Can ChatGPT Track Your Macros? An Honest Test

TL;DR

• A general chatbot is genuinely good at one thing: estimating the macros of a described or photographed meal. It is roughly as accurate as a dedicated AI scanner for that single question.
• It is bad at everything that comes after: remembering yesterday, adding up the week, keeping a consistent database, and not quietly changing its numbers between sessions.
• The failure mode is specific. Ask the same meal twice and you often get two different answers, with no flag that anything changed.
• Use a chatbot for one-off estimates and awkward restaurant meals. Use a tracker for anything you want to look back on.
• If you already pay for a chatbot, you do not need to pay for a tracker too. Chowdown is free and does the storage half properly.

What a Chatbot Is Actually Good At

Give ChatGPT a photo of a plate and ask for macros, and you will get a sensible answer. Roughly 450 kcal, 38 g protein, 40 g carbs, 15 g fat, with a short breakdown of how it got there. For a home-cooked meal with visible components, that estimate is usually within 15 to 20 per cent of the truth, which is about the same error band as any AI photo scanner, including ours.

It is even better at things scanners struggle with. Describe a meal in words and it copes: “chicken thigh curry, about two handfuls of rice, cooked with maybe a tablespoon of oil, coconut milk in the sauce.” A photo scanner sees a brown bowl. A language model understands the sentence. If you eat a lot of mixed dishes, stews, or food someone else cooked, this is a real advantage.

It is also excellent at explaining. Ask why your protein target is 1.6 to 2.2 g per kg of bodyweight, or what happens to your weight after a high-sodium meal, and you get a decent answer with the reasoning attached. As a nutrition tutor, a chatbot beats most apps. For the actual numbers, our guide on how much protein you need per day is a faster reference than asking each time.

Where It Falls Apart

It does not remember properly

Chat memory is not a food log. Depending on the product and your settings, a chatbot may remember a handful of facts about you across conversations, but it is not storing a structured record of every meal with a timestamp. Start a new chat and yesterday’s dinner is gone.

That matters because the value of tracking is almost entirely in the aggregate. One meal tells you nothing. Fourteen days of meals tell you whether your protein average is 110 g or 160 g, and that single number is usually the thing that needs fixing.

The same meal gets different numbers

This is the one that surprised us most. Ask for the macros of an identical meal in two separate conversations and the answers drift, sometimes by 80 to 120 kcal. Each estimate is defensible on its own. Together they are useless, because you cannot tell whether the difference between Monday and Thursday is your eating or the model’s mood.

A tracker solves this with a database. The same chicken breast entry returns the same 165 kcal per 100 g every single time. That consistency is boring and it is the entire point. You want your log to move when your eating moves, not when the estimator reshuffles.

Arithmetic over long conversations

Chatbots have improved enormously at maths, but adding up thirty log entries inside a conversation still goes wrong often enough to matter. Numbers get dropped, a snack from earlier in the thread disappears, or a total quietly disagrees with the items above it. You will only catch it if you check the sum yourself, which defeats the purpose.

No packaged food data

Ask about a specific supermarket ready meal and a chatbot will either give you a generic estimate or a number that sounds authoritative and is not on the label. It has no barcode lookup and no live product database. For anything that comes in a packet, scanning the barcode is faster and correct.

The Honest Comparison

Single meal estimate from a photo: roughly a tie. Both land within a sensible range.

Meal described in words: chatbot wins clearly. This is what language models are for.

Packaged and branded food: tracker wins. Barcode and database beats guessing.

Daily and weekly totals: tracker wins outright. This is not a close call.

Consistency across days: tracker wins. The database is the feature.

Explaining nutrition concepts: chatbot wins. Apps are bad teachers.

Cost: ChatGPT free tier is capable but rate-limited on image uploads, and the paid tier is around £20 a month. Chowdown is free with no premium tier, so there is nothing to compare on price. If you want the fuller breakdown of what “free” tracking apps actually cost you, see why free macro tracking apps aren’t really free.

How to Actually Use Both

The sensible setup is not choosing one. It is knowing which question belongs to which tool.

Use the chatbot when you are at a restaurant with no nutrition information and need a defensible estimate, when a meal is complicated enough that describing it beats photographing it, or when you want to understand something rather than record it. Then take the number it gives you and log it in your tracker, so the estimate survives past the conversation.

Use the tracker when the answer you want is a total, a trend, or a comparison with last week. That is most of the time. Snap the photo, accept or adjust the estimate, and move on. The point is that in ninety days you can look at a chart and see something true.

If you want a rule of thumb: the chatbot is a calculator, the tracker is the spreadsheet. Nobody manages their finances by asking a calculator what they spent last month.

What This Means for Tracking Apps

The interesting part of this test is not that the chatbot lost. It is how much of a tracker’s job a general model already does well. Estimation from a photo used to be the hard, proprietary part. It is now close to a commodity, as we’ve written about in how AI food scanner apps actually work.

What is left is the unglamorous half: a consistent database, a durable log, honest aggregates, and a way to see three months at a glance. That half was never the clever bit, and it is the bit that actually changes behaviour.

It is also why charging £40 a year for macro tracking is getting harder to justify. If the estimation is a commodity and the storage is a solved problem, a paywall around either is a business decision rather than a technical one. Chowdown is free because we think that is where this lands anyway. See how we compare on that front against MyFitnessPal and Cronometer.

Try the chatbot for a week if you like. Then look back and try to work out what your average protein intake was. That exercise usually settles the question.