Why calorie counts are ranges, not single numbers

The same food record, coded 53 ways#

Calorie counts are ranges because the number moves depending on who reads the meal, not only on what was eaten. The cleanest demonstration of this held everything else still: 53 members of Sports Dietitians Australia were handed seven-day food records from athletes on the 1996 Australian Olympic team and asked to code them using the same database and the same software package. Same records, same tools, trained professionals — the only thing that changed was which dietitian did the reading. Across 52 athletes and 1,456 athlete-days, their totals still disagreed with one another, and the coder's own decisions came out as a measurable source of variance in the data¹.

Nothing malfunctioned in that study. The spread is the measurement. When a tracker prints "650 calories" it has drawn one value out of that distribution and thrown the distribution away — and the discarded half was the part that told you how much to trust the 650. A range keeps both. What follows is what a width actually tells you, how wide it should be for a given food, and why the same uncertainty behaves completely differently across a week than it does at a single meal. For the audit of where the error enters in the first place — labels, portions, absorption — start with how accurate calorie counting is.

A width is a claim about evidence, not a hedge#

The usual objection to ranges is that they sound evasive: just give me the number and let me deal with it. But a width is not a mood. It is a statement about how much evidence an estimate is standing on, and it is measurable.

Because Braakhuis's team ran every nutrient through the same pipeline, they could rank them. The most variable — vitamin C, vitamin A, cholesterol — carried roughly threefold more variability than the least variable: energy, carbohydrate, magnesium¹. Energy sits at the stable end, which is worth pausing on, because it inverts the folk wisdom that calories are the flakiest figure in a food log. Calories are the sturdiest number in the panel; the micronutrients printed beside them are the guesses. Consumer software arrives at the same ranking from a completely different direction — calorie-tracking apps validate well on energy and fall apart on sodium.

So width is not a constant. It should move with what you logged.

Those are three different metrics and cannot be subtracted from one another, so read the table as a ladder rather than a scale. The ordering is the finding. A sealed package that a factory reproduced a million times earns a band a few percent wide; a plate a cook improvised earns one several times that; a photograph of that plate, with the oil absorbed and the underside hidden, earns wider still. Each rung is argued in full elsewhere — nutrition labels, restaurant calorie counts, and AI photo calorie counters.

A practical test falls out of it. A tool whose ranges are all the same width is not estimating. It has picked a number and dressed it in a margin. If a barcode scan and a described curry come back at the same plus-or-minus ten percent, the band is decoration — a different failure from merely getting the margin of error wrong.

Two kinds of error, and only one of them cancels#

Here is the part that changes what you do on a Tuesday. The uncertainty in a food log is not one substance. It is a random component that shrinks when you stack days and a systematic component that does not, and the two call for opposite responses.

The random part collapses quickly. In the Braakhuis data, variability in the mean of seven-day estimates ran two- to threefold lower than the variability of a single day¹. The coder who read Tuesday generously read Thursday meanly, and the noise partly cancels against itself.

How much averaging that takes is easy to underestimate. In a year-long study at the USDA's Beltsville Human Nutrition Research Center, 29 adults kept food records, and researchers calculated how many days of records are needed to land within 10 percent of a person's usual intake at 95 percent confidence. For food energy — again the easiest nutrient in the panel — the answer averaged 31 days for an individual, against just 3 days to pin the mean for the group. Vitamin A needed 433 days individually².

Even a flawless logger needs about a month of records to know their own usual calorie intake within ten percent. One day was never going to settle it.

That asymmetry — 31 days for you, 3 days for the group — is why population nutrition science works while your personal daily number feels like static. It is also the strongest argument here for reading the week instead of the meal: not because a day doesn't count, but because a day carries too little information to be read as a verdict.

Now the component that does not cancel. Averaging only helps against errors that change sign. It does nothing to a lean that runs the same direction every day, and self-report leans low: in the landmark demonstration, subjects who believed themselves diet-resistant were under-reporting their intake by 47 percent, with entirely normal metabolisms⁴. Stack thirty biased days and you get a precise estimate of the wrong number. So "log more days" is excellent advice against noise and no help whatsoever against bias — the first needs repetition, the second needs calibrating against what your weight actually does over a few weeks.

What the field says about its own numbers#

In 2015 a group of nutrition-measurement researchers answered critics who had argued that underreporting made self-reported diet data worthless. Their verdict was that the data are valuable and should keep being collected. Then they issued seven recommendations for using them, and the second is one sentence long:

"Do not use self-reported energy intake as a measure of true energy intake" — recommendation 2 of 7³.

That is the field defending food logging, not attacking it. The same paper asks researchers to "acknowledge the limitations of self-report dietary data and analyze and interpret them appropriately," and to design analyses that adjust for measurement error rather than assume it away³.

A range is what that instruction looks like when a consumer tool takes it seriously. A single printed figure cannot comply, because presenting one number is the claim to be the true intake. A band can: it reports the estimate and its own resolution in the same glance, and it leaves you able to tell a barcode from a guess.

Reading a range without obsessing#

• Spend your attention where the band is wide. Width is a report on the evidence, not a property of the food. A narrow band had something solid to stand on and needs nothing from you; a broad one is missing detail that only you hold — which is the one place a correction buys anything.
• Judge the week, not the meal. A band 80 calories wide on lunch is noise at the scale of seven days — and the seven-day mean is exactly where the random half of the error has already cancelled itself down two- to threefold.
• Anchor what you repeat. Weigh your five most-frequent meals once. Repeated foods carry repeated width, so collapsing it once collapses it permanently.
• Calibrate the bias separately. Your weight trend over three or four weeks is the only instrument that can catch a consistent lean. No amount of careful logging will find it, because careful logging is what produced it.

The goal was never the true number. A month of records is what the true number costs, and nobody is buying. The goal is a picture honest about its own resolution, cheap enough that you are still keeping it in six months.

FAQ#

Are calorie ranges less accurate than exact numbers?#

No — they are the same estimate with the uncertainty left visible. A confident "650" and a range of "600–720" typically come from identical data; the range simply stops discarding the part that says how much to trust it. The precision of a single number is typographic, not empirical.

How wide should a calorie range be?#

It should track the evidence behind the food. A packaged item scanned from a barcode justifies a band a few percent wide — measured energy across 24 snack products ran a median of 4.3 percent over the label⁵ — while an unaided estimate from a meal photo sat 26.9 percent off on average⁷. If a tool shows the same width for both, it is decorating rather than estimating.

Do ranges matter if I weigh all my food?#

Less, but they don't disappear. Weighing removes portion error, the largest single term. It does not touch the fact that a database entry is an average, that a coder's choices moved totals even for professionals working from identical records¹, or that pinning your own usual intake within 10 percent took about 31 days of records².

Sources#

1. Braakhuis AJ, et al. Variability in estimation of self-reported dietary intake data from elite athletes resulting from coding by different sports dietitians. Int J Sport Nutr Exerc Metab. 2003.
2. Basiotis PP, et al. Number of days of food intake records required to estimate individual and group nutrient intakes with defined confidence. J Nutr. 1987.
3. Subar AF, et al. Addressing current criticism regarding the value of self-report dietary data. J Nutr. 2015.
4. Lichtman SW, et al. Discrepancy between self-reported and actual caloric intake and exercise in obese subjects. N Engl J Med. 1992.
5. Jumpertz R, et al. Food label accuracy of common snack foods. Obesity (Silver Spring). 2013.
6. Urban LE, et al. Accuracy of stated energy contents of restaurant foods. JAMA. 2011.
7. O'Hara C, et al. An evaluation of ChatGPT for nutrient content estimation from meal photographs. Nutrients. 2025.