What the number actually is
Chronological age is a fact about the calendar. Biological age is a prediction: a statistical model is trained on a population, learning which measurable features correlate with age or mortality, and your measurements are fed in to produce an estimate.
This matters because a model is only as good as its training data and its target. A clock trained to predict chronological age is, at best, a very good age guesser. A clock trained to predict mortality risk is doing something more useful and something quite different.
So the first question about any biological-age result is not "what is my number" but "what was this model trained to predict, and in whom?"
Epigenetic clocks
The best-known approach reads DNA methylation — chemical tags attached to DNA that change in patterned ways over a lifetime. Horvath’s 2013 multi-tissue clock established the field; Hannum, PhenoAge and GrimAge followed, each with different training targets.
GrimAge is generally considered the most useful for health outcomes because it was trained on mortality and morbidity rather than on chronological age alone. It is also the one most sensitive to smoking history, which is either a feature or a distortion depending on what you are asking.
A practical caution: different clocks give different answers for the same sample. If a consumer test reports one number without saying which clock produced it, that omission is itself informative.
The variance problem
Test-retest reliability on consumer epigenetic tests is the least discussed and most important issue. Split a single blood sample in two, send both to the same provider, and results can differ by one to three years.
This has a direct consequence for anyone testing a supplement: if your intervention produces a two-year improvement and your assay noise is two to three years, you have learned nothing. The result is indistinguishable from re-testing on a different Tuesday.
Practical mitigations: use the same provider and assay throughout, standardise collection conditions, take two baseline readings rather than one, and treat differences smaller than the assay’s stated variance as noise no matter how much you want them to be real.
What actually moves it
Ranked roughly by evidence quality, the list is boring — which is usually the sign it is honest.
- Smoking cessation. The single largest documented effect in this space, by a wide margin.
- Body composition. Visceral fat in particular shows consistent associations across clocks.
- Sleep. Chronic short sleep tracks with accelerated epigenetic age across multiple cohorts.
- Exercise. Both cardiovascular and resistance training, with the effect appearing over months rather than weeks.
- Alcohol reduction. Dose-dependent and measurable at moderate intake levels.
- Supplements. Present in the literature, effects generally small, and dominated by everything above.
Testing a supplement against it, properly
If you intend to evaluate any supplement — ours included — against a biological-age panel, the protocol matters more than the product. Two baseline readings to establish your noise floor. One variable at a time. Ninety days minimum, because methylation patterns do not shift in three weeks. Matched conditions on re-test: same provider, same assay, same time of day, similar recent training and fasting state.
And write the prediction down before you start. Deciding afterwards what would have counted as success is how self-experiments produce confident conclusions from noise.