How the grading works.
Every supplement, drug and habit in your routine, checked against the biology of aging, the human evidence, and who paid for the studies.
This page shows how that check is made and why we built it this way. The short version comes first. Every section opens with the main idea; the full rules sit in panels marked “Go deeper”.
Four steps, the same for everything we grade.
- 1Find the studiesFixed searches under every name the product goes by; studies chosen by design, never by result.
- 2Weigh each oneSorted by what it tested (cells, animals or people) and how well it was run.
- 3Check the biologyIs there a documented route from the ingredient to an aging driver?
- 4Turn it into wordsBiology times evidence, translated into one plain rating for each aging driver.
Most health claims reach you with the evidence already filtered.
A label, an ad or a podcast segment usually shows you the studies that worked, often in mice, rarely with who paid for them. Almost nobody hands you the whole record, sorted and weighed. That is the gap this is built to fill.
Why grade against the biology of aging?
Medicine sorts the body by organ and by disease. Aging research has found a short list of shared processes that drive the decline behind many of those diseases at once. The science calls them the hallmarks of aging: twelve in the field's 2023 review, plus one candidate we add, the shrinking thymus. We call all of them the 13 aging drivers. Rating everything against the same 13 gives one map for a whole routine, from fish oil to strength training. It is young science, and the pages say so.
Why not just ask a chatbot?
- Answers from whatever it absorbed in training.
- The same question can get a different answer.
- Sources that are hard to check, or that do not exist.
- The same written rules, applied to every study of every intervention.
- Language models read the papers and propose mechanism scores under written rules. Code computes every rating and the words that report it. A person approves the rules, never a single score.
- Every study is checked against PubMed before it is allowed in.
No company can pay for a rating. The rules are the same for everything we grade, and they are written down on this page.
Find the studies, and choose them fairly.
For each intervention we search PubMed, the U.S. National Library of Medicine’s database of medical research, under every name the product goes by, then narrow to studies of the product as it is sold and the biology of aging. Tap a stage.
Chosen by design, never by result. When there are more human studies than a first pass can carry (120), we choose by design: pooled analyses first, then the largest and longest trials. Whether a human study found an effect is never a reason to keep it or leave it.
- Fixed searches of PubMedWe search PubMed under the intervention's name and its other names with a fixed set of searches built around study design (trials, pooled analyses, large cohorts and more), then remove duplicates. Cell and animal studies are added for mechanism support.
- The product as people take itWe keep studies of the product as people actually take it. A different route (an injection on a page about a pill), a combination product, a food that merely contains the ingredient, or the body's own level of the compound is set aside. Trials that test a popular claim stay in even when the claim is not about aging.
- Placed on the evidence ladderEach study is placed on the seven-level ladder from its design. For trials we record the number of people enrolled, not the smaller number who finished.
- Chosen by design, never by resultWhen there are more human studies than a first pass can carry (120), we choose by design: pooled analyses first, then the largest and longest trials. Whether a human study found an effect is never a reason to keep it or leave it.
- Graded, then re-auditedThe target is 50 to 80 studies a page, fewer when the research is thin. Pages are then re-audited one at a time in a fresh session that searches again for studies we missed, large no-effect trials included.
The rule that matters most: whether a human study found an effect is never a reason to keep it or drop it. Studies that found nothing are brought in on purpose and shown with a clear label.
Go deeperHow we choose when there are too many studies
Below the cap, every study in scope comes in. A first pass carries at most 120 studies; only when the human studies in scope would pass that number do we choose, by design, in a fixed order:
- every pooled analysis of the product as people take it;
- every randomized trial with 100 or more people, or 12 weeks or longer, or a main outcome registered in advance;
- the rest by design: double-blind and placebo-controlled first, then larger, then longer.
Then a check: the share of kept randomized trials that found no effect has to be at least the share in the full pool, less 10 points. If it is lower, no-effect trials are swapped in by the same order. Cell and animal studies are chosen for mechanism support: at least six of each where they exist, at most 20 in a large literature.
Go deeperWhat we set aside, and why
The page grades the product as people take it. So we set aside studies of a different route (an injection or a skin cream on a page about a pill), hospital and emergency uses, combination products where the ingredient cannot be separated from the others, foods that merely contain it, and studies of the body’s own level of the compound rather than of taking it. Retracted papers, preprints and confirmed predatory journals are left out too.
Studies in scope that we leave out are listed one by one with the reason. Search results that are plainly out of scope are logged as counts by category.
Go deeperHow we catch the studies we missed
Each page is re-audited in a fresh session that did not build it. Among other checks, it searches PubMed again for studies we missed. When the literature is small (40 or fewer human trials), every in-scope study not yet seen is added. When it is large, up to 20 are added, and at least five of those go to large trials that found no effect, when there are any. Additions and removals are listed in the page’s version note. The re-audit program began in October 2026 and is working through every page.
Weigh each study by what it can actually show.
A test in a dish is not a trial in people, and one small trial is not a pooled analysis of many. Every study goes on a seven-level ladder. Tap a level.
People are assigned by chance to the intervention or to a comparison group, so a result is far less likely to come from a hidden difference between the groups. Three well-run trials, or one with two supporting studies, reach the second rung.
The evidence score for each aging driver is set by how many good studies agree, not by the single best one. Then every randomized trial and pooled analysis gets a quality flag. A trial is read on three facts: how many people, how long, and whether it was blinded (neither the volunteers nor the researchers knew who got what).
- People enrolled120meets the bar (80 or more)
- Length12 weeksmeets the bar (8 weeks or more)
- Blindednot statedone deduction: we give no credit for what we cannot see
Go deeperNot stated is not the same as poor
Each fact is adequate, short of our bar (fewer than 80 people, shorter than 8 weeks, no blinding), or not stated in the paper’s summary. A short or unstated fact costs one deduction, because we give no credit for what we cannot see. But we never call a trial small, short or unblinded unless the paper says so; the card says “not stated” instead. A trial is excluded from scoring only when all three shortfalls are measured, and even then it stays on the page.
Go deeperHow pooled analyses are checked
A pooled analysis is read on its own facts: did it use a recognized quality tool, report how much the trials disagreed, pool at least five trials and 500 people, register its plan in advance, and avoid being funded by a single manufacturer. One shortfall makes it yellow; two or more of the serious ones (no quality tool, no measure of disagreement, no registered plan, funding by one manufacturer, among others) make it red. It is excluded if it appeared in a predatory journal, was retracted, or every author works for the manufacturer. Every pooled analysis stays on the page whatever its flag.
Go deeperTrials of exercise and other habits
Nobody can be blinded to whether they are lifting weights, so blinding is not held against a trial of a behavior. Habits are held to the same evidence bar as anything else.
Check that the biology adds up.
A study can show that something changed. The mechanism says whether there is a real reason to expect it. We trace every step, from the ingredient to the molecule it acts on, to a documented biological pathway, to an aging driver.
A mechanism score of 3 or more has to rest on a pathway from our curated map of 52 documented pathways. A pathway that is not on the map supports a 2 at most, until a person reviews the evidence and adds it to the map. Each link gets a mechanism score from 0 to 5, and every score shows on the page as a word.
Two readers set each score without seeing each other’s work. Later, in every audit, a third reader who did not score the page re-checks the top scores.
Go deeperWhat the top score requires
A 5 needs a primary research paper (never a review article) showing the molecule meets its target: a measured binding strength, a solved structure with the molecule bound, or an effect that disappears when the target is removed or switched off. Without that paper the score stops at 4. One molecule can hold a 5 on at most two aging drivers; a third would need a measured binding strength and a reviewer’s sign-off, and no page holds three today. Some of the oldest pages were scored before the two-reader and primary-paper rules were written; they are brought up to them as they are audited or re-run.
Go deeperWhat the audit checker looks at
For every score of 4 or 5, working only from papers fetched from PubMed for the check: do the cited papers exist and say what was claimed; are they primary research; has any been retracted; does the map allow the score; and are the citation rules met. It also records how far the concentration used in the lab sits from the blood level a person reaches at a real dose, and flags gaps larger than tenfold. A requirement it cannot confirm counts as not met. The checker can lower a score and can never raise one.
Multiply, then translate into words.
For each of the 13 aging drivers, the mechanism score and the evidence score are multiplied. A strong mechanism with no human evidence stays low. Strong human evidence with no known mechanism stays low too: it can reach Emerging, never Supported. A high rating needs both.
The grid shows every combination. Tap a square to see how it reads.
| 5 | ||||||
|---|---|---|---|---|---|---|
| 4 | ||||||
| 3 | ||||||
| 2 | ||||||
| 1 | ||||||
| 0 | ||||||
| 0 | 1 | 2 | 3 | 4 | 5 |
- StrStrong. A direct, well-established mechanism plus replicated, well-run human evidence, including at least two human studies from the top four levels and one with 100 or more people. Deliberately hard to earn.
- SupSupported. A direct mechanism and at least one randomized trial or pooled analysis, short of the Strong bar. A real signal in people, not a settled one.
- EmEmerging. An early signal, usually animal studies or small human studies. Worth watching; not established.
- ·None. Too weak or too speculative to rate. Very often the research simply has not been done yet, which is not the same as useless.
Any score above zero but below 0.5 is shown as 0.5, so some mechanism plus some evidence always reads at least Emerging. For the two exercise pages, Strong needs a mechanism score of 4 rather than 5; the evidence bar is the same.
Only studies that found an effect add to the evidence score. Studies that found no effect can hold a rating down: when at least two good trials (green or yellow flag) found no effect and they outnumber the good trials that found one, a Strong or Supported rating drops one step and the page carries a Mixed results tag. The number underneath does not change.
Go deeperEverything that can hold a rating down
- A design fact the paper leaves out
- One deduction for each missing fact (people, weeks, blinding). The card says "not stated"; we never call the trial small, short or unblinded unless the paper says so.
- All three design facts measured short
- The trial is excluded from scoring and stays on the page with its label.
- A weak (red-flagged) trial
- It can count as the randomized trial that Supported needs, and toward Strong's study anchors, but it cannot help reach the top two evidence rungs, however many there are.
- Observational or gene-based studies on their own
- The middle rung at most. A well-run trial or pooled analysis has to be among them to go higher.
- Several reports of one trial
- One vote. A follow-up paper on the same trial does not add a second. If its reports disagree on a driver (one found an effect, another did not), the trial counts on neither side.
- More good no-effect trials than good positive ones
- When at least two good trials (green or yellow flag) found no effect and they outnumber the good ones that found an effect, a Strong or Supported word drops one step and the page is tagged Mixed results. If a driver's only positive trials are red-flagged, no-effect trials of any flag count against them the same way. The number underneath is unchanged.
- A failed trial in a disease
- A trial that did not slow a disease (a neurological disease, for example) is shown on the page, but it does not count against an aging driver. Positive disease results still count.
- A change inside one group only
- A before-and-after change in the treated group, with no comparison against the control group, is not counted as a win.
- No difference against an active drug
- A trial that compared the ingredient with an active drug and had no placebo group cannot show "no effect", so it is not counted as a miss.
- A mechanism with no primary binding, structure or knockout paper
- The mechanism score stops at 4. With no pathway on our map, it stops at 2.
- The two mechanism readers disagree
- One point apart: the lower score stands. Further apart: the lower score is used, both views are kept, and the split is flagged for review.
- A top mechanism score on a third aging driver
- One molecule can hold the top mechanism score on at most two drivers. A third would need a measured binding strength and a reviewer's sign-off.
- Human evidence with no known mechanism
- However strong the trials, the rating stops at Emerging: Supported needs a mechanism score of 4 or more, and Strong needs 5.
Following the money, study by study.
Who paid for a study, and whether its authors have ties to a company that sells the product, changes how the study should be read. It is also the part most often buried in a footnote, and it is rarely checked study by study. We check it for every human study on every page, and we show what we found next to the evidence.
- PubMed recordThe paper's conflict-of-interest statement, its grant list and the authors' affiliations.
- PubMed CentralThe full text's funding, conflict and acknowledgment sections, and any sentence saying who supplied the product.
- Europe PMCFunder records, and full text that PubMed Central does not carry.
- OpenAlexThe funders linked to the paper.
- Crossref Funder RegistryThe funders the publisher registered for the paper.
- ClinicalTrials.govWho sponsored the trial, what kind of sponsor it is (company, university, government) and its collaborators.
- Open-access copiesLegal free copies of papers that are otherwise paywalled.
- The publisher's pageFunding and conflict sections that publishers show even when the paper itself is paywalled.
- University library accessThe full paper, for paywalled trials that drive a rating.
- US Patent and Trademark OfficeGranted patents on the ingredient's forms and uses, and whether each is in force.
- European Patent OfficeEuropean patents and their legal status.
- Who publishes the researchWhether one research group is behind most of the published work on the ingredient.
How a label is earned
- 1GatherEvery human study is checked in up to 12 places.
- 2QuoteThe paper's own disclosure is quoted, never paraphrased.
- 3VerifyEach label is traced to the exact field or sentence that states it, or it is dropped.
- 4ConfirmA person confirms every tie a machine finds, and every author matched to a patent.
- 5ShowThe label, the quote and the source sit beside the rating, never inside it.
The rule underneath: every label needs a source that states it, and that goes for “independent” as much as for a conflict. We would rather under-report a tie than accuse a researcher of one we cannot show.
Silence is reported as silence. It is never read as a conflict, and never read as independence.
What it looks like on a page
Each intervention page opens its funding section with a summary: how many studies we could check, how many report company ties, what the company-tied studies found compared with the ones reporting no ties, and whether one research group dominates the field. Then the patents, then every study with its label and the paper's own words.
Of the 24 human studies whose funding and conflicts we could check, 9 report ties to a company that makes or sells this ingredient (38%) and 15 report none. Of the other 6: 4 don't include a conflict-of-interest statement; 2 are still being reviewed.
Results by funding. Studies with company ties that tested an aging driver: 6 positive, 2 mixed, 1 no effect. Studies reporting no ties: 5 positive, 3 mixed, 7 no effect.
Who does the research. One research group is an author on more than half of the published studies of this ingredient.
Patents. 7 granted patents on forms or uses of this ingredient, 4 in force and 3 expired, checked October 2026.
A company tie is a fact to weigh, not proof a result is wrong. Funding never changes the evidence ratings on this page.
Who paid never changes an intervention's rating. A company tie is a fact to weigh, not proof that a result is wrong, so it is shown beside the rating and left for you to weigh.
Go deeperHow a label is verified
A proposed label must point to its evidence in one of three ways: a structured field in the study's record (a grant, a registry sponsor, a stated affiliation), an element of the full text (the funding or conflict section), or a sentence quoted word for word from the paper. A script then re-checks that the field or sentence really is in the source. If it is not, the label is dropped and the study waits for review. “Industry-funded, author ties” needs both halves shown separately: the funding and the author's tie.
A trial registered with a company as its sponsor is labeled company funded. A trial sponsored by a government, university or research network counts toward independent. Our own text never names a company or a researcher; the label describes the relationship, and the paper's quoted words carry the names.
Go deeperPatents and inventors
For each ingredient we search the US and European patent offices and list the granted patents on its forms and uses, with each one's status taken from the office record: in force, expired, or expired because the fees were not paid. When an author of a study is also named as an inventor, the match is shown only after a person confirms it against the author's affiliation or the paper's own disclosure. A matching name alone is never enough.
Go deeperThe two places funding is used by rule
A pooled analysis built almost entirely on trials funded by one manufacturer, or written entirely by its employees, gets a lower quality flag, because that is part of judging how well the analysis was done. And in your stack report, items whose studies are mostly company funded are named, and when they make up three in ten or more of the stack they keep it from the top grade. Nowhere else does funding enter a rating.
What else every page shows you.
Besides who paid, four more things matter to you but do not change how strong the evidence is. They sit next to the rating, each with its sources. Tap a card.
From single items to your whole stack.
The Longevity Intelligence Report takes the ratings of everything you take or do and reads them together: which aging drivers your routine reaches on human evidence, where the biggest gap is, and where several items work the same way. The result is one grade, in words, with the reason under it.
- ExemplaryBroad, well-evidenced coverage of the aging drivers you can act on today, with little waste.
- SolidA real evidence-backed core. Room to widen or tighten, but the foundation is there.
- MixedSome genuine signal, held back by a specific catch that is named under the grade.
- ThinReal coverage, but nothing yet on the strongest human evidence.
- SpeculativeMostly mechanism-stage bets; little human proof yet.
Four of the 13 drivers (telomeres, cellular recycling, stem cells and the thymus) are treated today as the frontier, where today’s evidence cannot meaningfully move them across a routine. They are left out of the breadth count, so they are never counted against you; a Strong rating there still counts in your favor. Every line in your report is fired by a number the engine computed, so a sentence cannot claim what the data does not show.
Go deeperWhat each grade takes
- Exemplary
- At least 8 of the 9 drivers you can act on reached on Strong or Supported evidence; at least 65 percent of items rated Strong or Supported somewhere; at least three drivers rated Strong by items other than exercise; under 12 percent of credit lost to overlap and no three items working the same way; poorly absorbed and mostly company-funded items below the limits in the next panel; exercise present.
- Solid
- A Strong rating somewhere; at least 4 of the 9 drivers reached on Strong or Supported evidence; at least 45 percent of items rated Strong or Supported somewhere; exercise present; no heavy overlap.
- Mixed
- A Strong rating somewhere, but short of the Solid bar, or held at Mixed by no exercise or heavy overlap. The reason is always named.
- Thin
- No driver rated Strong by anything in the routine, but at least one Supported.
- Speculative
- Nothing rated Strong or Supported on any driver.
Go deeperWhat can hold a stack grade down
- No Strong rating anywhere
- If nothing in the routine is rated Strong on any driver, the grade is Thin (or Speculative if nothing is even Supported), whatever else is true.
- Overlap
- Items that work the same way are grouped (15 named groups, such as NAD boosters or prebiotic fibers). In each group the strongest item counts in full and each extra one is discounted. Losing 12 percent or more of the stack's credit to overlap, or three items in one group, keeps the stack from Exemplary; losing 30 percent, or four items in one group, holds it at Mixed.
- Poor absorption
- Three or more red-absorption items taken by mouth, or a quarter or more of the stack, keep the stack from Exemplary. Absorption also shortens an item's coverage bar in your report (a red form counts 30 percent, a yellow form 60 percent), but it does not change breadth or quality.
- Items whose studies are mostly company funded
- An item counts as company leaning when at least half of its studies with readable funding (and at least three of them) were company designed or company funded. When such items make up three in ten or more of the stack, Exemplary is out of reach. They are always named in your report.
- No exercise
- Without resistance or aerobic training the grade is held at Mixed, whatever the supplements.
- Size
- Size by itself neither raises nor lowers the grade. Duplicates are ignored, and a grade needs at least two items.
What a rating means, and what it does not.
- How strong the evidence is that an intervention engages the biology of aging, one driver at a time.
- Computed by the same written rules for every intervention.
- A snapshot of the studies we found, dated, and revised as the evidence changes.
- A finding that it treats, prevents or cures any disease or symptom.
- Medical advice, a dose or a recommendation to take anything.
- Proof that something rated low is useless. Often the research has not been done yet.
We set no doses, prescribe no protocols, and are not your physician. Bring what you find here to a clinician who knows your history.
Checked again, page by page.
Pages are re-audited one at a time by a session that did not build them: trial facts rechecked against PubMed, the literature searched again, every top mechanism score re-checked, claims and cautions built or rebuilt, every sentence re-read against our plain-language rules. When a rule changes, it is re-applied to every page at once. Each page shows when it was last reviewed and its version number.
Every word on a page, and what earns it.
On an intervention page you see words, not scores. The rating words, the Evidence and Verdict words, the Independence word and every randomized-trial flag are computed by code from the data, by the same rules for everything. One tag, Form-dependent, is set by the writer from the product-form evidence. Pooled-analysis flags, claim labels, caution severities, the absorption grade and the safety color are assigned by researchers under fixed definitions; claim labels and caution severities are re-checked by a second reviewer. Pick a set.
The rating beside each of the 13 drivers on an intervention page.
- StrongReplicated human evidence on a direct, well-established mechanism.score 4 or more; mechanism 5 (4 for exercise); two or more human studies from the top four levels that found an effect, one with 100+ people
- SupportedReal human evidence on a direct mechanism, short of Strong.mechanism 4 or more; a randomized trial or pooled analysis that found an effect on this driver
- EmergingAn early signal: often animal or small human studies, or human evidence with no known mechanism.any mechanism and any study that found an effect, short of Supported
- NoneNothing to rate yet. Often the research has not been done.score 0: no mechanism, or no study that found an effect