The food that calorie apps forget

10 September 2026

Open a mainstream calorie tracker and search for waakye. You will probably get nothing. Search for goat meat and you may get a raw cut measured per 100 g, which is not what you ate. Search for a bottle of malt and you'll get a scattering of user-submitted entries with wildly different numbers and no way to tell which one is right.

This is not an oversight anyone forgot to fix. It follows directly from how those apps are built.

A database is a list of decisions

A conventional tracker is a search box in front of a food database. That database was assembled somewhere, by someone, from the products on shelves near them. Everything in it is a decision that a particular food was worth including.

Those decisions were mostly made in North America and Western Europe. So the database is excellent on breakfast cereal and supermarket ready meals, thin on anything cooked in a home kitchen, and close to useless for a dish that is ordinary in Accra or Lagos and unheard of in the dataset's home town.

The gap gets papered over with user submissions, which is how you end up with six entries for the same drink disagreeing by a factor of three.

What that actually costs

Not accuracy — abandonment.

If two meals a day can't be logged without a fight, the day's total is wrong, the app knows it's wrong, and you stop opening it. Nobody quits a calorie tracker because it was off by forty calories. They quit because logging dinner took four minutes and ended in a guess.

The failure mode of a food database isn't a wrong number. It's a person who gives up.

Describing instead of searching

Ragus doesn't have a food database to be missing from. You describe what you ate, and it works out what that means — the same way a dietitian does when you tell them about your day. “Goat meat, some rice and a bottle of malt” is a complete input. There is no list for it to be absent from.

The first time someone logs an unfamiliar item, that costs a real lookup. The answer is then remembered and shared, so the next person who mentions the same thing gets it instantly and for nothing. Over time this builds a record of what the people using the app actually eat — which, for regional food, is worth considerably more than any dataset you could buy.

What we're watching

The interesting failure is near-duplicates. “Goat meat”, “goat meat chunks” and “goat meat chunks with skin and bone” are three ways of saying roughly one thing, and treating them as three separate foods is both wasteful and slightly wrong. Teaching the system that they're the same food described at different levels of detail is an ongoing piece of work.

It's a better problem to have than an empty search result.