What counts as the same product
States are joined by a valid normalized GTIN. Records with invalid codes or sharply conflicting identities are quarantined. A shared barcode is strong evidence, but not proof: codes can be reused and upstream databases can contain errors.
What gets compared
We preserve raw ingredients and normalize only case, spacing, Unicode, trivial punctuation, and two conservative USDA wording variants: “made from X” versus “X,” and “sugars” versus “sugar.” A parenthesis-aware parser then compares top-level entries. Nutrition values are compared only on equivalent bases after unit conversion. Package and serving values must parse to compatible mass, volume, or count dimensions.
Noise thresholds
Formatting-only edits do not become events. Nutrients must cross both a 5% relative threshold and an absolute threshold: 5 kcal for calories, 10 mg for sodium, and 0.5 g for the tracked macronutrients. Mass or volume changes must exceed 1% and one gram or milliliter. A lone nutrient-field change cannot reach the public threshold by itself, and physically implausible normalized 100 g nutrient profiles are not compared.
Significance and confidence
Significance describes how much label information changed. Confidence describes how strongly the available source history supports the comparison. Public events require significance of at least 30/100 and confidence of at least 55%. Serving or package values that reverse within 30 days remain in the evidence ledger but are withheld from publication as likely transient source corrections.
What Label Delta does not claim
An observed change can be a new package, corrected data, or a possible formulation change. We do not infer ingredient quantities, health effects, dietary quality, or allergy safety.