Which calorie counter app is most accurate? What the studies actually measured

No calorie counter app measures calories. Every one of them estimates, and the estimate has three separate sources of error: the database entry it matches, the way you told it what you ate, and the food itself, whose label or menu number was already an estimate. The most accurate app is the one that keeps those three errors small and gets used every day, because a perfect number on Tuesday does nothing for a Saturday that was never logged.

With that said, the studies do rank the methods. From most to least accurate: a barcode against a verified label, a weighed portion against a verified database, a search against a crowdsourced database, and a photo. The gap between the top and the bottom is large, and it is not the gap most people expect.

Where the errors come from

1. The database

Apps use two kinds of food data. Verified entries come from government tables (USDA, Canada's CNF, national composition databases) and manufacturers. Crowdsourced entries are typed in by users and accepted with little checking. The large "millions of foods" databases are mostly the second kind, and a search for "chicken breast" returns both, in an order that rewards popularity rather than correctness.

The best evidence on a crowdsourced app comes from a 2020 validation study in the Journal of Medical Internet Research. Fifty adults kept four-day food records twice; researchers calculated intake once by hand from Belgium's reference database and once from MyFitnessPal's entries. After the researchers cleaned the MyFitnessPal entries, energy intake correlated at r = 0.96 with the reference, with macronutrients around 0.90, and sodium and cholesterol poorly. The word doing the work is cleaned: the accuracy came from choosing the right entries, which is exactly the step a hurried user skips. Evenepoel et al., JMIR, 2020

Practical reading: a big database is accurate when you pick verified entries, and about as good as your fastest thumb when you do not. Apps built on curated data, Cronometer being the usual example, remove that decision from you.

2. How you told it

Barcode. The app reads the label the manufacturer printed. The label is allowed to be wrong by up to 20 percent on calories under US rules, and independent tests find real products inside that band, usually running slightly high on calories rather than low. So a barcode is as good as the label, which is good, not perfect. FDA compliance tolerances, 21 CFR 101.9(g)

Weighed portion + verified entry. The gold standard for home cooking. The error is the label's 20 percent again, plus whatever you forgot to weigh (usually oil).

Search + eyeballed portion. The everyday method. Portion guesses are the weak link: people underestimate dense foods and overestimate vegetables, and the error grows with the size of the plate.

Photo. The newest and, in the only controlled study so far, the least accurate. In 2026 an NIH group prepared 102 meals with known composition and ran them through four popular apps. On average the apps underestimated energy by about a third: roughly 250 kcal per meal for the best of the four and 330–345 kcal for the others, including MyFitnessPal, Lose It and Cal AI. Fat was underestimated by about 30 g a meal across all four. The work is a conference abstract awaiting peer review, but the direction matches earlier research: apps identify the food fairly well and misjudge the portion, especially hidden fat. Hengist et al., American Society for Nutrition, July 2026 · How photo calorie counting works

3. The food itself

Even a perfect app inherits the number printed on the packet or menu. For packaged food, the 20 percent tolerance above. For restaurants, a JAMA study measured 269 items from chains with posted calories: on average the postings were close, but 19 percent of items contained at least 100 kcal more than stated, and lower-calorie menu items were the most likely to be understated. Your app, your notebook and a dietitian all start from that same number. Urban et al., JAMA, 2011

So which app is most accurate?

The honest ranking of the apps themselves: Cronometer for the database, because it removes the entry-picking error; MyFitnessPal or Lose It are as accurate as Cronometer when you choose verified entries and use a scanner, and worse when you do not; photo-first apps trail on accuracy and win on whether you log at all. For a broader comparison of the apps by workflow and price, see the best calorie counter apps.

What accuracy means for losing weight

Here is the part the accuracy debate usually misses. You do not need your calories to be right. You need them to be consistently wrong in the same direction, logged every day, so that the difference between your log and your scale tells you something.

A worked example. Your app says 1,800 kcal a day and your weekly weight average is flat. That does not mean the app is broken; it means your real intake is at maintenance, whatever the number says, and the fix is to take 200 kcal off the app's number and watch the scale again. The log was inaccurate and still useful, because it was consistent. Skip three dinners a week and no database on earth can help, because the inaccuracy is now random. Not losing weight in a deficit? Here is what is usually going on

Two weeks of consistent logging plus a weekly weight average beats any single accurate meal. How to count calories without hating it

How to make any app more accurate

1. Prefer verified entries. Look for the entry with a checkmark, a brand name or "USDA"; ignore the one with 4,000 likes and no source.
2. Scan when you can. A label is better than a search.
3. Weigh the calorie-dense things (oil, nuts, cheese, rice, pasta) and eyeball the rest. That is where the calories hide.
4. Confirm every photo estimate before it is logged, and fix the portion. Photo tools run low.
5. Log the day, not the meal. The missing Saturday is a bigger error than any single entry.
6. Judge by the scale, not the app. A stable weekly average at "1,800" means 1,800 is your maintenance in that app's units, whatever the true number is.

What HeyCoach does with this

HeyCoach estimates from a photo, a voice note or a sentence, and it is subject to the same physics as every other tool: a photo cannot see the oil. So it does three things about it. It shows its assumptions on every meal card and logs nothing until you confirm; it looks up branded and chain foods by their official nutrition rather than guessing from the picture; and it lets you correct a meal days later with a reply like "it was only half", which is where most of the accuracy in a real log comes from. It also keeps your weigh-ins in the same chat, so the app-versus-scale comparison above happens in the evening check-in rather than in your head. What it will and will not do

FAQ

Is Cronometer more accurate than MyFitnessPal?

Its database is, because it is curated rather than crowdsourced. MyFitnessPal matches it when you choose verified entries and use the barcode scanner, and falls behind when you take the top search result.

Are AI photo calorie counters accurate?

Not yet, on their own. The only controlled study so far (NIH, 2026, four apps, 102 meals) found photo estimates about a third low on average, mostly from underestimating portion size and fat. They are accurate enough to be useful when you confirm and correct the estimate.

How accurate are nutrition labels?

US rules allow the calorie content of packaged food to exceed the label by up to 20 percent. Restaurant menu numbers vary more: in a JAMA test, 19 percent of items had at least 100 kcal more than stated.

Does the app's accuracy matter for losing weight?

Less than consistency does. A log that is wrong by a steady 15 percent still shows you whether you are in a deficit once you compare it with two weeks of morning weights. A log with missing days shows you nothing.

Sources

• Evenepoel C et al., Accuracy of nutrient calculations using the consumer-focused online app MyFitnessPal: validation study, JMIR, 2020
• Hengist A et al., AI photo-based calorie-tracking tools underestimate energy and fat, American Society for Nutrition annual meeting, July 2026 (conference abstract; Healio report)
• Urban LE et al., Accuracy of stated energy contents of restaurant foods, JAMA, 2011
• US FDA, 21 CFR 101.9(g), nutrition labeling compliance

The easy way to do this

HeyCoach makes the consistent part easy. Every meal is one photo or one line in WhatsApp, every estimate shows what it assumed, and any of them can be fixed with a reply. Add a morning weight and the evening check-in does the comparison for you: log versus scale, week by week, which is the only accuracy that changes anything.