Your response to food is personal, but the payoff from diet algorithms is modest

Key takeaways

• In a 2015 Israeli study of 800 adults, the top tenth of responders to bread had blood sugar rises more than five times those of the bottom tenth.
• In ZOE’s own 18-week trial, its app beat a government diet leaflet by about 5 pounds of weight loss, but the company paid for the study and the comparison group got much less hands-on support.
• In the DIETFITS trial of 609 adults, matching people to a low-fat or low-carb diet by their genes did not improve weight loss.
• The NIH announced $170 million in 2022 to build food-response algorithms, and in 2026 its website still listed enrollment as open.
• Before you pay, ask what the app personalizes, what it was tested against, and who funded the test.

In 2015, scientists at Israel’s Weizmann Institute of Science fitted 800 adults with glucose sensors for a week and logged 46,898 of their meals¹. The same foods produced strikingly different results from one person to the next.

Take bread. In the tenth of people who responded least, blood sugar rose only a little. In the tenth who responded most, the average rise was more than five times larger¹.

That finding launched an industry. Apps and test kits now promise a diet built for you from your glucose readings, your gut bacteria or your genes, often for a monthly fee.

The science behind the first claim is solid: people do respond differently to the same food. The harder question is whether an algorithm that predicts those responses makes you healthier than good general advice would. So far, trials say the answer is yes, a little, sometimes.

This article walks through the key studies, who paid for them, what they did and didn’t show, and the questions to ask before you sign up.

The landmark finding: the same bread, very different blood sugar

The Weizmann team gathered more than 1.5 million glucose readings. They also collected blood tests, activity data, questionnaires and stool samples to profile each person’s gut bacteria¹.

They then used machine learning, in which a computer finds patterns in large data sets, to predict each person’s blood sugar response to a meal. The model drew on 137 factors, including the microbiome¹.

The model beat carbohydrate counting, the standard approach for people with diabetes. Scientists score predictions from 0 (useless) to 1 (perfect). The algorithm scored about 0.70 in a separate group of 100 people, compared with 0.38 for counting carbs¹.

Next came a small test. Twenty-six people each ate a “good” diet and a “bad” diet for one week apiece, chosen either by the algorithm or by experts. The good diets produced lower blood sugar after meals¹.

Two cautions apply. One week of glucose curves in 26 people says nothing about long-term health. And a low glucose response doesn’t make a food healthy. On average in the study, ice cream and beer raised blood sugar less than rice or potatoes did¹.

A British study of 1,002 people found the same, and genes explained little

In 2020, a UK and U.S. research team published PREDICT 1 in Nature Medicine. It tested 1,002 healthy adults in the United Kingdom, many of them twins, with a 100-person check group in the United States².

After identical meals, people varied widely in blood sugar, insulin and especially triglycerides, a type of blood fat². Between-person variation in triglyceride responses was even larger than for glucose².

The surprises were in what drove the differences. For blood sugar, the meal’s mix of carbohydrate, fat and protein explained about 15% of the variation, and gut bacteria about 6%². Genes had only a modest effect on the predictions: 9.5% for glucose and under 1% for blood fats².

That matters if you’re shopping for a DNA-based diet. In this study, your genes were a small part of the story.

A machine-learning model predicted glucose responses well, scoring 0.77 on the same 0-to-1 scale. It did worse for triglycerides, at 0.47².

PREDICT 1 was supported by the company Zoe Global, along with grants from the Wellcome Trust and the UK Medical Research Council. Several authors were company consultants or employees². That doesn’t make the data wrong, but it is worth knowing.

Predicting a spike isn’t the same as improving your health

These studies predict a short-term response: how high your blood sugar or blood fat climbs over a few hours after a meal. Health outcomes such as weight, heart disease and diabetes play out over years.

In 2021, the National Academies of Sciences, Engineering, and Medicine held a workshop on precision nutrition. Robert Califf, then at Verily and Google Health, argued that single biomarkers rarely predict how a food will affect your health, except in specific nutrient deficiencies⁹. He warned that recommendations built on unproven markers could cause harm that might not show up for years⁹.

At the same workshop, Josh Anthony of Nlumn said most personalized nutrition companies have no published data supporting their own programs⁹. Speakers also flagged gaps in who gets studied. Much genetic data comes from people of European ancestry, and apps are used mostly by younger, educated women⁹.

Glucose-based plans also depend on sensor readings, which can run high or disagree in people without diabetes. Our article on glucose sensors explains why.

So the real test of a personalized diet is a randomized controlled trial, where people are assigned by chance to the personalized plan or to another plan. Several have now been done.

The big test: ZOE’s app against a government leaflet

In May 2024, ZOE scientists and academic collaborators published the METHOD trial in Nature Medicine³. They randomly assigned 347 U.S. adults ages 41 to 70 to one of two groups for 18 weeks.

The 177 people in the personalized group did ZOE’s at-home test kit, including a glucose sensor, standard test meals and finger-prick blood tests. The app then scored foods for each of them³. The 170 people in the control group got a PDF leaflet on the U.S. Dietary Guidelines, a video lesson, online resources and weekly check-in emails³.

Here is what the personalized group gained compared with controls³:

• Triglycerides fell slightly more, one of the two main outcomes.
• LDL cholesterol, the other main outcome, showed no difference.
• Weight dropped about 2.5 kilograms (5.4 pounds) more, and waists shrank about 2.4 centimeters (under an inch) more.
• Diet quality scores rose more.
• Blood pressure, insulin and blood glucose showed no difference.

Those are real but modest gains. The bigger issue is the comparison. The authors acknowledge that, by design, the control group was not matched for contact or intensity³. A leaflet and emails are a low bar.

Nicola Guess, a dietitian and researcher at the University of Oxford, raised more concerns. Only the ZOE group had to record each meal as they ate it. No one was blinded, and a sensor on your arm keeps telling you you’re being watched⁴. In a 2025 disclosure, Guess listed advising Beyond Meat and giving input to the diet app MyFitnessPal’s behavior programs, and no ties to glucose-sensor makers⁵.

The paper states that ZOE Ltd funded the research and helped design it, collect and analyze data, and write the paper. Three company co-founders were authors³. The trial also tested the first version of the algorithm, so today’s app may differ³.

When personalization faced a real diet, the gap narrowed

What happens when the comparison group gets a solid diet, not a leaflet? The METHOD authors summarized several such trials in their own paper.

In an Israeli trial in people with prediabetes, a diet built to keep each person’s post-meal glucose low improved blood sugar control more than a Mediterranean diet. Weight loss, though, did not differ at six months³⁹. In another trial, a personalized diet aimed at post-meal glucose did not lead to more weight loss than a low-fat diet³. A trial called PREVENTOMICS, which grouped people by metabolic profile, found no benefit for weight or body fat³.

An earlier and larger test came from Europe. In the Food4Me trial, 1,269 adults in seven European countries finished six months of online advice⁸.

People who got personalized advice ate a little less red meat, salt and saturated fat and a bit more folate than people who got standard advice. Their diet-quality scores rose slightly more⁸.

The twist: advice based on blood tests or genes did no better than advice based only on what people already ate⁸. Personalizing around a person’s real diet did as well as adding biology.

Gene-matched diets have already been tested, but they didn’t work

Stanford’s DIETFITS trial, led by Christopher Gardner and published in JAMA in 2018, put a popular idea to a direct test. Researchers randomly assigned 609 adults with overweight or obesity to a healthy low-fat or a healthy low-carbohydrate diet for a year⁶.

Before the diets began, the team checked each person for a gene pattern and an insulin measure they thought might predict which diet would suit them⁶⁷. Neither one did⁶.

The low-fat group lost an average of 5.3 kilograms, about 12 pounds, and the low-carb group 6.0 kilograms, about 13 pounds. The difference wasn’t statistically significant⁶. People matched to “their” diet by genes did no better.

What did vary was people. Within both diets, some people lost more than 60 pounds while others gained 15 to 20 pounds⁷. The National Institutes of Health (NIH) was among the funders⁷.

Both groups heard the same basic message: build meals around whole foods and vegetables, and cut back on sugar and white flour. Gardner also urged people to eat in a way that didn’t leave them hungry or deprived, so they could stick with it⁷.

That is a familiar message to anyone in natural health. The idea that each body is unique, sometimes called bio-individuality, has real support: people respond differently, and Gardner’s team saw it⁷. But so far, the clearest gains in these trials came from basics that tradition has long taught: real food, vegetables and less refined sugar.

The government’s own algorithm is still being built

One clear sign that the science is unfinished: the NIH is still building its own algorithm. In January 2022, it announced $170 million over five years for Nutrition for Precision Health, a study of 10,000 volunteers drawn from its All of Us research program¹⁰.

The aim is to develop algorithms that predict how individuals respond to foods and eating patterns¹⁰. “We know that nutrition, just like medicine, isn’t one-size-fits-all,” said Holly Nicastro, the study’s coordinator¹⁰.

The design is demanding. All participants do diet tracking, lab samples and a test meal. Smaller groups of about 1,500 to 2,000 follow assigned diets at home, and about 500 to 1,000 live at a research center while eating study meals¹¹.

Enrollment opened in 2023¹². When the NIH reviewed the program’s homepage in May 2026, it still said enrollment was open and mentioned no results¹³. The NIH says the findings might someday help clinicians tailor nutrition advice¹².

A July 2026 perspective article in Nature Communications explains why patience is warranted. The most powerful models often can’t explain their choices, its authors note. Their training data also lean toward wealthier study groups, so a model may stumble when used on people whose diets, health or circumstances differ¹⁴. Adding layers of biological data has so far brought “modest and context-dependent” gains¹⁴.

Questions to ask before you pay

Use this scorecard before you buy a personalized nutrition program. It’s built from the trials above.

Question	A reassuring answer	A red flag
What exactly is personalized?	It names the inputs (food log, glucose, stool, DNA) and explains how each changes advice	“Your unique biology” with no detail
Was this program tested in a randomized trial?	A published trial of this product, with results you can read	Only testimonials or a study of a different version
What was it compared with?	A real diet plan with similar coaching	A leaflet, or no comparison group
Who paid for the study?	Independent funding, or company funding stated plainly	No funding information
What improved, and by how much?	Weight, blood fats, A1C or blood pressure, with numbers	Only “glucose spikes” or “felt better”
Does advice match the basics?	Vegetables, fiber, whole foods and less added sugar	Rewards foods like ice cream for a flat glucose curve
What happens to your samples and data?	Clear deletion rights; no selling of DNA or stool data	Vague privacy terms
Can you cancel easily?	Simple online cancellation and clear monthly price	Hard-to-find terms or auto-renewal traps

If you have diabetes, kidney disease, a history of disordered eating, or take medicine that affects blood sugar, talk with your clinician or a registered dietitian before following any app’s plan. If a program makes disease-cure claims or won’t let you cancel, you can report it to the Federal Trade Commission at ReportFraud.ftc.gov.

What the evidence supports

Claim	Evidence level	What the research shows
People’s blood sugar and blood fat responses to the same foods differ widely	Strong evidence	Seen in large studies in Israel, the UK and the U.S.12
Algorithms predict an individual’s short-term glucose response better than carb counting	Promising but early	Beat carb counting in research groups; real-world app accuracy unknown1214
Genes explain most of how you respond to a meal	Evidence says no, or unsafe	Genes had a modest effect in PREDICT 12
Choosing low-fat or low-carb by your genes improves weight loss	Evidence says no, or unsafe	No benefit in the 609-person DIETFITS trial6
Adding blood or gene data beats personalizing from your diet alone	Evidence says no, or unsafe	No added benefit in Food4Me8
A personalized app program beats minimal standard advice	Promising but early	Modest gains in one company-funded trial with an unmatched control34
Personalized plans beat a well-designed diet for long-term health	Not enough evidence	Mixed short trials; no long-term outcome data39
Commercial algorithms work equally well for all ages and backgrounds	Not enough evidence	Study groups were narrow3914

Questions readers ask

The bottom line

You really are different from the person next to you at the table, and your blood sugar proves it. But so far, algorithms that turn that difference into a diet have delivered modest extra benefits in short trials, often funded by the companies selling them. Gene-matched diets have failed head-to-head tests.

Before you pay, look for a published trial of the exact product against a real diet plan, and check who paid for it. Meanwhile, the advice that held up in every trial is free: whole foods, vegetables, fiber, and less added sugar and refined flour.

Related reading

• Glucose sensors catch every spike, and in healthy people most spikes are normal
• Who paid for the study matters, even when the science is real
• When the company reading your labs also sells the fix
• Health apps share more than you think, and you can lock yours down
• The wellness subscription you meant to cancel

Sources

1. Zeevi D, et al. “Personalized Nutrition by Prediction of Glycemic Responses.” Cell, 2015. wisdom.weizmann.ac.il
2. Berry SE, et al. “Human postprandial responses to food and potential for precision nutrition.” Nature Medicine, 2020. nature.com
3. Bermingham KM, Linenberg I, Polidori L, et al. “Effects of a personalized nutrition program on cardiometabolic health: a randomized controlled trial.” Nature Medicine, 2024 (open-access full text). iris.unitn.it
4. NutraIngredients. “Zoe hails personalized nutrition trial success, results come under scrutiny.” May 10, 2024. nutraingredients.com
5. Science Media Centre. “Expert reaction to study looking at continuous glucose monitors and measuring blood sugar levels in people without diabetes” (declared interests of Nicola Guess). February 26, 2025. sciencemediacentre.org
6. Gardner CD, et al. “Effect of Low-Fat vs Low-Carbohydrate Diet on 12-Month Weight Loss in Overweight Adults and the Association With Genotype Pattern or Insulin Secretion: The DIETFITS Randomized Clinical Trial.” JAMA, 2018. pubmed.ncbi.nlm.nih.gov
7. Stanford Medicine. “Low-fat or low-carb? It’s a draw, study finds” (news release via ScienceDaily). February 20, 2018. sciencedaily.com
8. Celis-Morales C, et al. “Effect of personalized nutrition on health-related behaviour change: evidence from the Food4Me European randomized controlled trial.” International Journal of Epidemiology, 2017. pubmed.ncbi.nlm.nih.gov
9. National Academies of Sciences, Engineering, and Medicine. “Challenges and Opportunities for Precision and Personalized Nutrition: Proceedings of a Workshop—in Brief.” National Academies Press, 2021. ncbi.nlm.nih.gov
10. National Institutes of Health. “NIH awards $170 million for precision nutrition study.” NIH news release, January 20, 2022. nih.gov
11. NIH Common Fund. “Nutrition for Precision Health: Frequently Asked Questions.” NIH. commonfund.nih.gov
12. NIH Common Fund. “NIH launches largest precision nutrition research effort of its kind.” NIH, May 16, 2023. commonfund.nih.gov
13. NIH Common Fund. “Nutrition for Precision Health” (program homepage). NIH, last reviewed May 27, 2026. commonfund.nih.gov
14. Massara P, et al. “Applying Artificial Intelligence and machine learning in precision nutrition.” Nature Communications, 2026. nature.com