The 20% rule: why food labels can be legally off

'Misbranded' is a lab verdict, not a broken promise#

The calorie number printed on a package is not a measurement of the item in your hand. It is a legal declaration, and US regulation spells out exactly when that declaration becomes illegal: a food is "misbranded" for calories only when a laboratory, testing a composite of twelve units, finds the food runs more than 20 percent over the printed figure¹. Below that line the label is fully compliant. So the printed number is already a one-sided range by construction — reading it as the middle, or the floor, of a band a fifth wide is precisely correct, not sloppy.

Two facts make the 20 percent unavoidable rather than a loophole, and this article works through both. Real food genuinely varies from one unit to the next, so no single number could ever be right for every package. And most labels were never measured at all — the regulation lets the number be computed from a database instead of assayed from the finished food. Once you see how the figure is manufactured and how rarely it is checked, treating it as an estimate with a legal tolerance stops feeling like cynicism and starts feeling like literacy. (The empirical question of how far labels actually drift, and the rounding rules that turn cooking spray into a zero, are the neighbouring subject of how accurate nutrition labels are; this piece is about the rule itself.)

The test is never run on your package#

Start with what "compliance" physically means, because it is not what most people picture. The regulation does not imagine an inspector weighing your specific candy bar. It defines a sampling procedure: "The sample for nutrient analysis shall consist of a composite of 12 subsamples (consumer units), taken 1 from each of 12 different randomly chosen shipping cases, to be representative of a lot"¹. Twelve units are pooled and blended, and the blend is what gets analysed. The verdict is about a lot's average, not any individual package — the variation between your bar and the next is deliberately averaged away before the test even begins.

What that averaged composite is measured against depends on the nutrient, and the rule splits into classes¹:

The class split matters because it reveals what the tolerance is for. Nutrients a manufacturer adds and advertises — the vitamins in a fortified cereal — must be present at 100 percent of the label or more. Calories, fat, sugar, and sodium get the 20 percent excess allowance. The law is protecting the eater from overstated benefit and understated cost; the direction the slack runs is worked through in full in how accurate nutrition labels are.

The slack leans the unfriendly way#

There is a clause most people never read that settles which direction a wrong label tends to point. The regulation states that "reasonable deficiencies of calories, total sugars, added sugars, total fat, saturated fat, trans fat, cholesterol, or sodium under labeled amounts are acceptable within current good manufacturing practice"¹.

Read that plainly: a package containing fewer calories than its label claims is explicitly fine, and a package containing more is fine up to the 20 percent ceiling. For a calorie counter the consequential half is the ceiling — the printed number is a figure your food is legally allowed to exceed but only quietly allowed to undercut. When a calorie label is wrong, the structure of the rule makes it likelier to be wrong low than high, which is the opposite of the direction a person trying to eat less would hope for.

The calorie number on a package is a legal claim about a twelve-unit lab composite, held to a one-sided 20 percent tolerance. It was never a measurement of the item in your hand, and it never promised to be.

Why the tolerance is 20 percent and not 5#

A fifth sounds enormous until you ask what a tighter number would demand. It would demand that broccoli, flour, a cut of beef, and the filling squirted into a cookie all hit the same value in every batch — and they do not, because food is a biological product, not a machined part.

The scale of that variation is documented by the reference system the labels themselves lean on. The USDA's Foundation Foods dataset publishes not just a mean for each food but "the individual data points behind the mean values," specifically so users can see "the variability in the nutrient and food component values provided as well as the potential effects of the production site, procedures, season, post-harvest processing, analytical methods, and other factors"⁴. Even the gold-standard database treats a food's composition as a distribution with a spread, not a fact. Season, soil, cut, ripeness, and the assay itself each move the number, and they move it by more than a few percent. A 5 percent tolerance would make a compliant label physically impossible for real food; the reg's own language concedes as much when it accepts "reasonable" variation within good manufacturing practice. The 20 percent is an engineering allowance for the fact that the thing being labelled refuses to hold still.

Most labels were never measured at all#

Here is the fact that reframes the whole tolerance. The 20 percent legal room sits on top of a number that, more often than not, was never the result of burning food in a calorimeter. The regulation permits compliance to be "provided by use of an FDA approved database that has been computed following FDA guideline procedures"¹. A manufacturer may build its calorie figure by adding up the database values of the ingredients in the recipe, adjusting for expected processing losses — no assay of the finished product required.

So the printed calorie count is usually a calculation, and calculations fork. The same recipe can be computed several legal ways, which is one reason two apps hand you different numbers for the same food (why calorie estimates vary takes those forks apart). Stack that on top of the sampling tolerance and the picture is complete: a computed number, rounded by rule, sold under a 20 percent enforcement band, tested against a twelve-unit composite that is rarely pulled. The wonder is not that labels drift — it is that they land as close as they do.

What regulators find when they actually test#

They do occasionally pull products and burn them, and the results are the most useful part of this whole topic because they tell you which direction to worry about. Canada's food inspection agency sampled 1,010 products from retailers between 2006 and 2010 and analysed them against a ±20 percent tolerance analogous to the US rule. Overall, 16.7 percent came back "unsatisfactory," and for calories specifically 14.2 percent of tested items exceeded the band — with every single unsatisfactory product sitting on the excess side, meaning the label understated what was actually in the food. Sodium and calories were consistently under-reported².

US packaged food looks better, and the reason is standardization. When 24 popular American snack products were measured, metabolizable energy ran a median of just 4.3 percent over the label, and most sat comfortably inside the 20 percent allowance³. Two studies, two countries, one direction: when a sealed, machine-made product's label misses, it tends to miss low, and usually by a single-digit percentage rather than the full legal margin. The 20 percent is the outer wall of the room, not the size of the average error inside it.

So how should you read the number?#

As the centre of a band, graded by how the food was made. A barcoded, sealed staple is the most trustworthy calorie figure in your day — a median 4.3 percent overshoot beats every other estimate you will make³ — so log it and move on. But log it knowing the printed figure is the middle-to-floor of a legally sanctioned range, not a fact, and that the same tolerance which keeps honest manufacturers legal also lets the number drift a few percent under what you actually eat.

That is not a reason to distrust labels; it is a reason to stop asking them for a precision they were never built to carry. The tolerance is a feature — the only workable response to a food supply that varies — and treating the printed number as one point in a range is the accurate way to use it, the same logic that makes every calorie count a range rather than a single number, and one layer of the broader accuracy stack beneath any total.

FAQ#

How does the FDA actually check whether a calorie label is accurate?#

By analysing a composite of 12 units "taken 1 from each of 12 different randomly chosen shipping cases," and declaring the food misbranded only if that pooled sample runs more than 20 percent over the label¹. It is a test of a lot's average, not your package — and one that is run only rarely, since the regulation also lets a label's value come from an approved database instead of an assay.

Why is the legal tolerance as wide as 20 percent?#

Because the food genuinely varies that much. The USDA's own reference dataset publishes the spread behind each food's average, driven by "production site, procedures, season, post-harvest processing, analytical methods, and other factors"⁴. A tighter tolerance would make a compliant label impossible for a biological product, so the rule allows "reasonable" variation within good manufacturing practice.

Does a 20% tolerance mean my snack has 20% more calories than it says?#

No. The 20 percent is a ceiling on a 12-unit lab composite of a whole lot, not a prediction about your one package — your unit could sit over or under the label by an unknown amount. What the evidence does show is the direction: when packaged labels miss, measured calories tend to run over the printed figure, a median of about 4.3 percent in US snacks³ and further in less standardized foods.

Sources#

1. 21 CFR 101.9 — Nutrition labeling of food. Code of Federal Regulations (Cornell Legal Information Institute).
2. Fitzpatrick L, Arcand J, L'Abbe M, et al. Accuracy of Canadian food labels for sodium content of food. Nutrients. 2014.
3. Jumpertz R, Venti CA, Le DS, et al. Food label accuracy of common snack foods. Obesity (Silver Spring). 2013.
4. USDA FoodData Central. Foundation Foods Documentation.