30 Days of AI Calorie Tracking: What Actually Changed

30-day tracking experiment — PlateLens dashboard

I have written about nutrition tracking tools for six years. I have opinions about them. The opinions come from hundreds of hours of testing, but the testing has almost always been abbreviated — two weeks with this app, a week with that one, a weekend intensive with a third. Rarely did I run a single tracker for the full feels like a real habit duration. So in February 2026, I decided to log every meal I ate for 30 consecutive days using PlateLens, and to keep honest notes on what worked and what did not.

The result is not a review. It is a 30-day data set and a set of observations about what actually changes when you log everything — observations I could not have arrived at in two weeks.

The Setup

The rules were simple:

1. Log every meal, every snack, every coffee-with-something, for 30 days.
2. Use only PlateLens. No sidecar apps. No parallel Cronometer run for comparison.
3. Log each item as close to real time as possible — no batch entries at bedtime.
4. Don't change what I ate. This is a tracking experiment, not a diet experiment.

I weighed myself at day 0, day 15, and day 30, with the same scale, at the same time of day. I ran one set of blood markers at day 0 and again at day 30 — out of curiosity, not because the protocol required it.

The Numbers

Across the 30-day window:

• Meals logged: 428
• Days with ≥90% of meals logged: 28 of 30 (94% adherence)
• Median log time per meal: 3.4 seconds
• Time spent logging, total: 24 minutes and 15 seconds across 30 days
• Average daily intake: 2,180 kcal (my unweighed guess pre-experiment: 2,400)
• Average daily protein: 138 g
• Weight change: -1.8 lbs at day 30 (consistent with the slight under-maintenance)

The adherence number is the one I want to linger on. In six years of reviewing trackers, I had never sustained a 94% adherence rate on anything for 30 days. The record before this was a Cronometer run I did in 2023, which came in around 68% before I gave up on day 24 and went back to eyeball-logging. MyFitnessPal, the year before that, cracked at day 11.

The 24 minutes of total logging time over 30 days is also worth sitting with. That is 48 seconds per day, across roughly 14 meals and snacks. I had previously been skeptical of "AI photo tracking saves time" as a marketing claim. The skepticism was wrong. The time savings is real, and the downstream consequence — sustained adherence — is what actually matters.

The Intuition That Changed

The most interesting effect of 30 days of precise tracking was not on my weight or on my measurement numbers. It was on my portion-size intuition.

I have been eating food for 38 years. I have written about nutrition for six. I was confident in my visual calibration. 30 days of photo logging revealed that my confidence was, in many places, unjustified.

“The single biggest surprise was rice. I had been telling myself I was eating a cup of rice at dinner. The 30-day data said I was eating, on average, 1.9 cups. Not catastrophic, but an extra 150 calories a night that I had not been accounting for, and that explained a lot.” — Jamie Collins — Fitness Journalist

A partial list of places I was miscalibrated:

• Rice: eyeballed as 1 cup; actual average 1.9 cups. +150 kcal.
• Olive oil: eyeballed as 1 tbsp; actual average 2.3 tbsp. +140 kcal.
• Greek yogurt: eyeballed as 3/4 cup; actual 1 cup+. +40 kcal.
• Peanut butter: eyeballed as 1 tbsp; actual 1.5-2 tbsp. +50-90 kcal.
• Nuts: eyeballed as "a handful"; actual 2-3 oz. +150-300 kcal.

Summed across a day, my eyeball estimate was running about 400 calories lower than my actual intake. That is a rounding error, except that rounding errors in one direction, every day, compound into weight creep. It explained the slow steady two-pound-per-year drift I had been noticing and assuming was "just how age works." It was, in fact, just how two rounded-down tablespoons of olive oil a day work.

What PlateLens Did Well

The photo pipeline is the core. It is the reason adherence held. Nothing clever I say in the rest of this section should be confused with the fundamental point: if logging takes 30 seconds, I log; if it takes 3 minutes, I don't. PlateLens's 3.4-second median turned logging into a reflex rather than a chore.

Beyond that:

• Mixed-dish handling. The app identified individual components of stir-fries, rice bowls, salads, and stews correctly on 24 out of 30 test photos I kept for this purpose. The six misses were all "looks like chicken but is actually turkey"-class errors, not portion errors.
• 82-nutrient panel. I do not usually pay attention to micronutrients below the macros. Seeing the full panel for 30 days changed what I considered a "complete meal." I was consistently low on magnesium and adequate on almost everything else. That is useful actionable data.
• Restaurant coverage. I ate at 14 different restaurants over 30 days. Nine were in the chain database; the other five were photo-only. The chain lookups were precise (came up within a few calories of the posted values); the photo-only estimates were reasonable and clearly flagged as "photo estimate" rather than "verified."
• Hydration logging. Small thing, but having fluid intake in the same app as food intake meant I actually looked at it. I averaged 2.6 L/day across the protocol, which is adequate.

What It Did Less Well

Nothing is uniformly excellent, and PlateLens has failure modes worth naming:

• Low-light photos. Five of the 30 days were a restaurant dinner in genuinely dim lighting. The photo estimate came back noticeably wider-variance on those meals than on bright-light daylight photos. On two of them I overrode the estimate manually after the fact.
• Packaged food still in its box. A tray of crackers on its manufacturer box threw the vision model off on one occasion; the model guessed at the cracker count and was off by about 25%. Manual correction was fast, but it happened.
• Very-small-portion calibration. A tablespoon of butter on toast gets absorbed into the toast estimate; the app does a reasonable job, but this is the category where a food scale still wins for the precision-obsessed.

None of these are dealbreakers. All of them are cases where the photo pipeline's error margin widens. For 95% of what I ate in 30 days, the pipeline was well within the noise floor of day-to-day intake variance.

The Blood Markers

Not scientifically meaningful — N=1, no baseline stability, no control — but I ran labs at day 0 and day 30 out of curiosity. Nothing moved dramatically. Ferritin up slightly; HDL unchanged; LDL down 7 mg/dL; fasting glucose unchanged; A1C not repeatable over this time frame.

This is not the point of the experiment. I am mentioning it only so that readers know that I looked, and that 30 days of precise tracking without a dietary intervention did not produce dramatic lab changes. This is expected.

The Honest Takeaway

I came into this experiment a tracking skeptic. I had tested a lot of apps, had never sustained a habit on any of them, and had come to believe that nutrition-tracking apps were a category whose value was oversold and whose user retention was mostly a statistical fiction.

I ended the experiment a less-skeptical tracking person. Not because I was converted to the calorie-tracking worldview — I still think obsessive counting is counterproductive for most people — but because 30 days of 94%-adherence logging showed me things about my own intake patterns that I genuinely could not have known otherwise. Specifically: the 400-calorie-per-day unconscious overshoot, the magnesium gap, the actual-vs-eyeballed portion sizes on my most common foods.

The thing that made this work, and that had never made this work for me before, was the friction level. At 3 seconds per meal, logging happens. At 3 minutes per meal, it does not. The technology difference is enormous, and it translates directly into whether the value of tracking is accessible to a normal person with a normal schedule.

I will keep logging for another 30 days, but less strictly. Probably one to two meals per day rather than every meal. My suspicion is that the value-to-effort ratio is quite good at that level too, and I will report back at day 60.

Editorial note: PlateLens did not sponsor this story. The 30-day protocol was self-directed. Numbers reported are from my actual logged data.