 1 of the remaining gap. Setting p = 0 collapses Jaro–Winkler back to plain Jaro. The Jaro distance is 1 − Jaro and the Jaro–Winkler distance is 1 − JW. This tool implements the standard prefix bonus with no boost threshold — the most widely cited definition — and reproduces the published reference values for MARTHA/MARHTA, DWAYNE/DUANE and DIXON/DICKSONX exactly.

Worked examples

MARTHA vs MARHTA — the canonical textbook pair

Lengths 6 and 6; window = max(0, ⌊6/2⌋ − 1) = 2
All 6 characters match ⇒ m = 6
Matched order M A R T H A vs M A R H T A differ at 2 positions ⇒ t = 1
Jaro = (6/6 + 6/6 + (6−1)/6) / 3 = 0.9444
Prefix MAR = 3 (T ≠ H); JW = 0.9444 + 3·0.1·(1 − 0.9444) = 0.9611

DWAYNE vs DUANE — Winkler's own example (different lengths)

Lengths 6 and 5; window = max(0, ⌊6/2⌋ − 1) = 2
Matches D, A, N, E ⇒ m = 4 (W and Y find no partner)
Matched order D A N E vs D A N E agree ⇒ t = 0
Jaro = (4/6 + 4/5 + 4/4) / 3 = 0.8222
Prefix D = 1 (W ≠ U); JW = 0.8222 + 1·0.1·(1 − 0.8222) = 0.8400

abc vs xyz — the zero-similarity edge case

Lengths 3 and 3; window = max(0, ⌊3/2⌋ − 1) = 0
No characters match within the window ⇒ m = 0
By definition Jaro = 0, and the prefix bonus cannot lift a zero score
Jaro = 0.0000, Jaro–Winkler = 0.0000 → No similarity

Frequently asked questions

Sources & references

The formulas on this page were last cross-checked against these sources on 2026-07-09. Jaro–Winkler is a fixed algorithm definition, so the tool needs no rate or schedule updates — only the reference examples are periodically re-reconciled.

Related tools

LiveAI

Cosine Similarity Calc

Compute the cosine similarity, cosine distance, and angle between two numeric vectors or two short texts, with the full dot-product and magnitude working. Matches scikit-learn, runs entirely in the browser.

Open tool

LiveAI

Wasserstein Distance Calc

Compute the Wasserstein-1 distance (Earth Mover's Distance) between two 1-D distributions or samples, entirely in the browser — the drift metric that stays finite when KL divergence blows up. Full per-interval working, a CDF chart, and a quantile cross-check; matches scipy.stats.wasserstein_distance.

Open tool

LiveAI

Levenshtein Distance Calc

Computes the Levenshtein edit distance between two strings with a full Wagner–Fischer DP matrix and step-by-step edit operations, entirely in the browser.

Open tool

Rate this tool

Be the first to rate

Comments & feedback

Spotted a bug or want an improvement? Tell us — our team reviews every comment, and good ideas get built. Comments are public and anonymous.

Found a bug, edge case, or want to suggest an improvement?

Email me at [email protected] — most fixes ship within 24 hours.