Personalized Nutrition: How It Actually Works

I sell meal plans, so read the rest of this with that in mind. Every number below sits next to the study it came from, and where a paper cuts against my own business I have left it in rather than around it.

How differently people actually respond

Start with the part that is not in dispute.

Zeevi and colleagues, writing in Cell in 2015, fitted 800 people with continuous glucose monitors for a week and logged 46,898 real-life meals, with standardized test meals interspersed among them.[1] Blood sugar responses to the same food varied widely between people: bread that sent one person's glucose climbing barely moved another's. Their own conclusion was that universal dietary recommendations may have limited utility. Two things to keep straight about that study, because they are usually lost: the intervention arm was 26 people for one week, and what it measured was post-meal glucose, not weight.

Berry and colleagues put numbers on the same phenomenon at scale in Nature Medicine in 2020, in the PREDICT 1 study of 1,002 UK adults with an independent US validation group of 100.[2] After identical meals, the population spread in response was 103% for triglyceride, 68% for glucose and 59% for insulin. Person-specific factors including the gut microbiome accounted for more of the triglyceride response than the meal's own macronutrients did, 7.1% against 3.6%, though for glucose the meal mattered more. That one cuts against me and I am leaving it in: biology is doing real work here. The next section is about why measuring it still does not tell you what to eat. The study was funded by and run with Zoe Global Ltd, a company that sells personalized nutrition.

Where personalization does not pay

Variation being real does not mean the tests sold to measure it can tell you what to eat.

What the panels actually predicted

Gardner and colleagues ran DIETFITS, published in JAMA in 2018: 609 adults randomised to a healthy low-fat or a healthy low-carbohydrate diet for 12 months, with both groups coached to raise food quality rather than to hit an extreme.[3] The low-fat group lost 5.3 kg and the low-carbohydrate group 6.0 kg, a difference of 0.7 kg with a confidence interval spanning zero. Then the interesting part: the trial had pre-specified a genotype pattern and an insulin-secretion measure as predictors of who would do better on which diet. Neither worked. No significant diet-genotype interaction, no significant diet-insulin interaction.

Berry's data points the same way from a different angle. Genetic variants contributed modestly to the predictions in PREDICT 1, explaining 9.5% for glucose and 0.8% for triglyceride. Genetics is in the picture. It is not the lever.

So the honest reading is that macronutrient ratio is not where your leverage sits, and neither is a swab. Two sensible diets, both built on real food, land in the same place after a year.

Where it does pay: the Food4Me trial

There is one trial that is the closest published analogue to a service like this one, and it is the reason I am comfortable selling what I sell.

Celis-Morales and colleagues, in the International Journal of Epidemiology in 2017, ran Food4Me: an internet-delivered randomised trial across seven European countries with 1,269 people completing six months.[4] Four arms. One got conventional dietary advice. The others got personalized advice based on their own baseline diet, then baseline diet plus phenotype, meaning body measurements and blood biomarkers, then all of that plus genotype from five diet-responsive variants.

The personalized arms ate less red meat, less salt and less saturated fat than the control arm, took in more folate, and scored higher on the Healthy Eating Index. And the layers made no difference. Adding phenotype did not improve the advice. Adding genotype on top of that did not improve it either. The thing that worked was the cheapest, least glamorous input in the study: what the person was already eating. That is the same thing an intake form asks you for, and it is the part that did the work.

Adherence is the variable that moves

If diet composition is not the lever, something else has to be, and two older papers say what it is with unusual bluntness.

Johnston and colleagues pooled 48 randomised trials covering 7,286 people in a network meta-analysis in JAMA in 2014, comparing named diet programmes head to head.[5] Weight loss differences between the individual named diets were small, and the authors' own conclusion was that this supports recommending whichever diet a person will stick to. So the question is not which diet is best. It is which one you will still be doing in March.

Dansinger and colleagues had shown where the association sits, nine years earlier, in JAMA in 2005, by assigning 160 people to Atkins, Ornish, Weight Watchers or the Zone for a year.[6] Weight loss correlated with self-reported adherence at r = 0.60, and with diet type at r = 0.07, which is another way of saying not at all. Completion across the four arms ran between 50% and 65%, so between a third and a half of the people who signed up did not finish the year. If you have abandoned a diet partway, you were the statistically normal case.

That is the real target. Not the ratio on the label, but whether the plan survives contact with a Tuesday.

What we do, and what we do not

What I would watch for is the difference between personalized and merely segmented. A form that returns the same plan to everyone who ticks the same boxes has sorted you into a bucket, not built you a plan. So here is ours, and you can judge it against that.

A SlimFitNut plan is built by a certified nutritionist from an intake form: your weight, height, age, activity level, goals, the foods you like and the ones you will not touch, and the shape of your week. Out of that comes a calorie target, a protein target of 1.6 to 2.2 g per kg, a structured week of meals and portions, and a schedule that fits the day you actually have. If breakfast has to be eaten before an early commute, the plan respects that. If lunch has to travel, that is a design constraint, not a preference to be talked out of. It is a one-time plan, not ongoing coaching.

We do not sequence your microbiome. We do not run continuous glucose monitoring. We do not order genetic testing, and after DIETFITS and Food4Me I would not know what to do with the results if we did. I cannot infer your microbiome, your hormone status or your glucose response from a questionnaire, and I do not pretend that I can. If you want the version of personalization that involves a lab, that exists, and it is not what you are buying here.

The field itself is more cautious than the marketing

I would rather admit the limits of what we do than imply otherwise, because the field's own reviewers are cautious about it. Ordovas and colleagues surveyed personalised nutrition for the BMJ in 2018 and concluded that most of the supporting evidence comes from observational work using risk factors as outcomes rather than randomised trials with clinical endpoints, and that a good deal more research and regulation is needed before the field delivers what it has been promising.[7] That is a review rather than an experiment.