How Sift works

Private beta. This page explains what happens to your photo and where the answers come from.

Two steps, not one

A scan runs in two separate halves, and they do different jobs.

1. An AI model reads the photo. Google Gemini looks at your picture and writes out the words printed on the label, and reads the product name off the front of the pack. It is never asked whether the food is safe for you, and it never decides what you are warned about.

2. Ordinary software checks the words. Sift's own code takes that text and compares it against fixed lists of ingredient terms. Same text in, same answer out, every time. There is no judgement call and nothing is made up.

Splitting it this way is deliberate. The part that can be creative only transcribes; the part that decides what to warn you about cannot be creative.

Receipts: every match shows its wording

When Sift flags something, the result carries a Receipts section. Open it and you see the exact phrase printed on your label, which of your allergens it matched, and which part of the label it came from: the ingredient list, a "contains" statement, a "may contain" warning, or wording too ambiguous to settle. Nothing is reported that cannot be traced back to words on the package.

If no printed phrase matched, the receipts say that too, plainly, instead of implying the product was checked and cleared.

Where the lists come from

• Allergens. The 16 allergens Sift knows about are the ones named in law: the United States major food allergens under FALCPA (21 U.S.C. 321(qq)) and the FASTER Act of 2021, and the substances listed in Annex II of EU Regulation 1169/2011. Every one of them carries its own citation in the data, so each can be checked against the law it came from.
• Nutrition percentages. Reference amounts are the FDA Daily Values printed in the table at 21 CFR 101.9(c)(8)(iv), for adults and children aged 4 and over. They are copied from that regulation. Nothing is estimated.
• Diet preferences and ingredient notes. These lists are Sift's own work. No government defines what counts as vegan, and Sift does not pretend otherwise. Each group records where its terms came from so it can be reviewed.

The law says which allergens exist. It does not list every word a manufacturer might print for one: "casein" and "whey" both mean milk. Those word lists were drafted with software and are being checked by hand, group by group, against the source each one names. Some groups have had that check and some are still waiting on it, and the data records which is which. A group that has not been checked yet can only ever add a warning, never remove one, so it is safe to use while the check is outstanding.

Reference lists last updated

These are the dates the underlying data files were last generated, read from the files themselves.

• Allergen terms: 24 August 2026
• Diet ingredient terms: 28 August 2026
• Ingredient note terms: 28 August 2026
• Nutrition reference amounts: 9 September 2026

What this does not mean

Sift reports what it could read. A photo can be blurred, cropped or out of date, transcription can be wrong, and a package can change. An ingredient that is not flagged has not been proven absent.

Sift gives no score and no safety rating, and it does not certify any product as vegan, gluten-free, halal or kosher. It is not medical advice. Read the package, follow the advice of your own clinician, and ask the manufacturer when wording is unclear.

More detail on handling and storage is in the Privacy Policy, and on the limits of a result in the Terms and Conditions.