When estimates go wrong

TrakMac's estimator targets ±20% accuracy on common foods (see how accurate AI calorie tracking apps are). For most meals, that's the actual outcome. But specific items occasionally come back with estimates that are clearly off — sometimes by 50% or more.

Three common patterns:

1. The estimate is wildly low. You logged 'chicken stir fry' and got 350 calories when you know the dish was closer to 700. Often happens with home-cooked dishes that include significant cooking oils or cream-based sauces.

2. The estimate is wildly high. You logged 'protein bowl' and got 1,200 calories when the chain's published nutrition says 600. Sometimes happens when the description triggers an over-estimation of portion size.

3. The estimate is consistently off in the same direction for a regular meal. Every time you log your morning smoothie, the protein number is 8-10g lower than it should be.

Each of these has a clear fix.

How to correct an estimate

The immediate correction:

  1. Tap the logged meal on the dashboard.
  2. Edit any of the four macros directly. Calories, protein, carbs, and fat are independently editable.
  3. Save.

The daily totals update immediately. The streak math uses the corrected numbers.

For more on the editing flow, see how to edit a logged meal.

What happens after you correct

This is the part that matters for future estimates.

When you edit an estimate, two things happen:

Per-user calibration. TrakMac records your correction against the food description. The next time you (specifically) log the same or similar item, the estimate reflects your historical edits. Your typical portion sizes, your usual modifications, your typical preparation method.

This kicks in after 2-3 corrections of the same item. By the third or fourth time you log 'my morning smoothie,' the estimate is usually within ±10% of your actual macros without further editing.

Global feedback loop. Your correction also feeds into the broader estimator. When many users correct the same item the same way (e.g., 50 users all say 'Chipotle steak bowl' is consistently estimated 80 calories low), the underlying model adjusts.

This means your edit isn't just about your today. It's about everyone's tomorrow.

When to edit vs when to leave it

Not every imperfect estimate is worth editing. A few rules:

Edit when:

  • The estimate is more than ±15% off from your honest read
  • The food is something you eat regularly (the calibration benefit compounds)
  • You have actual nutrition data (chain restaurant, package label, recipe with macros)

Don't edit when:

  • The estimate is within ±10-15% of where you think it is — that's inside the noise of any tracking system
  • You don't actually know what the right number is — guessing replaces a model estimate with your guess, which may be worse
  • The food is one you'll never log again — the calibration benefit doesn't compound

The goal isn't to make every estimate perfect. It's to keep your tracking inside the ±20% accuracy band that produces real body composition outcomes.

Specific patterns and what to do

Restaurant chain meal estimate is off

TrakMac maintains a cache of menu items for the major fitness-friendly chains (currently 9 brands, 2,600+ items). When the cache has the item, the estimate uses the chain's published nutrition data. When it doesn't, the model falls back to general estimation.

If you're logging a chain meal and the estimate is far off:

  1. Check the chain's website for the actual nutrition (most publish per-item PDFs).
  2. Edit the macros to match.
  3. Note the exact item name in the description so the cache lookup matches next time.

For chains the cache doesn't cover yet, your edit teaches both the per-user system and feeds into chain-cache expansion priorities.

Home-cooked recipe estimate is off

Home recipes vary enormously. Your bolognese is not the same as everyone else's bolognese. Your protein smoothie has specific ingredients in specific quantities. The estimator can only guess at typical preparations.

The cleanest fix:

  1. Calculate the actual recipe macros once. Add up the ingredient macros via package labels.
  2. Edit the logged meal to the calculated macros.
  3. Use a consistent name for the dish (e.g., 'my morning smoothie' or 'my chicken bolognese') so the per-user calibration locks onto it.
  4. After 2-3 logs with the same name, the estimate will usually pull from your historical edits.

Voice transcription got the wrong words

If the speech-to-text mis-heard you ('chip butty' instead of 'chipotle'), the estimate runs on the wrong food entirely. Fix:

  1. Edit the description text to the right words.
  2. The system offers a 'Re-estimate' option for significantly changed descriptions. Tap it.
  3. The new estimate runs on the corrected description.

For more on voice recognition issues, see voice log not recognizing food.

A particular food consistently estimates wrong (across many users)

If you've logged a popular item 5+ times, edited each one, and the estimate still doesn't improve — and you suspect this is a global issue, not just your specific recipe — flag it via [email protected].

Include:

  • The exact food name as you logged it
  • The estimated macros TrakMac returned
  • What you believe the correct macros are
  • Source if you have one (chain nutrition page, package label, recipe with macros)

This goes into the model improvement queue. Common items that need adjustment get prioritized.

What you should NOT do

Don't "compensate" with a different meal. If TrakMac estimated your dinner 200 calories low, don't add 200 fake calories to lunch. Edit the actual meal.

Don't lie about portions to make the estimate look right. If you ate 8 oz of chicken, log 8 oz of chicken even if the estimate looks high. Inventing inputs corrupts your own data.

Don't abandon tracking when one estimate is wrong. A wrong estimate, edited, is functionally a correct entry. The system learns. The trend stays accurate.

Don't expect perfect accuracy on uncommon foods. Regional dishes, ethnic cuisines, niche recipes, weird brand names — these have higher variance because the model has less training data on them. The fix is editing, not expecting perfection.

The honest framing

The accuracy stack at TrakMac (USDA cross-reference + restaurant cache + per-user calibration + global feedback loop + Claude Haiku 4.5 estimation) targets 80% of estimates within ±20% of the user-corrected number. That target holds for the items the system has seen most.

For items it has seen less, accuracy is lower at first. Your edits are how that improves.

You are not pestering the system by editing. You are making the system better. Every correction you make is a small contribution to the model that benefits you on the next log and benefits every other user logging the same food next week.

The goal isn't perfect estimates on day one. It's a system that learns from real eating to produce accurate-enough tracking over time. Your edits are the mechanism. Use them.