This is not medical advice. General information for self-tracking only. If you are under 18, pregnant or breastfeeding, have ever had an eating disorder or a difficult relationship with food, or are managing a medical condition, talk to a doctor or a registered dietitian before you change what you eat. Reading about measurement error is not a reason to eat less to be safe. The full list is here.

The short answer, with the numbers

Nobody has measured how accurate a calorie tracking app is for you, and nobody can. What has been measured is every link in the chain between the food on your plate and the number on your screen, and each link has a published error attached to it.

  • What you tell it is low. Measured against doubly labelled water across five pooled validation studies, people under-reported their energy intake by about 15% on a single 24-hour recall and 28% on a food frequency questionnaire.1
  • The label on the packet has legal room in it. A US food may contain up to 20% more calories than it declares before the label counts as misbranded.2
  • An estimate from a photo is good at calories and bad at fat. Against weighed meals, the better AI models correlated above 0.8 for energy and often landed inside 10%, while every model tested overestimated fat by more than 20%.3
  • The target you compare it against is a prediction too. The equation almost every app uses lands within 10% of measured resting metabolic rate more often than its rivals, which is another way of saying it misses by more than 10% a good deal of the time.4

Stack those up and the honest position is uncomfortable for anyone selling a tracker, mine included. The biggest error in the chain belongs to the person operating it, and that was true long before phones.

No calorie tracker is accurate in absolute terms. What one can do is apply a roughly similar error to every day, so the difference between Tuesday and Wednesday means something even though neither number is true. That is a smaller claim than the one on the App Store, and it is the one the research supports.

Error one: what you tell the app

The reference method here is doubly labelled water. You drink a measured dose of water whose hydrogen and oxygen have been replaced with heavier isotopes, then the rate at which each one washes out of your urine over the following couple of weeks gives your total energy expenditure. It asks nothing of your memory and depends on nothing you say. In someone whose weight is stable, expenditure equals intake, which is what makes it the reference the self-report methods get judged against.

Held against it, self-reported intake comes up short almost universally. The largest pooled analysis, covering five US validation studies run between 1999 and 2009, found average under-reporting of 15% with a single 24-hour recall and 28% with a food frequency questionnaire. Body mass index was one of the strongest predictors of the gap, along with education and age.1

The OPEN study, which put 484 adults through the same comparison, found men under-reporting energy by 12% to 14% on recalls and women by 16% to 20%. One detail from it matters for anyone tracking macros rather than calories: there was little under-reporting of the percentage of energy coming from protein.6 The composition of what people report survives better than the total does, which is a quiet argument for paying more attention to your macro split than to your calorie sum.

The number everybody quotes comes from a 1992 New England Journal of Medicine paper. Ten people with a history of failing to lose weight on a self-reported diet of under 1,200 kcal were measured properly. Their metabolic rates were normal. They were under-reporting their intake by an average of 47%, and over-reporting their exercise by 51%.5 That figure gets waved around as though it applies to everyone, and it should not: the group was ten people, selected precisely because they were the hard cases. Treat 15% as the typical gap and 47% as the far end of it.

None of this is lying. It is the oil the pan was cooked in, a portion guessed low because guessing low is what people do, and the four crisps taken from somebody else's bag. No software fixes it, and a tracker that claims to has not read the literature.

Error two: the number printed on the packet

People assume the packaged food is the trustworthy part of the log. The regulation says otherwise, and it says it precisely. Under 21 CFR 101.9(g)(5), a food declaring calories "shall be deemed to be misbranded ... if the nutrient content of the composite is greater than 20 percent in excess of the value for that nutrient declared on the label".2 A bar labelled 200 kcal can contain 240 and remain compliant.

The common shorthand for this is that the FDA allows a tolerance of plus or minus 20%, and that is not what the rule says. The 20% is a ceiling in one direction only. Deficiencies below the labelled amount are handled by a separate paragraph, which calls reasonable ones "acceptable within current good manufacturing practice" and attaches no percentage at all.2 The asymmetry runs the unhelpful way for anyone counting: the legal room is on the side of there being more calories than you logged.

Restaurant food is worse, and it has been measured with a bomb calorimeter rather than estimated. Researchers bought the 42 most frequently ordered meals from randomly selected independent and small-chain restaurants around Boston, 157 meals in total, and burned them. Mean energy was 1,327 kcal per meal. In the subset they could match directly, the real meals contained 19% more energy than the national food database entries for the equivalent items.7

The most useful detail in that study is the variation rather than the average: the same dish from the same restaurant varied with an average standard deviation of 271 kcal.7 Even a database entry that is exactly right about the recipe is describing a dish that the kitchen builds differently on a Tuesday than on a Saturday.

Error three: the database entry behind the search result

Searching a food database feels like measurement in a way that photographing a plate does not, but you are still choosing somebody else's entry for a food that resembles yours.

An Italian team ran five of the largest tracking apps against a validated food-composition reference over two weeks of real eating. FatSecret, Lifesum, MyFitnessPal, Yazio and Melarossa all under-estimated total energy, and the authors concluded that leading nutrition apps "present critical issues in assessing the intake of energy and nutrients". The mechanisms they blamed were food items that did not match the country's composition tables, and user-created entries that nobody checks.8

A second study cuts the other way, and it is the one that stops me writing this section one-sidedly. When a different group took the 50 most frequently eaten foods from a weight-loss study, mostly unprocessed and minimally processed items, and compared four commercial apps against the Nutrition Data System for Research, agreement was good to excellent for CalorieKing, Lose It! and MyFitnessPal across energy, macros and most other nutrients. Fitbit was the outlier, dropping as low as an intraclass correlation of 0.16 for fibre in vegetables.9

Those two results are not in conflict. Plain foods are well covered and the databases handle them fine. The errors arrive with mixed dishes, restaurant meals, regional foods and the user-submitted entries that fill the gaps. The same study is a good illustration: MyFitnessPal's mean energy difference from the research database was only 8 kcal, but the spread around that mean had a standard deviation of 133 kcal.9 That makes it unbiased across a lot of foods and unreliable on the one you just ate.

Operator skill is part of the error too. In a study of 37 Filipino adults with obesity, participants logged five days in MyFitnessPal while three dietitians logged the same paper food records into the same app. The averages did not differ significantly, but the agreement between the two, item by item, was very weak to moderate.10 Two people entering the same food into the same app did not arrive at the same numbers. The authors were working with 37 people in one country whose cuisine a US-centred database covers thinly, so read it as a demonstration rather than a measurement of your own logging. It is still the most direct evidence I know of that the app is not the whole instrument.

Error four: estimating a meal from a photograph

This is what my own app does, so treat the next paragraph as a disclosure as much as a description.

The cleanest test to date photographed 15 standardised hospital meals at a fixed 90-degree angle under 500 lux, weighed every component for ground truth, and had 10 registered dietitians and 10 AI models estimate them. For energy and carbohydrate, the better models performed like the dietitians: correlations above 0.8, frequently inside 10%. Protein and fat were significantly worse across every model, and all of them overestimated fat with a mean bias above 20%. The authors call it the invisible nutrient problem.3

The reason is not mysterious. A photograph cannot see the oil the vegetables were tossed in, the butter under the fish, the sugar dissolved in the sauce, or the half of the portion sitting underneath the visible half. It also cannot see the plate size unless something in frame gives it away.

The limits of that study matter before anyone generalises from it: fifteen meals, hospital plating, controlled lighting and a fixed camera angle. Your kitchen at 8pm is none of those things, and there is no published equivalent for a phone held at a random angle over a takeaway container. So it describes a best case rather than a typical one.

For what it is worth, in MyPlate the number I would trust least on any given scan is fat, which is exactly what the research predicts. The app has no food database and no barcode scanner, so a packet with the numbers printed on the side still has to be described rather than scanned, and a photo with a short written description of what is in the dish gives it more to work with than the photo alone.

Error five: the target you are counting toward

Every error above concerns the left-hand side of the comparison. The right-hand side, the daily target, is an estimate too, and it is the one people forget is an estimate at all.

Almost every app, this one included, starts from the Mifflin-St Jeor equation. A systematic review of the four equations in common clinical use found it the most reliable of them, predicting resting metabolic rate within 10% of the measured value in more people than any of the alternatives and with the narrowest error range. The same review flagged noteworthy errors when it is applied to an individual, and noted that older adults and US-resident ethnic minorities were under-represented both in the studies that produced these equations and in the studies that validated them.4

Ten percent of a 1,500 kcal resting rate is 150 kcal, and that error is then multiplied by an activity factor between 1.375 and 1.9 before you ever see it. MyPlate's arithmetic is public for exactly this reason: 10 times your weight in kilograms, plus 6.25 times your height in centimetres, minus 5 times your age, plus 5 if you are male and minus 161 if you are female, scaled by activity and then by 0.85 to lose or 1.1 to gain, and never allowed below 1,200 kcal for women or 1,500 for men.11 You can run the whole thing yourself in the calorie calculator and check it against the citations on the sources page.

One limitation of mine while I am listing everyone else's: logging a weight in MyPlate does not recompute your calorie target. Your expenditure falls as you get lighter, and the app will keep showing you the target it worked out when you set the goal until you re-run the setup. That is on my list and it is not fixed yet.

So which is the most accurate calorie tracker?

The question has no winner, and it is worth knowing why. No consumer calorie tracker has been validated against doubly labelled water in the way the question implies. What the studies compare is an app against a reference database, using food that somebody entered carefully, which measures the database and not the app-plus-human system you are actually asking about.

Sorting trackers into accurate and inaccurate bins by category does not hold up either. Database apps under-estimated energy in one study8 and agreed well with a research database in another.9 Photo estimation matched registered dietitians on energy and failed badly on fat.3 Neither method is measuring anything; both are looking things up or guessing, with different failure modes.

Two things move accuracy more than the choice of app. The first is whether you log the item at all, since an unlogged snack is a 100% error that no database quality compensates for. The second is whether the food is simple and packaged or a mixed dish someone else cooked, because that distinction is bigger than the gap between any two trackers on this market.

If you are shopping for a calorie and nutrient tracker rather than a calorie counter, ask the two questions separately, because the error bars are not the same per nutrient. Energy and carbohydrate are the well-estimated ones. Fat is the worst case in photo estimation.3 Fibre showed the widest variation between databases.9 An app can be perfectly adequate at calories and poor at the nutrient you actually opened it for, and for micronutrients specifically a photo cannot help you at all.

Why consistency beats accuracy

If your log is wrong by roughly the same amount every day, the difference between days still carries information even though the level does not. A systematic bias mostly cancels when you compare Tuesday with Wednesday, and what is left is the random part.

That argument is real but it needs a caveat that gets left off. The bias is not constant. It grows with meal size, and it grows with restaurant food, which is why a week of eating out is under-counted more than a week of cooking. So the log is a decent instrument for "more than yesterday" and a poor one for "1,847 exactly".

What the practical evidence supports is frequency rather than thoroughness. Among 142 people in a 24-week online weight-control programme, those still logging at month six spent no more time doing it than the people losing less weight, but they opened their food journal significantly more often: 2.4 times a day against 1.6 for those who lost under 5% of their body weight.12 This is an observational finding inside a trial, in a sample that was 91% female, so it shows an association rather than proving that more logging causes more loss.

The measurement that closes the loop happens on the bathroom scale: your weight, averaged over two to four weeks. If the scale is not moving on 1,800 logged calories, the useful conclusion is not that the app lied. It is that 1,800 is your reading on an imperfect instrument, and the number to adjust is the one you are aiming at, not the one you recorded.

One thing matters more than any of this. If tracking is making you anxious, if you find yourself eating less to make a number look right, or if logging has stopped being something you chose, put it down and talk to a doctor or a registered dietitian. Getting the number right is not worth that.

Common questions

How accurate are calorie tracking apps?

Not accurate in absolute terms, and the largest error is not in the software. Measured against doubly labelled water, people under-report their own intake by about 15% on a single 24-hour recall and 28% on a food frequency questionnaire.1 On top of that, a US food label may legally understate calories by up to 20%,2 restaurant meals have been measured at 19% above the database values for equivalent dishes,7 and the daily target the app compares your log against comes from an equation that lands within 10% of measured resting metabolic rate at best.4 A realistic expectation is that your daily figure is in the right neighbourhood and the trend across weeks is what carries the information.

Which calorie tracker is the most accurate?

There is no validated answer, because no consumer app has been tested against doubly labelled water in the way the question assumes. The studies that exist compare an app against a research food database using carefully entered food, which measures the database rather than the app-and-user system you are asking about. Results also cut both ways: five leading database apps under-estimated energy against a validated reference,8 while a separate comparison found good to excellent agreement for CalorieKing, Lose It! and MyFitnessPal on commonly eaten unprocessed foods.9 Choose on whether you will keep using it, because logging frequency is the variable with evidence behind it.12

Are AI photo calorie counters less accurate than food database apps?

Not for calories. Against 15 weighed hospital meals, the better AI models estimated energy and carbohydrate about as well as registered dietitians, correlating above 0.8 and often landing within 10%. They failed on fat, which every model overestimated by more than 20%, because oil and butter are invisible in a photograph.3 Database apps have a different weakness: the entry you pick was written by somebody else for a dish resembling yours, and user-created entries that nobody checks were the mechanism researchers blamed for the errors they measured in five popular apps.8 Both are estimating rather than measuring.

Does the FDA allow food labels to be 20% off?

In one direction only, and the common "plus or minus 20%" shorthand is wrong. Under 21 CFR 101.9(g)(5), a food declaring calories is misbranded if the measured content is greater than 20 percent in excess of the declared value, so a bar labelled 200 kcal can legally contain 240. Amounts below the label are covered by a different paragraph, which calls reasonable deficiencies acceptable within current good manufacturing practice and sets no percentage.2 The legal room sits on the side of there being more calories than you logged.

If the numbers are wrong, is calorie tracking pointless?

No, because you are not using the number the way a laboratory would. An error that repeats at a similar size each day mostly cancels when you compare one day with another, which leaves the log useful for direction even when it is unreliable for level. The evidence for what makes tracking work points at frequency rather than precision: in a 24-week programme, people who lost 5% or more of their body weight opened their food journal 2.4 times a day against 1.6 for those who lost less, while spending no more total time on it.12 Weigh yourself regularly and treat the weight trend over two to four weeks as the real measurement.

Is a calorie and nutrient tracker accurate for micronutrients too?

Treat those as two separate questions, because the error differs by nutrient. Energy and carbohydrate are the best estimated, fat is the worst case in photo-based estimation,3 and fibre showed the widest variation between commercial databases and a research reference.9 For vitamins and minerals specifically, a photo cannot help at all and you need an app drawing on a curated database such as USDA FoodData Central. If you have a clinical reason to watch a particular micronutrient, that is a conversation for a dietitian rather than a feature comparison.

The bottom line

App Store copy promises accuracy. What the research supports is consistency, which is smaller and more useful. Your log is an instrument with a bias you cannot see or remove, reading a quantity that varies by hundreds of calories between two servings of the same restaurant dish.7

So log the same way every day, including the days you would rather not, since a skipped snack is a bigger error than any database can introduce. If your app takes a photo and a description, give it both. And treat the number as an input to a decision rather than a fact about your body: if two weeks of an 1,800 kcal log has not moved your weight, adjust the target and carry on, instead of arguing with the app.

Where a target comes with a floor, respect it. MyPlate will not show a daily target below 1,200 kcal for women or 1,500 for men, a limit taken from the NIH clinical guidelines on obesity.11 An imprecise instrument is a reason to leave yourself more margin.

I would rather tell you all of this than let you find it out in month three. Every figure above is linked below, and if one of them does not say what I claim it says, I want to know.

Sources

Every figure above traces to one of these. If you find a number that doesn't match the source it claims, tell me and I'll correct it.

  1. Freedman LS, Commins JM, Moler JE, et al. (2014). Pooled results from 5 validation studies of dietary self-report instruments using recovery biomarkers for energy and protein intake. American Journal of Epidemiology, 180(2):172-188. Read on PubMed →
  2. US Food and Drug Administration (2026). Nutrition labeling of food, 21 CFR 101.9(g). Code of Federal Regulations, Title 21. Read the regulation →
  3. Isobe T, Zhang LW, Murakami H, et al. (2026). Accuracy of AI-Based Nutrient Estimation from Standardized Hospital Meal Images: A Comparison with Registered Dietitians. Nutrients, 18(6):966. Read on PubMed →
  4. Frankenfield D, Roth-Yousey L, Compher C (2005). Comparison of predictive equations for resting metabolic rate in healthy nonobese and obese adults: a systematic review. Journal of the American Dietetic Association, 105(5):775-789. Read on PubMed →
  5. Lichtman SW, Pisarska K, Berman ER, et al. (1992). Discrepancy between self-reported and actual caloric intake and exercise in obese subjects. New England Journal of Medicine, 327(27):1893-1898. Read on PubMed →
  6. Subar AF, Kipnis V, Troiano RP, et al. (2003). Using intake biomarkers to evaluate the extent of dietary misreporting in a large sample of adults: the OPEN study. American Journal of Epidemiology, 158(1):1-13. Read on PubMed →
  7. Urban LE, Lichtenstein AH, Gary CE, et al. (2013). The energy content of restaurant foods without stated calorie information. JAMA Internal Medicine, 173(14):1292-1299. Read on PubMed →
  8. Tosi M, Radice D, Carioni G, et al. (2021). Accuracy of applications to monitor food intake: Evaluation by comparison with 3-d food diary. Nutrition, 84:111018. Read on PubMed →
  9. Lin AW, Morgan N, Ward D, et al. (2022). Comparative Validity of Mostly Unprocessed and Minimally Processed Food Items Differs Among Popular Commercial Nutrition Apps Compared with a Research Food Database. Journal of the Academy of Nutrition and Dietetics, 122(4):825-832. Read on PubMed →
  10. Banal MG, Bongga D, Angbengco JM, et al. (2024). MyFitnessPal smartphone application: relative validity and intercoder reliability among dietitians in assessing energy and macronutrient intakes of selected Filipino adults with obesity. BMJ Nutrition, Prevention & Health, 7(1):54-60. Read on PubMed →
  11. National Heart, Lung, and Blood Institute, NIH (1998). Clinical Guidelines on the Identification, Evaluation, and Treatment of Overweight and Obesity in Adults. NIH Publication 98-4083. Read the guidelines →
  12. Harvey J, Krukowski R, Priest J, West D (2019). Log Often, Lose More: Electronic Dietary Self-Monitoring for Weight Loss. Obesity, 27(3):380-384. Read on PubMed →

The formulas and limits behind MyPlate's own numbers, with their citations, are on the health sources page.