We weighed everything for a week before logging it. The apps were not the biggest source of error.

Two of us weighed and recorded every item we ate for seven days, converting weights to nutrients against reference composition data. Then we logged the same food the way we normally would — estimating portions, tapping through an app.

The finding

The gap between our weighed record and our estimated log was substantially larger than the gap between the best and worst app in this category.

Our portion estimates were off by more than any app’s estimation engine was. Both of us under-estimated, both of us under-estimated the calorie-dense items specifically, and neither of us had any sense that we were doing it.

This is well documented in the nutrition literature and it was still surprising to experience.

Where the error came from

Fats and oils. The single worst category, and by a distance. A “splash” of olive oil weighed between 11g and 24g depending on the day and the pan. That is a spread of over a hundred calories from an ingredient neither of us was consciously tracking.

Anything poured. Cereal, rice, pasta, nuts. Volume estimation is bad and gets worse with density.

Things eaten standing up. Neither of us logged the pieces of cheese eaten while cooking until we were weighing deliberately. That is not an estimation error, it is an omission error, and it may be the largest one of all.

What this means for the app question

It reframes it rather than dismissing it.

Estimation error and portion error compound rather than substituting for each other. An app with a 2% error band applied to a portion you over-poured by 30% gives you a very precise number about the wrong food.

So the practical order of operations is: weigh what you can weigh, which for most home cooks is the majority of their calories, and use a good estimator for the rest — the restaurant meals, the shared dishes, the things you did not cook.

That also means the app property that matters most in daily use is whether entering a known weight against a real record is easy. Apps built around a verified catalogue make this trivial. Apps built around a photograph with a thin database behind them fight you when you try, which is a design choice that quietly pushes you toward the less accurate path.

The cheerful conclusion

A kitchen scale costs about twelve pounds and has no subscription.

For the foods you can put on it, it beats every estimation figure any lab has published for any app, because it is not estimating. Given that the best independently verified figure in this category is ±1.1%, that is a low-cost way to beat the state of the art for a large share of your week.

Use the app for what a scale cannot reach. That is a division of labour, not a criticism of either.