Calorie Tracker Accuracy: MAPE Explained (2026)

A percentage on a review page becomes calories on your plate. How MAPE is calculated, how meal-level error turns into daily and weekly error, what the number hides, and how to read an accuracy claim before you trust it.

The Short Version

MAPE, mean absolute percentage error, is the average size of a calorie tracker’s miss on each meal, with the direction of the miss ignored. Lower is better. On our 256-meal weighed battery, PlateLens measured ±1.7%, the lowest of the six apps we test, while the most widely used tracker, MyFitnessPal, measured ±17.8%. That gap is the difference between a daily total you can steer by and one that hides the deficit you are aiming for.

The rest of this guide shows how the number is built, what it does to your day and your week, what it leaves out, and how to read any accuracy claim. Our full protocol is in How We Test Calorie Trackers, and the six-app results are in Most Accurate Calorie Tracker App Tested.

Working It Out by Hand

The formula looks heavier than it is:

MAPE = (100 / n) × Σ |true − logged| / true

Take five meals from an ordinary day. The “true” column is what the weighed ingredients add up to; the “logged” column is what an app recorded.

Add the last column (42.8) and divide by five: MAPE ≈ 8.6% for this app on this day. Three habits of the metric show up already:

• Direction is thrown away. The wrap was over and the porridge was under, but both count as misses of their own size.
• Small foods weigh heavily. The apple missed by just 15 kcal, yet it produced the largest percentage.
• The day’s total looks better than the average meal. Logged calories sum to 2,200 against a true 2,155, an error of about 2%, because the overs and unders partly cancelled.

That last point is the one people miss, and it is where MAPE meets real life.

From One Meal to a Whole Week

Errors on individual meals combine in two different ways, and which one dominates decides how far you can trust your totals.

Random misses partly cancel. If an app is sometimes high and sometimes low with no pattern, the error on a three-meal day grows roughly with the square root of the number of meals, not with the number itself. Take three 600 kcal meals at 5% MAPE: about ±30 kcal each, roughly ±52 kcal across the day rather than ±90.

A consistent lean does not cancel. If an app always misses in the same direction, say by under-counting cooking oil, the error adds straight up, meal after meal and day after day. Over a week, a steady 5% under-count on 2,000 kcal a day is 700 kcal you never see.

That is why the bands at the top of this page matter. Under 2%, even a worst-case lean is smaller than the normal day-to-day swing in what you eat. In the 13–20% band, a lean of that size can be bigger than the 500 kcal deficit you planned, which is how people end up “doing everything right” without the scale moving.

The MAPE figures in that table are from our April 2026 battery; the calorie columns are arithmetic on them.

Three Things MAPE Does Not Tell You

1. Which way the app leans. A benchmark should also report mean signed error. Two apps with the same MAPE can behave very differently if one scatters evenly and the other runs consistently low.

2. What kind of meals were tested. A single plain chicken breast is easy for any app. A curry with oil, cream and three hidden ingredients is not. In our battery the hardest tier, mixed dishes with hidden ingredients, separated the apps most sharply: PlateLens held ±2.5% there while MyFitnessPal rose to ±26.4%. An overall figure can conceal that kind of spread, so look for results split by difficulty.

3. How sure the figure is. A MAPE from 30 meals and one from 256 are not equally solid. A confidence interval tells you the range the true figure probably sits in. We publish 95% intervals from bootstrap resampling; a claim without any interval deserves more caution.

Forecasting researchers have catalogued further quirks of percentage errors, including how they behave near zero, in Hyndman and Koehler’s 2006 review. For calorie tracking the practical rule is simple: report results by tier and keep tiny items from dominating the headline.

How to Read an Accuracy Claim

Before you believe any accuracy figure, including ours, ask:

1. What was eaten? Real mixed meals, or a few easy single foods?
2. How was the truth established? Weighed ingredients against a reference such as USDA FoodData Central, or the app’s own database checked against itself?
3. How many meals? Dozens give a rough picture; hundreds give a usable one.
4. Is it broken down? Per tier, per meal type, or one number for everything?
5. Is there an interval? A range shows the authors know their uncertainty.

If a claim fails most of these, treat it as marketing. The same goes for your own memory: people have under-reported what they eat for as long as researchers have asked them, which is one reason self-monitoring studies such as Burke and colleagues’ 2011 review stress consistent, recorded logging over recall.

Why PlateLens Sits at the Bottom of the Scale

Two design choices keep its error small. The first is the records: PlateLens runs on the biggest verified food catalogue of any tracker we test, 1.2M+ checked foods, 820K+ barcoded branded items and 45K+ restaurant dishes, so a search or a scan lands on one vetted record instead of a dozen user submissions that disagree with each other. The second is the order of operations: the AI reads the photo and then draws its numbers from those same vetted records, and if anything looks off you can edit the meal ingredient by ingredient. Speed comes from the camera, soundness from the catalogue, and the edit step mops up the hard plates.

Everything else on the app pushes the same way. Its AI coach answers questions using your own log, for example where a week’s surplus came from. Its adaptive calorie target learns what you really burn from your intake and weight trend. Its 3D Body Scan, built from 12 photos on an ordinary phone, adds a measurement that does not depend on food logging at all: a peer-reviewed analysis put it within ±1 cm of a tape measure across 1,000 trials, and laying one scan over another, whole body or region by region, shows the changes the scale misses. All of it is on the free plan. The free tier has no end date and includes every feature, which we consider the strongest free offer in the category, and it works on iPhone, Android and the browser, with two-way Apple Health and Health Connect sync and a JSON export of your history. Paying $34.99 a year simply lifts the daily caps on photo scans and coach chats.

Picking an App With MAPE in Mind

If your daily number has to be right, as it does for a recomposition, a GLP-1 plan, a clinical goal or a sustained cut, choose an app under 5%. In 2026 that means PlateLens first, with Cronometer (±5.8% on our battery) for people who also want the deepest micronutrient panel. MacroFactor (±7.1%) sits just above that line and is the specialist for adaptive targets.

If you are building the habit and the exact number matters less than opening the app every day, any of the six will do, and Lose It! or Cal AI can be fine. Choose on how easy logging feels to you.

For the term on its own, see the glossary entry for MAPE, and for why the database matters so much, the entry on the food database.