diff --git a/server/app/goauto/aimatching/normalizer.go b/server/app/goauto/aimatching/normalizer.go
index 3f31584..1c5daac 100644
--- a/server/app/goauto/aimatching/normalizer.go
+++ b/server/app/goauto/aimatching/normalizer.go
@@ -15,6 +15,7 @@ var (
 	traditionalToSimplified *gocc.OpenCC
 	traditionalErr          error
 	weightPattern           = regexp.MustCompile(`(?i)(\d+(?:\.\d+)?)\s*(?:[-~～至到]\s*(\d+(?:\.\d+)?))?\s*(公斤|千克|kg|斤)`)
+	sizeTokenPattern        = regexp.MustCompile(`(?i)^(?:xxxs|xxl|xxxl|xxxxl|[2-9]xl|xs|s|m|l|xl|\d+xl)\b`)
 )
 
 // Normalize makes a deterministic comparison key. It intentionally removes
@@ -80,7 +81,7 @@ func normalizeWeight(raw string) string {
 // deliberately treated as no match instead of picking the first label.
 func DeterministicMatch(request MatchRequest) (MatchResult, bool) {
 	color, colorOK := exactCandidate(request.TargetColor, request.Colors)
-	size, sizeOK := exactCandidate(request.TargetSize, request.Sizes)
+	size, sizeOK := exactSizeCandidate(request.TargetSize, request.Sizes)
 	if !colorOK || !sizeOK {
 		return MatchResult{}, false
 	}
@@ -88,6 +89,27 @@ func DeterministicMatch(request MatchRequest) (MatchResult, bool) {
 	return result, true
 }
 
+func exactSizeCandidate(target string, candidates []string) (string, bool) {
+	key := sizeTokenPattern.FindString(strings.ToLower(Normalize(target)))
+	if key == "" {
+		return exactCandidate(target, candidates)
+	}
+	matches := make([]string, 0, 1)
+	seen := map[string]bool{}
+	for _, candidate := range candidates {
+		candidate = strings.TrimSpace(candidate)
+		if candidate == "" || sizeTokenPattern.FindString(strings.ToLower(Normalize(candidate))) != key || seen[candidate] {
+			continue
+		}
+		seen[candidate] = true
+		matches = append(matches, candidate)
+	}
+	if len(matches) == 1 {
+		return matches[0], true
+	}
+	return "", false
+}
+
 // ManualMatch records an already-confirmed mapping without reinterpreting it.
 // The caller owns the human-confirmation check; this helper only creates the
 // same bounded task snapshot shape used by automatic decisions.