How accurate are calorie counts, really?

People tend to hold two contradictory beliefs about calorie counts at the same time: that the number on a packet is essentially exact, and that calorie counting does not work because "not all calories are equal". Both are wrong, and understanding why is oddly liberating.

Where the number on a label comes from

Almost nobody burns food to find its calories. The standard method is calculation. The food is analysed for protein, fat, carbohydrate and alcohol, and each is multiplied by an average energy factor — roughly 4, 9, 4 and 7 kcal per gram. The products are summed.

These are the Atwater factors, published in the late nineteenth century and still in use because they are good enough for the job. They are averages across broad food categories, not constants. Protein in an egg does not yield precisely the same energy as protein in a lentil.

Carbohydrate is often not measured at all. It is calculated "by difference": take 100 g, subtract the measured water, protein, fat, ash and sometimes fibre, and whatever is left is called carbohydrate. Every measurement error in the other components lands in that figure.

How much tolerance labels are allowed

This is the part that surprises people. Food labelling regulations in most jurisdictions permit a difference between the declared value and the measured one, and for energy that tolerance is commonly in the region of 20 per cent.

That is not a loophole exploited by bad actors. It exists because food is a biological product: two batches of the same biscuit genuinely differ, and requiring exactness would be requiring the impossible. But it does mean a 200 kcal bar might really be 230, and a 500 kcal ready meal might be 580.

The practical consequence is that a log built entirely from labels still carries a margin you cannot remove.

Then there is what your body does with it

A calorie count describes the energy in the food. What matters is the energy you extract, and those differ.

Whole nuts are the clearest example: their intact cell walls resist digestion, and the metabolisable energy is meaningfully below the calculated figure. Grind them into butter and more becomes available. High-fibre foods behave similarly, which is one reason fibre is assigned a lower energy factor in most labelling systems.

Cooking works the other way. Heating starches and denaturing proteins makes them easier to digest, so a cooked food can yield more usable energy than the same food raw — the subject of does cooking change the calories in your food.

Home-cooked food is where the error gets large

Everything above concerns packaged food, which is the well-measured end of the spectrum. Most people eat mostly food that never had a label.

A home-cooked curry has no correct answer. The oil alone can swing it by 40 per cent, as set out in how much oil is actually in your curry. The database entry someone else typed in used a recipe that was not yours. Serving sizes are named rather than weighed.

This is why counting calories in home-cooked food needs a different method from scanning barcodes, and why two apps can hand you different numbers for the same plate, which is unpacked in why two apps give the same roti different calories.

Why none of this breaks calorie tracking

Here is the part that gets lost. A food log does not need to be accurate. It needs to be consistent.

Suppose your log reads 2,000 kcal and the truth is 2,300. If that 13 per cent gap is roughly the same every week — the same foods, the same estimates, the same habits — then holding your log at 2,000 and watching your weight for three weeks tells you everything. If weight is flat, your maintenance is 2,000 as you measure it. Subtract 300 from that and you have a real deficit, regardless of what the true numbers were.

What ruins this is inconsistency: logging oil some days and not others, weighing rice on weekdays and guessing at weekends, using a careful entry for breakfast and a random one for dinner. Variable error hides the signal; constant error does not.

How big the error is, roughly

Putting numbers on it helps, as long as they are treated as orders of magnitude rather than facts.

A packaged food logged from its label: within about 10 to 20 per cent, usually closer on high-volume products from large manufacturers, further on small-batch items.

A single ingredient weighed raw and matched to a national composition table: within about 5 to 10 per cent. This is the most accurate thing you can do.

A home-cooked dish where you counted the ingredients yourself: within about 15 per cent, dominated by how well you accounted for cooking fat.

A home-cooked dish logged from a generic database entry: anywhere from 25 per cent low to 40 per cent low, which is a different category of wrong.

A restaurant meal estimated by eye: 30 to 50 per cent low is common, and almost always low rather than high.

The pattern is consistent. Errors grow as you move away from ingredients you handled yourself, and they grow in one direction.

The asymmetry is the real problem

If errors were random, they would cancel. They do not. Under-reporting is the overwhelming pattern in dietary research, and the mechanism is not dishonesty — it is that the forgettable items are almost all calorie-dense.

Nobody forgets a plate of rice. People forget the oil it was fried in, the ghee on the roti, the sugar in three cups of tea, the handful of namkeen at six o'clock, and the two bites of a child's leftovers. Every one of those is invisible and energy-dense.

Which means the useful correction is not "be more precise". It is "look in the places errors hide", and that list is short enough to memorise.

What to do with this

Four things, in order of value.

Log the same way every day. One entry per food, reused, not a fresh search each time.

Prefer sourced entries. An entry naming a national food composition table or a manufacturer beats one typed in by a stranger, for reasons covered in how to judge a calorie app's food database.

Weigh the high-density items. Oil, ghee, butter, nuts, cheese. A 10 g error in oil is 90 kcal; a 10 g error in cucumber is nothing.

Let the scale do the calibration. Your weekly weight trend is the only instrument that measures the thing you actually care about, and it corrects for every error above at once.

Precision in a food log is mostly theatre. Consistency plus a weight trend is the real measurement system, and it works with numbers that are known to be imperfect.