The Calorie Count on Your Meal Plan Is Probably Made Up

Ask an AI meal planner for a recipe and it will hand you a calorie count, a protein figure, maybe a full macro breakdown. It all looks precise. Here is the uncomfortable part: in most cases nobody checked those numbers against anything. A language model produced a recipe, then produced some plausible-looking figures to sit underneath it, in the same breath and by the same method. The recipe is a guess. The numbers are a guess about the guess.

That is not a swipe at anyone in particular. It is how these tools are built, and until recently it was how ours worked too. If you had asked Nutridly in July how many calories were in a recipe, the honest answer would have been "the model reckons roughly that." Which is fine for a rough idea, and not fine at all if you are making decisions about iron or protein or fibre off the back of it.

So we changed it. Every calorie, macro and micronutrient figure in a Nutridly plan is now calculated from McCance and Widdowson's Composition of Foods Integrated Dataset, the UK government's own food composition database. It is the reference the nutrition profession here actually works from: nearly 2,900 foods, each one measured in a laboratory, published by Public Health England. We match your recipe's ingredients against it one by one, scale by the weight in the recipe, and add it up. There is no model in that path at all. Ask twice, get the same answer.

The interesting design question was what to do when we cannot match something. Chia seeds, for instance, are not in the 2021 dataset. Neither is almond milk. Our first answer was to show nothing at all, on the principle that a blank is more honest than a number you cannot trust. That was half right. The real problem was never estimating, it was estimating quietly. So now, where too little of a recipe can be accounted for, you get a figure with a tilde in front of it and the word "Estimated" next to it, and the recipe says plainly that we could not match enough of its ingredients. You can still use the number. You just always know which kind of number it is.

One place we do not budge: micronutrients are never estimated. A calorie figure that is roughly right is useful. A nineteen-nutrient spectrum assembled from guesses would look authoritative and mean nothing, and it is exactly the sort of thing someone might act on. So the spectrum only ever draws on measured values, and stays quiet about the rest.

What this unlocks is the part we are most pleased with. Once the figures are real, you can do real things with them. Your plan now shows a nutrient spectrum for the week: all nineteen vitamins and minerals we track, each one against the target calculated for your age, sex and life stage. Not a tick saying "balanced diet." An actual shape you can look at, where a short spoke means a genuine gap in the week ahead, in time to do something about it.

And it closes a loop. Mark the meals you actually cooked, and next week's plan reads what you really ate rather than what was offered. If your logged meals ran low on B12, the following week leans toward foods that carry it. That only works because both halves of the sum are real, and it is the sort of thing you simply cannot build on top of numbers a model invented.

The wider point is not about calories. It is that "personalised nutrition" has become an easy claim to make and a hard one to check. A plan can reference your blood results, use your name, sound convincing, and still be built on figures nobody verified. Worth asking of any tool you are trusting with this: where does that number come from, and what happens when the honest answer is that you do not know?