The missing test in an AI meal plan
Predicting a meal response and improving health require different evidence. The METHOD trial and newer reviews show why the exact programme, outcome, and comparator must remain visible.

A meal score can predict a response without proving that following the score improves long-term health. Those are separate claims, and a nutrition product needs evidence for the one it sells.
Personalised nutrition is appealing because it offers a specific answer to a familiar problem: general advice can be difficult to turn into an ordinary meal. A system that uses personal measurements may make that decision feel less arbitrary. The useful question is what changes when people actually follow its recommendations.
Prediction needs a defined target
An algorithm might predict blood glucose after a meal. Another might rank foods using several biological measurements. A more accurate prediction is evidence about the target the model was trained to predict. It does not automatically cover hunger, digestive symptoms, or later disease.
A lower reading on one measure can be relevant without being a complete description of a meal. The interpretation depends on the person and the health question. Treating every rise in glucose as a problem would make a much broader claim than predicting the rise.
I would ask a product to name the outcome it expects to improve. The explanation should then connect that outcome to a trial of the actual intervention. An attractive personal dashboard can make a measurement easier to inspect while leaving the health claim unresolved.
Read the trial as a test of a whole programme
The 2024 METHOD trial tested an 18-week app-based personalised dietary programme against advice based on U.S. dietary guidance. The programme used biological measurements and personal food scores alongside general advice. It reported a greater improvement in triglycerides, while the between-group result for LDL cholesterol, the other primary outcome, was not statistically significant.
That is more specific than saying a personalised diet beat generic advice. Different outcomes had different results. The trial also evaluated the programme as delivered, so its findings cannot isolate the effect of a scoring algorithm from the rest of the support experience.
The participant mix limits simple generalisation. Most participants were women, and the trial's average body mass index was in the obesity range. The results need to be interpreted in that population and over that duration. A claim about lifelong prevention for everyone would require different evidence.
Commercial interest belongs in the reading of the study too. The product being evaluated is sold by a company involved in the work. That relationship does not erase the findings; it makes independent replication and the exact comparison especially relevant to a purchasing decision.
The product can change after the trial
ZOE's evidence page, updated in August 2026, notes that the original study used its original tests. That detail is important when reading a current product page beside an older clinical trial.
A change in the tests or the recommendation process can change what evidence applies. A company should be able to identify which parts of the present programme were evaluated and which rely on an extrapolation. The date on a marketing page cannot turn an earlier intervention into a trial of every later version.
The same issue affects a competing product that quotes another company's results. Similar language about personalisation does not establish that the inputs, recommendations, and support are equivalent. The intervention needs to be specified closely enough that a reader can tell what was actually tested.
The wider evidence still varies
A 2026 review of randomised personalised-nutrition trials examines approaches that differ in the information used for tailoring and the support participants receive. That variety makes the phrase personalised nutrition a broad research category. A result for one approach cannot be assigned to every app using the label.
NIH's Nutrition for Precision Health programme describes a research goal of developing algorithms that predict individual responses to food and dietary patterns. A research programme pursuing that goal is evidence of an active question. Its existence cannot validate a commercial product's particular promise.
The research question is worth investigating. Better matching a recommendation to a person's circumstances may improve its usefulness. The outcome still needs to be observed, with a comparison that shows what the extra measurement and personalisation contributed.
Judge the service by the decision it improves
My preference would be a product that explains the measurement in ordinary language and makes its uncertainty visible. It should distinguish a food suggestion from a medical conclusion, with a way to correct information that the recommendation depends on.
The price also needs to be related to a defined service. Paying for coaching, a test, and an app creates several possible sources of value. A claim that AI removes all guesswork makes those differences harder to examine.
The strongest case for a personalised plan is a measured improvement in an outcome relevant to its intended users, compared with a credible alternative. Until that connection is shown, a personal score remains information to interpret within the health question it was designed to address.