No way of counting calories is exact. Labels are allowed a legal margin, food databases differ in quality, people tend to underestimate what they eat, and AI photo estimates add their own error. That does not make counting pointless: logged the same way day after day, the numbers are better at showing trends than at pinning down a single day, and systematic reviews link self-monitoring with better weight outcomes, although that evidence is modest and some of it points the other way.
How close is the number on a food label?
Usually close, but never promised to be exact, and the rules say so openly.
In the EU, the European Commission’s guidance on label tolerances starts from the fact that foods cannot always contain the exact nutrient levels labeled, because of natural variation and variation from production and storage. Declared values are meant to be averages, based on the manufacturer’s analysis, a calculation from the ingredients, or generally accepted data. Its table for foods other than supplements (the margins already include measurement uncertainty) allows:
| Nutrient, per 100 g | Below 10 g | 10 to 40 g | Above 40 g |
|---|---|---|---|
| Carbohydrate, sugars, protein, fiber | ±2 g | ±20% | ±8 g |
| Fat | ±1.5 g | ±20% | ±8 g |
Saturated fat is allowed ±0.8 g below 4 g per 100 g and ±20% from 4 g up. The table has no separate row for energy, and the document itself notes that it “has no formal legal status”.
In the US, the rule is 21 CFR 101.9(g). Compliance is checked on a composite of 12 units taken from 12 different shipping cases of one lot, so the rule applies to an average, not to each package. Naturally occurring protein, carbohydrate, fiber, vitamins and minerals must reach at least 80 percent of the labeled value. A food is misbranded if its calories, sugars, total, saturated or trans fat, cholesterol or sodium are “greater than 20 percent in excess” of the label. Reasonable shortfalls in those, and reasonable excesses of nutrients such as protein and fiber, are accepted within good manufacturing practice.
What happens when researchers test labels and menus?
In lab tests, packaged snacks came close to their labels, while restaurant food was accurate on average but varied a lot from item to item.
- Packaged snacks. A 2013 study of 24 popular US snack foods found metabolisable calories a median 4.3% above the label, which the authors judged “well within FDA limits”. Servings slightly heavier than stated and extra carbohydrate explained most of the gap.
- Reduced-calorie meals. A 2010 study found 29 reduced-energy restaurant foods averaged 18% more energy than stated and 10 supermarket frozen meals 8% more, with some single restaurant items up to twice the stated value. Because results varied widely between items, the differences were not statistically significant.
- Restaurant menus. Across 269 items from 42 US restaurants in a 2011 study, stated and measured energy matched on average (a difference of about 10 kcal per portion), but 19% of items held at least 100 kcal per portion more than stated. Foods listed as lower in calories tended to contain more than stated.
The pattern is that averages hold up reasonably well, while any single item can be noticeably off.
How much do people misjudge what they eat?
This error is separate from labels and databases, and it tends to run in one direction. The NIDDK, part of the US National Institutes of Health, notes on its Body Weight Planner that “people often underestimate how much they eat”.
The best-known study is small and specific. In 1992, researchers measured intake and energy expenditure in ten people with obesity (nine women and one man) who reported being unable to lose weight on a low-calorie diet. They under-reported what they ate by 47% on average and over-reported their activity by 51%, while their measured energy expenditure was close to predicted values. That is a selected group and should not be read as a figure for everyone, but the direction of the error matches the NIH’s general observation.
Are AI photo estimates accurate?
They are useful as a starting point and still imprecise. A 2026 study compared one AI meal-recognition app with weighed food records in 36 university students in Japan (35 of them women) over ten days. The two methods correlated, but agreed only moderately: the app systematically overestimated energy and the main macronutrients, underestimated fiber, and the gap between methods varied widely for several nutrients. The authors suggest such an app may be useful for following overall trends over time, with caution where precise individual numbers matter. It covered one app and one population, so its numbers will not transfer to other tools. A 2024 scoping review of 84 articles, covering studies from 2008 to 2021, describes a field that has improved quickly and still has clear weaknesses.
A photo cannot show oil that has soaked into food or a layer of cheese hidden under something else, and it has to judge the weight of a portion from how it looks, which is why confirming or correcting an estimate matters.
What is a barcode, and where do the numbers come from?
A barcode on food usually encodes a GTIN, a Global Trade Item Number. According to GS1, the standards body behind it, a GTIN identifies a trade item at any packaging level, comes in four formats (GTIN-8, GTIN-12, GTIN-13 and GTIN-14) and ends in a check digit. The barcode carries that identifying number, not the nutrition facts. An app scans the number and looks it up in a database, so the result is only as good as the entry behind it.
How reliable are food databases?
Two kinds of database are common, and they make different trade-offs.
| Open Food Facts | USDA FoodData Central | |
|---|---|---|
| What it is | “A collaborative, free and open database of food products from around the world” | A US government database run by the Agricultural Research Service |
| Strength | Wide coverage of packaged products from many countries | Separates lab-analyzed foods (Foundation Foods) from label data (Branded Foods) |
| Weak point | Entries come from its community, some are incomplete, and the project gives no accuracy guarantee | Focused on US foods; Branded entries are label data rather than lab analyses; SR Legacy was last updated in 2018 |
We did not find a published audit of the accuracy of Open Food Facts nutrition values. A 2020 validation study of a crowd-sourced calorie app (50 adults in Belgium) is a useful comparison: after the researchers removed implausibly high values (2.8% of entries), energy totals correlated strongly with a Belgian reference database (r = 0.96), while cholesterol and sodium agreed only weakly.
Does counting still help if the numbers are imperfect?
The research suggests it can, with honest limits.
- A 2011 systematic review of 22 studies found a consistent link between self-monitoring (food, activity and weighing) and weight loss, but the authors rated the evidence weak, mainly because samples were mostly white women and the data were self-reported.
- A 2021 meta-analysis of 12 randomized trials in adults with overweight or obesity found that digital self-monitoring of diet and activity was linked with about 2.9 kg more weight loss on average than comparison groups, with large differences between trials; tailored advice appeared to matter.
- A 2024 Cochrane review of 18 trials is more cautious: compared with no or minimal intervention, smartphone apps may make little to no difference to body weight at 12 months (low certainty) and probably make little to no difference at 24 months (moderate certainty, from a single trial).
- In a 24-week online program with 142 participants (91% women), time spent logging fell from 23.2 to 14.6 minutes a day between the first and sixth month. How often people logged, not how many minutes they spent, was linked with weight loss. This is an association within one program.
Because the same habits tend to produce the same biases, a consistent log is better at showing direction and change over weeks than at giving a precise figure for any single day. The authors of the 2026 photo study draw the same distinction for their app.
Counting is not neutral for everyone. In a survey of 493 college students, users of calorie trackers reported more eating concerns and dietary restraint. The study shows an association, not cause and effect, and the authors call it preliminary. If counting starts to feel stressful or controlling, stepping back from the numbers and talking to a doctor or dietitian is a reasonable choice.
What tends to reduce the error?
These are practical habits rather than tested interventions:
- Logging the same way each day. A consistent method makes week-to-week comparisons fairer, even if every day carries some error.
- Weighing a food now and then. Occasionally checking a typical portion on a kitchen scale shows how far eyeballed portions drift.
- Checking entries that look unusual. A yogurt with 900 kcal, or a product with no fat listed, is more likely a database error than a surprise.
- Looking at weekly averages. Single days are noisy; trends over weeks are more informative.
- Treating restaurant and photo estimates as approximate. They are reasonable starting points, not exact counts.
How IQFood handles this
Scanning a barcode in IQFood looks the product up among about 4.6 million products from Open Food Facts. Some of those entries are incomplete, so a scanned value is worth a quick check. Nutrition values for generic, unpackaged foods come from USDA FoodData Central. With AI logging, part of Premium, the app suggests foods, portions and macros from a photo, your voice or a sentence, and you confirm them before anything is logged. The calorie totals it shows are estimates built on these sources and on your confirmations, with all the limits described above.