Not what we buy, and not what we tell surveys. This report reads the food diaries themselves, across our two apps: ParrotPal (global, AI food diary) and Turtle Method (UK, women's health). The brands, the foods, the weekend, and the one measurement most reports skip: whether the diaries add up against the scale.
Built on the full history of both apps: every food log since Turtle Method began in 2019 and since ParrotPal launched in 2023. Every count on this page is a food-diary entry with full item detail: 25.4 million of them (15.6M ParrotPal, 9.8M Turtle), from more than 189,000 people who logged food; calorie and macro averages use complete logged days. Anonymised and aggregated throughout; every group shown has at least 50 people and no share rests on fewer than 10 people.
Two different apps, two different audiences, one grocery truth. When people write down what they actually ate, the same five own-label names top both diaries, and the biggest advertised brand reaches about 40% as many people as Tesco in each.
Tesco, Sainsbury's, Aldi, Asda and M&S hold the top five in both diaries, in slightly different orders.
The biggest CPG name in each diary is Cadbury. By people reached it gets to 40% of Tesco's members in ParrotPal and 44% in Turtle Method. By log volume its share is under 10%, which makes own-label's lead even larger.
ParrotPal's 2026 top 30 carries five protein brands: MyProtein, Optimum Nutrition, Grenade, Quest and Premier. Turtle's longer UK history carries one. The era shows up in the leaderboard.
Brand names appear exactly as users logged them. Family variants ("Tesco Finest", "by Sainsbury's", "M&S Food") rank separately. ParrotPal's item-level records begin in September 2023 and cover a share of users that grew as server-side logging rolled out; the rankings are stable across every era we can cut.
No logos on this podium. In both apps, across every year we hold, the most-logged foods are whole foods.
Share of all 5.8 million named food logs in the window. One food in every 71 logged anywhere in the world is a banana.
Average of each brand's rank across both apps. Five retailers come before any brand with a TV advert.
In Turtle, people pick from a database, so roughly 100 brands cover 71% of branded logs and 2,000 foods cover two thirds of everything. In ParrotPal, people talk, and the AI writes "chicken breast", "grilled chicken breast" and "cooked chicken breast" as three foods. The name space explodes to 703,760, and only one log in nine carries a brand at all. Concentration is a property of the logging instrument, not the diet.
A canonical dictionary of about 2,000 foods would cover most of the human diet in both apps. The diaries have been quietly building that dictionary for five years.
Chicken and turkey supply 22.9% of every protein gram logged. Supplements are now the second biggest identifiable source, level with all red meat combined. Plant proteins and pulses supply 1.5%. On the fibre side, one fruit leads everything: banana alone carries 3.9% of every fibre gram in the diary.
The plant-based era is loud online and nearly absent in the diaries.
Yes, chips out-deliver most vegetables. Volume beats virtue. The typical logger still lands well under the 25 to 30g guideline (median around 12g a day), which is why fibre, not protein, is the open goal in food right now. The stealth fibre brands: Mission wraps, Quest and Huel.
Logged calories on complete days are almost identical: 1,887 in the UK, 1,877 in the US. Americans log 9g more protein and more half-finished days. The food gap is cultural: Britain logs tea, semi-skimmed milk, honey and cucumber. America logs coffee, white rice and scrambled eggs. Both countries put a banana first.
About 9,700 UK users and 5,300 US users. These are logged figures, not measured intake. See the honesty gap below.
🇬🇧 Tesco · Aldi · M&S · Sainsbury's · Asda. One protein brand in the top ten (MyProtein).
🇺🇸 Trader Joe's · Kirkland · Great Value · Fairlife · Chobani. Four protein brands in the top ten: Fairlife, Premier, Quest, Built Bar.
The American diary is a protein aisle with a grocery store attached. The British diary is a supermarket with a protein shelf.
Men log 2,080 calories and 131g protein a day; women log 1,791 and 102g. The one number the sexes share is fibre: 20.0g and 20.2g. The food skews below are how many times more likely one sex is to log the item at least once.
At least 27% of ParrotPal users logged alcohol in the last six months, and at least 19% of Turtle's women across 2025 and 2026. "At least" is doing honest work: alcohol is the most under-logged category in any food diary. Nearly all of it is logged generically. The brands that do appear are beers, and an alcohol-free one is already in the top five.
Beer brands logged by name: Stella Artois, Peroni, BrewDog, Birra Moretti, San Miguel, and Lucky Saint, the alcohol-free beer already sitting among them.
Weekends account for 22.9% of all logs. Any food far above that line lives on the weekend; far below it, the food belongs to the working week. The results read like a national portrait.
Brands follow the same clock. Weekend: Chipotle 34%, McDonald's 32%, Haribo, Magnum, Pringles. Weekday: Pret a Manger at 11%, the purest commute signal in the data, then Uncle Ben's and Activia.
One more layer: engaged users under-log weekends (their weekend share is 23% against a 28.6% calendar benchmark), and what they do log runs about 6% more calories. The true weekend is even more roast-shaped than the diary shows.
A food diary records behaviour. Bodyweight over 90 days records energy balance, at roughly 7,700 kcal per kilogram. Run both against each other and you can measure something no survey can: how far the diaries drift from reality. Under-reporting is universal in nutrition research; the doubly-labelled-water literature puts it at 30 to 50% in some groups. Our diaries are no exception, and now we know by how much.
The energy deficit their scales prove: 6.0kg lost over 99 days.
Their diaries claim 1,891 kcal a day. If their maintenance sits near 2,350, the claim and the scale agree. These diaries add up.
Their scales prove maintenance: +0.1kg over 97 days.
Their diaries claim 1,851 kcal a day. A group averaging 84.7kg does not maintain on 1,851. These diaries are missing roughly 450 to 500 calories a day, and that allows adaptive thermogenesis a generous 15% of the deficit.
Same protein per kilo (1.36 vs 1.35), same fibre, same claimed calories. Across the whole report, treat every calorie figure as "logged", never as "eaten". Turtle Method's own evidence work reached the same conclusion independently with an implied-intake method, which makes this a two-app replication.
Both apps come from the same founder, and their data lakes stay fully separate: no user-level data ever crosses between them, only checksummed aggregate statistics. That separation is what makes this section possible. When two separately instrumented datasets, with different products, audiences and eras, produce the same number, the finding stops being "one app's users" and starts being how people eat. We call it cross-product replication.
| Finding | Turtle Method | ParrotPal | |
|---|---|---|---|
| Engaged women's weekend share of logged days (calendar benchmark 28.6%) | 23.0% | 23.5% | Replicates |
| Light users with zero weekend logging days | 65.4% | 62.7% | Replicates |
| Weekend calories run higher on logged days | +5.3% | +6.4% | Replicates |
| Own-label supermarkets hold the top five brand spots | 5 / 5 | 5 / 5 | Replicates |
| Cadbury's member reach vs Tesco's (first true brand in both; by log volume it is ~10%) | 44% | 40% | Replicates |
| Whole foods top the diary: banana, berries, eggs in both top tens | Yes | Yes | Replicates |
| Women's mean daily steps, 2026 | 7,740 | 7,016 | Same ballpark |
Fibre traces the same U-shape: 16.7g down to 14.1g, back to 17.5g. The top foods barely moved in five years, but the accessories flipped: 2021's diary had "mini light cheese" and "mini babybel light" in its top 25; 2026's has chia seeds at #7. Complete logged days, women; each year is a different mix of members, so read the shape, not tiny year steps.
Logged figures on complete days (at least 800 kcal recorded). ParrotPal's base is younger, more global and more protein-era; Turtle's is UK women on a structured method. The point is not which number wins. It is that both sit in the same physiological neighbourhood while the behavioural constants above are nearly identical. Turtle figures come from its checksummed evidence-base aggregate pack of 25 August 2026; its demographic feeds are not yet in the live lake.
LeanShield is the muscle-protection score inside both apps. Since it arrived in the ParrotPal app in March 2026, more than 2,700 users have opened it. The lazy comparison would flatter it, because engaged people use every feature more. So the comparison below is stratified: people who opened LeanShield's training pages against non-users at the same level of food-logging commitment.
A 17 to 21 point gap that holds in every band: more than 1,000 LeanShield users vs more than 5,000 non-users at the same logging levels, since the March launch. Any recorded workout shows the same pattern at +21 to +25 points. Train-page visitors are self-selecting people who already care about training, so this is association, not proof of cause.
Protein is flat. The differences by band are −1.7g, −2.8g and +2.0g. The score travels with training, not the plate, so far. An honest null we publish anyway.
Retention is too early to call. The group of new LeanShield users in the heaviest logging band is still under 50 people, so we don’t publish a figure for it yet. We re-run as the cohort grows.
Trial conversion cannot be measured yet. Fewer than 1 in 100 of more than 5,000 trial users opened LeanShield in their first week. Our hypothesis is that the feature sits too deep in the app to be found during a trial; the data measures the discovery rate, not the reason.
The nudge machine is live in Turtle Method. Its JITAI attribution system holds more than 1.1 million decision events across more than 53,000 members, and the explicit view-to-click rate on recommendations is remarkably stable across three policy versions: 10.3% (V4), 11.0% (V5) and 10.0% (V6), on about 3,600, 2,800 and 4,900 viewed recommendations. That is strong interaction evidence, not yet causal evidence. The randomised LeanShield study schema exists, but no participants have been enrolled, so there are no arm-level outcomes to publish yet. They will appear here when there are.