How GoodSwap scores.

On this page

Principles

1. AI never decides a score. The score is a pure function of the product’s facts, our knowledge base and published parameters. Same input, same output, every time.
2. Everything is on the receipt. Every product starts at 100. Each deduction or bonus is a line with a reason, an evidence tier and its references.
3. One universal score. The score is the consensus “Balanced” view: regulators and strong science. Personal standards, like seed-oil-free, are lenses. They add a verdict and never change the score.
4. Missing data is never punished silently. Unknown parts are skipped and noted on the receipt, and confidence drops instead. When we can’t stand behind a number, we show “Not scored yet” and ask for what’s missing.

Bands

Five bands. Each has a color, an icon and a word, so color is never the only signal.

Band	Score
Excellent	80–100
Good	60–79
Okay	40–59
Poor	20–39
Bad	0–19
Not scored yet	no score

On the Swap Meter, a pin in its band’s color marks the product’s score. When a better swap of the same kind exists, the same pin, outlined in lime, marks the swap’s score. We never use rings or circular gauges.

The Receipt

The Receipt is the score’s ledger. Read it top to bottom:

• START 100. Every product starts with full marks.
• Lines. One per deduction or bonus, grouped as ingredients, processing, nutrition, lab tests and bonuses, in that order. Each line has its points, a reason and an evidence tier. In the app, tap a line to read its sources.
• Subtotals and TOTAL. The total is 100 plus every line, rounded to a whole number (halves round away from zero), kept between 0 and 100, then limited by any cap.
• Footer. The barcode, the date, and the versions of the score engine and the knowledge base. Those two versions are all you need to reproduce the number.

Points are always multiples of 0.5, so the arithmetic is exact. Here is a worked example with an illustrative product, scored with the rules of score v1.3.0. Two dyes and corn syrup are each a high-concern ingredient (−20), so the ingredient group reaches its −50 cap on the third line, and the rest of the list (sugar, modified corn starch, natural flavor) adds nothing more:

Food and drinks

Ingredients

Each distinct ingredient we recognize has a concern level from 0 to 4 in our knowledge base, and each level costs a set number of points:

Concern	Food	Beauty & home	What it takes
c4 Very high	−25	−25	Banned or revoked by a major regulator, or classified by IARC as Group 1 or 2A at relevant exposure. Needs regulatory evidence.
c3 High	−20	−12	A regulatory warning or restriction somewhere, or strong evidence of harm at typical intake.
c2 Moderate	−7.5	−5	Strong evidence at high intake, emerging evidence of harm, or a common allergen or irritant.
c1 Low	−2.5	−1	Minor or very limited concern, such as being an ultra-processing marker only.
c0 None	0	0	A whole food or a benign function. No credible concern. Never shown as a deduction.

• Diminishing returns. Deductions are ranked from largest to smallest (ties by label order): the first three count in full, the 4th to 6th count half, and the 7th onward count a quarter. A long list adds up, but no single extra ingredient dominates.
• Group cap: ingredients can take at most −50 points in food (calibration range 40–60) and −85 in beauty and home. Food concern points may be calibrated by up to ±30%.
• Each line’s reason comes from the ingredient’s knowledge-base entry; its tier is the strongest tier among that ingredient’s effects.

Processing

• NOVA 4 (ultra-processed): −18 (calibration range 10–18). NOVA 3 (processed): −8 (range 3–8). NOVA 1–2 or unknown: 0.
• If NOVA is unknown and the product contains at least 3 distinct ultra-processing marker families (such as colors, flavorings, emulsifiers, modified starches or industrial sugars), we treat it as NOVA 4, and the receipt says “estimated”.
• Marker families: for NOVA 4 products, each distinct marker family beyond the first costs −1.5, up to −8.

Nutrition

Per 100 g for food and per 100 ml for drinks. A product counts as a drink when it’s a beverage or its nutrition is given per 100 ml. Sugars and energy have separate drink tables; saturated fat, salt, fiber and protein use the same thresholds for both.

• Bonuses: fiber of 3 g or more +2, 6 g or more +4. Protein of 8 g or more +2, 16 g or more +4, except when the sugar deduction is 8 points or more.
• Group caps: nutrition deductions stop at −30 and nutrition bonuses at +6. Group caps are applied line by line in receipt order; a line that would add nothing is dropped.
• Added sugars replace total sugars when they’re known and the product is a dairy or fruit-based food, so the natural sugars in plain yogurt or fruit don’t count against it.
• Unknown nutrition adds no points. The receipt shows an info line, and confidence can’t be high.

“High in” fat, saturated fat, sugars or salt

Above the front-of-pack “high” levels used in the UK, a product is high in that nutrient. Being high in any of them caps the score at 79 (it can’t be Excellent) and removes the food bonuses below (fiber and protein bonuses still count). Fat and saturated fat aren’t checked for the whole foods listed under the exemptions below.

Nutrient	Food, per 100 g	Drinks, per 100 ml
Fat	more than 17.5 g	more than 8.75 g
Saturated fat	more than 5 g	more than 2.5 g
Sugars	more than 22.5 g	more than 11.25 g
Salt	more than 1.5 g	more than 0.75 g

Whole-food exemptions

Fat in whole foods isn’t treated like fat in a snack:

• No saturated-fat or energy deduction for oils and olive oils, butter, ghee, nuts, seeds, nut butters (no added sugar or oil), avocados and eggs.
• Half the saturated-fat and energy deductions for cheeses, milks, plain yogurts and unprocessed meats.

Lab tests

Only when lab data exists with a benchmark (for example a regulator’s action level). A result above the limit costs −20 (group cap −30) and also caps the score at 20. Results below the limit, at trace level or not detected cost nothing and are shown for context. Each result shows the lab, the date and whether the data came from an independent lab, a regulator or the brand itself (“brand-supplied”). We never claim a phone camera can measure contaminants.

Packaging

Shown as context only in v1. No points.

Bonuses

• Certified organic (for example USDA Organic, EU organic, Canada Organic, Soil Association): +5.
• Simple whole food: at most 3 ingredients, all recognized, all of concern 1 or less, and NOVA 2 or lower: +5.
• No additives: a known, fully recognized ingredient list with no food additives: +3.
• Bonuses together are capped at +10. A product that is high in fat, saturated fat, sugars or salt gets none of them.

Caps

Some facts matter more than any sum. After adding up the lines, the lowest applicable cap wins, and the receipt names it:

The product contains…	Maximum score
An ingredient banned or revoked for food in the EU, the UK, Canada or the US	25
Any ingredient of very high concern (c4)	49
A lab result above its limit	20
Any ingredient of high concern (c3)	79
Food or drink high in fat, saturated fat, sugars or salt	79
Food, drink or baby food with no nutrition facts (except plain water, coffee, tea, herbs, spices and fresh produce, which need none). Without an ingredient list there is no score at all.	79
Baby food with an EU warning-label dye or a non-nutritive sweetener	49
Beauty: an ingredient prohibited in EU cosmetics (Regulation 1223/2009, Annex II)	15
Beauty and home: any ingredient of very high concern (c4)	39

A cap of 79 is the top of “Good”: those products can’t be rated Excellent, whatever else is on the receipt.

“Banned” means explicitly prohibited or withdrawn after review. An ingredient that was simply never authorized somewhere is shown as “not permitted in” that place on its ingredient page, and doesn’t trigger the cap. We use the regulators’ own words: “banned” only when something is banned, otherwise “not permitted in”, “requires a warning label in” or “restricted to”.

Beauty, home, supplements, pet and baby

Kind	How it’s scored
Beauty & personal care	Ingredient deductions on the beauty scale above (c4 −25, c3 −12, c2 −5, c1 −1) with the same diminishing returns, up to −85. Fragrance (“parfum”) counts as moderate concern because it’s an undisclosed mixture; the EU requires dozens of fragrance allergens to be named on labels. Certified organic or natural cosmetics (COSMOS, Ecocert, USDA Organic) +5.
Household	The beauty method with household ingredients. EPA Safer Choice certified +8.
Supplements	The beauty method applied to excipients and additives, plus lab tests. Third-party certified (NSF Certified for Sport, USP Verified, Informed Sport or Informed Choice) +8.
Pet food	Ingredients and processing only: human nutrition tables and bonuses don’t apply.
Baby food	The food method, plus a cap of 49 for EU warning-label dyes and non-nutritive sweeteners.
Other	No score.

Evidence tiers

Every line carries a tier, which says how sure we are, separately from the points, which say how much it matters. Tiers never reuse score colors.

Regulatory

A law or regulator acts on it somewhere: banned, revoked, restricted, or a required warning label. IARC Group 1 or 2A classifications also count.

Strong evidence

An authoritative body’s opinion or a consistent body of research, such as a meta-analysis.

Emerging evidence

Limited or mixed studies. Worth knowing, never treated as settled.

Preference

A lifestyle choice, such as avoiding seed oils or palm oil. Lenses handle these.

Info

Context only. No points.

The tier limits the concern level. Very high concern (c4) needs regulatory evidence. High concern (c3) needs regulatory or strong evidence. Emerging evidence alone can reach moderate (c2) at most, and a preference alone stays at c1 or below; preferences belong in lenses.

Sources follow a hierarchy: the law and its date, then authoritative classifications (such as IARC groups), then agency risk assessments with disagreements shown, then peer-reviewed studies, then dated independent tests. Lawsuits and legal notices are labeled as allegations and never scored as fact. When authorities disagree, the ingredient page shows each position with its source and date.

Not scored yet, and confidence

No score (“Not scored yet”) when:

• a food, drink or baby product has no ingredient list we can read. Nutrition facts alone can’t tell what’s in a product: scored on nutrition only, every one would land at 70–79 whatever it is. We show what we do know (allergens, high sugar or salt) and ask for a photo of the label;
• a pet, beauty, household or supplement product has no ingredient list we can read;
• the product kind is “other”.

A photo of the label from any shopper fills in what’s missing for everyone: AI reads the text, these rules score it, and it’s labeled “being checked” until a second photo or our team confirms it.

Every score also has a confidence level, based on how much of the ingredient list we recognized:

• High: at least 85% of ingredients recognized, nutrition present (for food, drinks and baby food), and a label photo from the last two years.
• Low: fewer than 60% recognized, an unknown ingredient list, or facts read by AI from shoppers’ photos (they stay low until the open sources carry the product).
• Medium: everything else.

Same product, same score

The score engine is deterministic: the same facts with the same knowledge base and parameters give the same number, on every device, every time.

• One specification, several implementations: a reference engine on our servers and the engine inside the iPhone app (Android later). Each must reproduce a shared set of test cases, built from real products and edge cases, exactly.
• Your phone computes scores itself from the offline pack, so a scan gives the same answer with or without signal.
• Every receipt prints score v1.3.0 · kb 2026.10.09. A score only changes when the product’s facts change or a newer knowledge base ships, and the newer knowledge base always wins.

Lenses

A lens is a standard you choose. It runs on your phone, from the same product facts, and gives a personal verdict. It never changes the universal score.

Lens	What it checks
Balanced (default)	Regulators and strong science. Dealbreakers: an ingredient banned or revoked somewhere, or of very high concern (c4). Warnings: high-concern ingredients (c3), EU warning-label dyes, ultra-processing (NOVA 4), high sugar or salt.
Real Food	Seed-oil-free and minimally processed. Dealbreakers: seed oils, synthetic dyes, artificial sweeteners. Warnings: natural flavors, ultra-processing, refined sugar.
Ultra-Processed Free	Dealbreakers: NOVA 4, or the ultra-processing markers used to define ultra-processed food.
Low-Tox	For beauty and home. Dealbreakers: fragrance, parabens, formaldehyde releasers, PFAS, phthalates, chemical UV filters.
Kids	Dealbreakers: EU warning-label dyes, caffeine, non-nutritive sweeteners. Warning: more than 10 g of added sugar per 100 g or ml.
Pregnancy	Dealbreakers, in food and cosmetics: retinoids, high-dose salicylic acid, hydroquinone, raw-milk cheese, high-mercury fish, alcohol and certain herbs.
Blood-Sugar Friendly	Dealbreakers: added sugars and syrups above set thresholds. Warning: refined starches.
Allergy-Safe	Driven by the allergens in your profile. “May contain” is a dealbreaker by default (fail closed); you can relax it to a warning.

For each profile in your household, GoodSwap checks, in order: allergens (“contains” is a dealbreaker, and so is “may contain” unless you relax it), diet (a violation is a dealbreaker; when we can’t confirm, say, vegan, it’s a warning), your own avoid list, then the lens. Child and pregnancy profiles add the Kids and Pregnancy rules. We never claim a product is halal or kosher without a certification label; we only flag known incompatible ingredients.

Lenses explain labels against a standard. They aren’t medical advice. Always check the package, because recipes change.

What AI does, and doesn’t

• It reads. When a product isn’t in our database, you can photograph the front and the ingredients. An AI model reads the photos and returns facts: name, brand, ingredients, nutrition. Before the first photo is sent, a consent screen names the AI provider.
• It doesn’t score. The facts go through the same rules as every other product. AI never writes a receipt line, picks a tier or chooses a swap.
• It’s labeled. Results read by AI are marked “AI-read”, carry low confidence, and are confirmed by later scans or by review.

Data and updates

• Product facts come from Open Food Facts contributors (food, with Open Beauty Facts, Open Products Facts and Open Pet Food Facts for other kinds), USDA FoodData Central branded foods, and labels users photograph. See open data.
• The knowledge base (ingredients, evidence, regulatory status, references) is ours. Every reference is a primary source: a law, a regulator, IARC, an agency assessment, a peer-reviewed paper or a dated independent test.
• Updates: offline product packs are rebuilt weekly with daily updates; the knowledge base and scoring parameters reach the app weekly, without re-downloading products.

Calibration

Before a version ships, we run it on a fixed set of real US reference products, from plain bottled water, rolled oats and extra-virgin olive oil to soft drinks and flavored chips, each with a target range. Parameters may only move within the ranges published above. We also check the overall picture: across US foods with known ingredients, the median should land around 40–55, and fewer than 20% should be Excellent. What stays Excellent is whole or minimally processed food, which the rules never deduct points from.

Corrections

Found a wrong ingredient, a missing product or a score that looks off? Tap Report an issue on the result, or email contact@goodswap.app with the barcode. Fixes to product facts reach everyone with the next data update. Changes to the method get a new version number and an entry below. Brands can’t pay to change a score.

Changelog

1. 2026-10-09 · swaps

   Swaps are the same kind of product. Tortilla chips for tortilla chips, cola for cola, shampoo for shampoo: every product belongs to one of 220 product types, or to none when we can’t tell, and then it gets no swap rather than a wrong one. Swaps used to come from a database category, which was often missing or too broad. Products of its kind that score at least 10 points higher are ranked by how much better they score, how close they are (same flavor and form, similar calories) and how easy they are to find in stores, with at most two per brand. A product with no score shows the best rated of its kind.

2. 2026-10-09 · score v1.3.0

   No number without the ingredient list. Food, drinks and baby food with no ingredient list we can read are “Not scored yet” instead of scored on nutrition alone, which put every one of them at 70–79. The result shows what we do know, and one photo of the label fills it in for everyone.

   No score for alcoholic drinks. The WHO’s cancer agency classifies alcoholic beverages as carcinogenic to humans (Group 1). Scored on their ingredient lists, spirits came out near 100, a number that would read as “good for you”. Beer, wine, spirits and cider now get no score; their non-alcoholic versions are scored like any drink.

   Unnamed colors aren’t called synthetic dyes. US law requires certified dyes such as Red 40 to be named on the label; “color added” and “artificial color” are only allowed for colors exempt from certification, such as caramel color or annatto. Those wordings no longer count as synthetic dyes.

3. 2026-10-01 · score v1.2.0

   Incomplete facts can’t be Excellent. Food, drinks and baby food with no ingredient list, a list we can’t read, or no nutrition facts are capped at 79. Missing data used to look like a clean label. Plain water, coffee, tea, herbs, spices and fresh produce need no nutrition facts and aren’t affected.

4. 2026-10-01 · score v1.1.0

   High-concern ingredients weigh more. In food, a high-concern (c3) ingredient now costs −20 (was −15.5) and a very high one (c4) −25 (was −20). Two new caps at 79 keep these products out of Excellent: any c3 ingredient, and being high in fat, saturated fat, sugars or salt (the UK front-of-pack levels). A product that is high in one of them gets no food bonus.

5. 2026-10-01 · score v1.0.0

   First publication of the GoodSwap method: five bands, the receipt, evidence tiers, caps, whole-food exemptions, confidence, “Not enough data”, and eight lenses. Calibrated against the reference products: NOVA 4 −18 and NOVA 3 −8 (the top of their ranges), and a high-concern ingredient −15.5.