← All food signals
Wellness turn

The Biomarker Diet

Algorithmic diet advice runs on body readings, and the readings it trusts most are the least steady — a stool sample reads differently at every gut-test service.

What is it?

Diet advice built from body readings, a stool test or a glucose sensor, fed through a model that picks foods for one person. The glucose reading holds up well; the gut report varies by lab, and the models turning either into menus are proven mainly on predictions.

Why now?

In 2026 NIST showed that seven gut-test services disagree about one reference stool as much as people differ from each other. The same year, an LLM meal agent appeared, a glucose model lost accuracy on new people, and a review found AI-guided eating evidence inconsistent.

Who’s building it?

Gut-test services, glucose-model researchers and LLM meal-agent teams are building it, with measurement scientists checking.

NISTHuman Phenotype ProjectGut-test servicesGlucose-model researchersLLM meal agents
The full read
Why now, who is building it, where the market goes, and what stays unsettled.
Get weekly →
01 · The market
Who’s in it
Builders — Test services and research teams turning readings into menus.
Buyers — People paying for a report that tells them which foods to eat or avoid.
Gatekeepers — Measurement scientists, and a regulator mostly absent so far.
What’s at stake
Whether a diet built from a gut test reflects the body or the laboratory that happened to process the sample.
Whether glucose models work for people they never met.
Whether meal agents ever get tested on real bodies.
What’s unsettled
Whether gut-test services will adopt a shared reference and quality control.
Whether predicted glucose gains can be measured in real people.
Whether a regulator treats stool-test diet advice as a medical claim.
02 · How the demand matureda progression, not a timeline
Novelty
A personal diet drawn from your own stool or sugar curve.
Retention
A glucose sensor gives live feedback; a gut report arrives once.
Willingness to pay
People pay for the sense of personal fit, before anyone has checked the underlying readings.
Category arc
From a lab report with food lists to an agent that rewrites the meal.
03 · Where the market goesthe forks, not a prediction
→Gut-test services agree on a shared reference material and quality checks, and their results start to converge across labs.
→Glucose models are tested on new people in real trials, not only on held-out data.
→LLM meal agents move straight to consumer apps while their evidence stays on predicted glucose alone.
→A regulator reclassifies stool-based diet advice as a claim that needs validation.
The read turns on whether each biomarker is checked before the model trusts it, or whether the model’s fluency is taken as proof of the reading.
04 · The Reading Check deskwhere the inputs are being tested
The Reference Stool
One sample, seven labs, wide spread.
Provider variation on the scale of donor variation.
Communications Biology
The Newcomer Drop
Glucose models lose accuracy on unseen people.
R 0.731 by meal versus 0.590 by person.
Current Developments in Nutrition
The Predicted Plate
An LLM agent rewrites meals against forecasts.
Evaluated on public datasets, a stated first step.
arXiv (ACL 2026 Findings)
05 · The Wire
01
One stool sample, seven services, results differ like different people.
Communications Biology ↗
02
Glucose data predicts meals well; new people, worse.
Current Developments in Nutrition ↗
03
An LLM agent rewrites meals to cut predicted glucose
arXiv (ACL 2026 Findings) ↗
04
Forty-five glucose-guided diet studies show inconsistent methods
Studies in Health Technology and Informatics ↗
05
AI in precision nutrition is limited by data quality and validation
Nature Communications ↗
06
Gut tests sold with diet advice sit in a regulatory gap
Journal of Law and the Biosciences ↗
The panel on this signal
They agreed on
–Gut-test results currently vary between providers about as much as between people, which undercuts diet advice built on them.
–Glucose sensor data is the strongest input these models have.
–Meal-rewriting AI is judged on predicted glucose, not on measured outcomes in real people.
The clinical-laboratory metrologist disagreed on
–Lab spread is a fixable calibration problem; a shared reference stool plus proficiency testing would shrink it considerably.
Know more about the method →
✦How this signal was built
01 · Panel
A Delphi panel of 4 AI analysts, each reasoning independently.
→
02 · Debate
They challenge and revise each other’s reads over two rounds.
→
03 · Evidence
Every claim is tied to a cited, linked source on The Wire.
→
04 · Synthesis
A moderator writes it into one read, disagreement and all.