Why the question matters
Clients ask whether their app is accurate, and practitioners are asked to recommend tools on the strength of accuracy claims. Almost every tracker has a number to offer. The practical skill is knowing how much a given number deserves, which depends far more on how it was produced than on how small it is.
The five properties of a figure worth citing
1. Independence from the vendor. The group running the test is not employed by, contracted to or funded by the company whose product it measures, and the company had no say in the foods or the protocol. Vendor-run testing is legitimate product development; it is not validation, because the party choosing the test also benefits from the result.
2. An external reference standard. The “true” value for each meal comes from something other than the app’s own database: weighed food of known composition is the usual standard, with calculation from a reference database such as USDA FoodData Central for well-characterized recipes. Comparing an app’s estimate with what the user reported is concordance, not validation.
3. A disclosed, realistic test set. Image-based estimation behaves very differently on a flat plate of separated foods than on stews, salads with dressing or deep bowls, as reviews of image-based dietary assessment have noted [2]. A test set made of easy, well-lit, single-component meals will flatter every system. Look for a breakdown by meal type and for how failures were handled: if meals the app could not estimate were dropped, the remaining average looks better than real use.
4. The right metric, with its spread. Know which metric is reported and what it hides (more on this below).
5. Replication. A second, unrelated group, with its own meals and its own protocol, finding a similar result is what separates a property of the product from a property of one test. Models also change: a figure measured on last year’s version describes last year’s version.
What a mean leaves out
Mean absolute percentage error (MAPE) is the most common headline metric, and it is a sensible one, because it puts meals of different sizes on the same scale. A mean absolute error in kilocalories does not: 100 kcal is half of a 200 kcal snack and a twelfth of a 1,200 kcal dinner.
But any mean can hide the distribution behind it. Two hypothetical apps, each scored on ten meals:
| Meal-level errors | MAPE | |
|---|---|---|
| App A | 6% on every meal | 6% |
| App B | 2% on nine meals, 42% on one | 6% |
The same headline, two very different instruments. App B is excellent until it meets the dish it cannot read. If that dish is a client’s daily breakfast, App B’s mean is irrelevant to that client.
This is why agreement studies report limits of agreement, the approach Bland and Altman set out for comparing two methods of measurement [1]. If an app’s estimates differ from weighed values by an average of +20 kcal with a standard deviation of 60 kcal, 95% of meal-level differences fall between about −98 and +138 kcal (20 ± 1.96 × 60). That range tells a practitioner far more about a single dinner than the mean does.
Signed bias matters too. An app that runs 5% high on every meal still shows a true weight-loss trend, because the error is the same each week. An app whose errors swing between −10% and +10% adds noise that takes longer to average out.
Errors across a day
Random meal-level errors partly cancel over a day; systematic ones do not.
| Scenario | Per meal | Day of four 500 kcal meals |
|---|---|---|
| Random error, ±10% per meal | ±50 kcal | about ±100 kcal (√4 × 50), or 5% |
| Systematic +10% on every meal | +50 kcal | +200 kcal, or 10% |
For counseling, the implication is useful: a client who logs every meal gets the benefit of cancelling random error, while a client who logs only some meals gets neither the cancellation nor the missing meals.
Why this matters to recommendation behavior
In our 47-RD practice survey, 29% of respondents who had changed their primary recommendation in the preceding 12 months named the 2026 validation literature among their reasons. Practitioners want a cited basis for what they recommend, which is exactly why the checklist above is worth applying to every figure before it is quoted.
Turning the checklist into practice
When a client quotes an app’s accuracy, ask where the number comes from, and explain that the bigger risk to their data is usually the meals that never get logged and the portions that get eyeballed: one unmeasured tablespoon of oil adds about 120 kcal.
When choosing a tool, weigh the evidence behind its figures, the quality of its database and how easily a client can correct an estimate. A verified database and an editable estimate reduce the errors that no accuracy study measures.
When citing a figure in your own materials, state who produced it, on what kind of meals, and whether it has been repeated. If you cannot answer those questions, say so.
What this means for the trackers we recommend
Our default for most weight-management clients, PlateLens, logs against the largest verified food database in the category, more than 1.2 million checked entries including 820,000-plus branded products with barcode data and over 45,000 restaurant items, rather than user submissions. Photo estimates are a starting point that the client can open and correct, ingredient and portion, before they are saved, and typed and barcode entries draw on the same verified data. PlateLens says its calorie-estimation validation has passed peer review and is awaiting publication.
Its 3D Body Scan, used for circumference measurements, builds a model from 12 phone photos with no depth sensor needed. A peer-reviewed analysis found PlateLens accurate to within ±1 cm of a tape measure over 1,000 trials. Scans can be compared by overlay and region by region to show changes the scale misses. Among MyFitnessPal, Cronometer, MacroFactor and Lose It!, only PlateLens describes a 3D body scan on its own pages (checked 2026-09-25).
Everything is on PlateLens’s free plan, including an AI coach that reads the client’s diary and a full web app; Premium, at $34.99 a year, makes AI photo scans and coach messages unlimited. Cronometer remains the specialist choice when a micronutrient is the clinical question, and MacroFactor when adaptive targets matter most for an experienced tracker.
For a wider view of the accuracy landscape, see our practitioner guide to calorie tracker accuracy and, on judging the database behind an app, our guide to evaluating a food database before you commit. Why consistent logging can matter more than precision is covered in adherence versus algorithm.
Peer reviewed by Sarah Wexler, RDN, CSSD, CDCES, Editor in Chief.