How Aurascan decides what to tell you
No black box. Every verdict starts from trusted product data, is researched ingredient by ingredient against authoritative sources, and carries the citations behind it. Here is exactly how it works, and where we draw the line on what we'll claim.
Universal coverage · Verifiable citations · Deterministic results · No sponsors
The short version
What we grade
From beneficial to concern
Findings sit on one scale. Benefits and concerns are tracked independently, so one never cancels the other out.
- Beneficial
- Neutral
- Mild concern
- Concern
- Serious concern
- Serious harm
Evidence tiers
What we don't claim
From barcode to verdict
Five stages turn a number on a package into something you can act on. Each one is auditable.
- 1
Scan the barcode
You point your camera at any packaged product's barcode. That number, a GTIN, is the only thing we need to start. - 2
Resolve the product
We look the product up across Open Food Facts, USDA FoodData Central, and a barcode database, and merge what each one knows into a single set of facts. We also decide, once per product, how it reaches your body: eaten, applied to skin, or another route. - 3
Research each ingredient
Every ingredient is researched individually against authoritative sources. Each claim we surface carries the citations behind it. - 4
Verify the citations
A verifier checks each cited authority against the sources actually retrieved. Anything it can't tie to a real source gets rewritten in neutral terms or removed, never left implying false authority. - 5
Deliver a transparent verdict
You get plain-English findings scoped to how the ingredient actually reaches your body, tiered by how often you use it. Never a black-box grade.
No dead-ends. No “untested.”
Some tools stop cold when a product isn’t in their hand-curated list. Aurascan researches every ingredient, so a scan returns a real answer instead of a shrug. When an ingredient isn’t in our research library yet, it reads Researching while it is researched in the background, and the result updates when that research lands. When the evidence is thin or missing, we say so plainly, which is different from having nothing to show you.
And when no database has the product at all, you can add it by photographing the label, so a missing barcode isn’t a dead-end either.
Where the product data comes from
We combine four sources and keep the best field from each. Every scan credits only the sources that actually contributed to it, not a generic list.
Open Food Facts
ODbLA collaborative, open database of packaged foods. We use it for standardized additive, allergen, and processing tags, plus product identity.
USDA FoodData Central
Public domain (CC0)The U.S. government's authoritative branded-food dataset. When it has a product, its label ingredients and nutrition numbers take precedence.
Go-UPC
LicensedA barcode lookup that fills the long tail: product name, category, and image for items the open datasets haven't catalogued yet.
Community submissions
User contributedWhen no database has a product, a shopper can photograph its front, ingredient list and Nutrition Facts panel to add it. It stays private to them until we review it; approved products join the catalog, credited and clearly marked as user-contributed.
How we research each ingredient
AI does the research, in the open
Each ingredient is researched by a large language model (Google’s Gemini) with live web grounding. We’re deliberately transparent about that: other tools either bury the fact that AI is involved or hide their methodology entirely. We’d rather show you the machinery and the sources, so the research is something you can check, not something you have to take on faith.
Every citation is checked, and false authority is stripped out
A claim without a source is just an opinion. When the research says “according to the FDA,” a verifier checks that an FDA source was actually retrieved to back it. If it can’t tie the authority to a real source, it rewrites the phrase to a neutral one (“according to research”) or drops the sentence entirely, never leaving language that implies an authority we can’t point to.
The allowlist is tiered, with regulators and health agencies ranked above journals: FDA, EFSA, the EU scientific committees, CIR, EWG, IARC, JECFA (WHO/FAO), NIH/PubMed, EMA, MHRA, and Health Canada, plus peer-reviewed hosts like Nature, The Lancet, and the BMJ.
Research continues after the scan, and shows its confidence
Ingredients missing from our research library are queued and researched in the background, one at a time, and ingredients below high confidence are queued to be researched again. Every graded ingredient carries a research confidence meter (low, medium or high), scored from the stored research: how many sources it draws on, how authoritative they are, and how well its claims are cited.
When there’s no good evidence, we say so
The most important thing we do is refuse to make things up. When an ingredient has no authoritative sources, or the research can’t be grounded, we return an unresolvedresult with the reason, not a confident-sounding guess. A blank we’re honest about beats a verdict we can’t stand behind.
Not all evidence is equal
A claim backed by a randomized human trial is not the same as one backed by a mechanistic hypothesis. We tag research by its evidence tier so you see not just what we know, but how well we know it.
Tier 1
Human clinical trials
Randomized controlled trials in humans: a direct measurement of cause and effect in real people. When available, this is what we cite first.
Tier 2
Human observational studies
Prospective cohorts and case-control studies. Real people, real outcomes, but correlational: we can measure what happens, not always prove why.
Tier 3
Animal studies
Controlled studies in animal models. Directional evidence, useful for mechanism, but not conclusive for humans.
Tier 4
Mechanistic hypotheses
In-vitro experiments or plausibility-based mechanisms without in-vivo data. Interesting signals, but we flag these so you know the evidence is thin.
Two independent signals
A verdict draws on two channels that don't depend on each other, so a gap in one doesn't silence the other.
Deterministic regulatory data
A hand-curated regulatory seed: FDA-prohibited substances and EFSA/IARC-cited additive flags. The same inputs produce the same signal every time, with no model in the loop.
Includes genotoxicity classifications from EFSA OpenFoodTox 3.0, licensed CC-BY-4.0.
Grounded AI analysis
Benefits and concerns are independent
An ingredient can carry a benefit and a concern at the same time. We don’t collapse them into one number that cancels the nuance out.
Scoped to how you’re exposed
A concern about inhaling something doesn’t apply to eating it. We decide once, per product, how it reaches your body (eaten or drunk, applied to skin, and so on) and scope every signal to that route. For products you don’t eat, such as cosmetics and personal care, food-only details like nutrition are left out.
Dose matters
“Should you care?” is tiered by how often you eat it. An occasional treat and a daily habit are not the same question. The amount counts too: alcohol is graded by its ABV, an ingredient at a high dose per serving is graded more cautiously, and nutrition is shown for the whole container when its size is known.
The same product, the same answer
AI is famously non-repeatable: ask twice, get two answers. We engineer that out. Every analysis is cached against the product’s normalized barcode, the methodology version it was produced under, and your profile. Scan the same product again and you get the identical result, byte for byte, until we explicitly publish a new methodology version and regenerate. When the science changes, we bump the version on purpose. Results never drift on their own.
1:1
same product + profile, same result
GTIN-14
normalized barcode in the key
Versioned
every result is stamped
Independence pledge
The only thing that shapes what we tell you is the evidence. Nothing else is for sale.
No paid placements
No ads, no affiliate cut
Industry funding counts against a source
Funded by subscriptions
Where we draw the line
- AI can be wrong. Research is a starting point, not the final word. We show sources and flag weak evidence so you can judge for yourself.
- This is not medical advice. Aurascan is educational and is not intended to diagnose, treat, or replace guidance from licensed clinicians or nutrition professionals. If an ingredient interacts with a medication you take, talk to a doctor, not a phone app.
- Found a mistake? Tell us. Every ingredient has a report-a-problem path, and every result is stamped with the methodology version behind it, so a correction is traceable, not lost.
Frequently asked
Does AI write the analysis?
Yes, and we're upfront about it. Each ingredient is researched by a large language model (Google's Gemini) with live web grounding, then every claim is checked against an allowlist of trusted sources. We show you the citations so you can verify the research yourself, rather than asking you to trust a hidden model.
What happens when there's no good evidence?
We tell you. If an ingredient has no authoritative sources, or the research can't be grounded, we return an unresolved result and say why. We never invent a verdict to fill the gap.
Will the same product give the same result every time?
Yes. Results are cached against the product's normalized barcode, the methodology version they were produced under, and your profile, so a given product returns the same analysis for the same profile until we explicitly publish a new methodology version.
Do you give products a single score or grade?
No. A single letter or number hides the reasoning and treats one bite the same as a daily habit. We show what each ingredient does, scoped to how it reaches your body and how often you use it, and let you decide.
How sure is the research on each ingredient?
Every graded ingredient shows a research confidence meter, low, medium or high, scored from the research itself: how many sources it has, how authoritative they are, and how well its claims are cited. Ingredients that aren't in our research library yet read Researching while they are researched in the background, and ingredients below high confidence are queued to be researched again.
Can the analysis be wrong?
It can. Science evolves, databases have gaps, and AI can make mistakes. That's why we show our sources, flag weak evidence, stamp every result with a methodology version, and give you a way to report a problem.
Published April 19, 2026 · Updated October 3, 2026
See what's in the products you buy
Scan a barcode and read what every ingredient does, in plain English.