How to read a supplement study (without a science degree)

Supplement labels and marketing are written by the seller, not by the people who ran the trials. In the US, FDA does not have the authority to approve dietary supplements before they are marketed [17]. That leaves the published research as the main public check on whether an ingredient does what people say it does.

This guide walks through the questions our editors ask of every study before it counts toward a grade. Each example comes from one of our compound pages, so you can see the same study in context. The grades themselves are explained on How we grade evidence.

You do not need statistics training to get most of the way. Five questions cover the main points:

1. Who took part? Healthy adults, older adults, athletes or people with a medical condition.
2. Compared with what? A placebo, another product, or nothing at all.
3. How many people, and for how long? A single dose, a few weeks, or a year.
4. What was measured? A test score or blood measure, or how people said they felt.
5. Who paid, and who ran it? A university grant, the ingredient maker, or a mix.

A well-reported trial answers all five in its methods section and its funding and competing-interests statements. The CONSORT 2025 reporting guideline lists a 30-item checklist of what a randomized trial report should include, plus a diagram showing how participants moved through the trial [7]. When a paper leaves these out, that gap is itself worth noting.

The standard design for a single trial is a randomized controlled trial: people are assigned by chance to the supplement or a comparison, usually a placebo that looks the same. "Double-blind" means neither participants nor researchers know who got which.

The placebo group shows how much people change for reasons other than the ingredient. In a randomized, double-blind trial of 30 healthy adults with moderate stress, Perceived Stress Scale scores fell by 17.98% after 28 days with 400 mg/day of AlphaWave L-theanine and by 17.88% with placebo [2]. The trial was funded by Ethical Naturals, which employed one of its authors [2]. There was no statistically significant difference between the groups in change in salivary cortisol [2]. Without the placebo arm, the first figure would look like a clear result.

An open-label study, where everyone knows what they are taking, cannot separate the ingredient from expectation. As of November 11, 2026, the only human mood data we found for uridine is an open-label case series of seven adolescents with bipolar disorder (a clinical population) given uridine for 6 weeks, in which uridine was associated with lower symptom scores [8]. With no comparison group, that result tells us a trial might be worth running, not that uridine works. Results from clinical groups may not apply to healthy adults. The full record is on our uridine page.

A result holds for the people tested. In a Cochrane review of 35 trials, cocoa flavanol products lowered blood pressure slightly compared with control in 1,804 mainly healthy adults, but in people whose blood pressure was already normal there was no significant difference [9]. The review authors note that the test for this subgroup difference was of borderline significance and needs confirming [9].

So "cocoa lowers blood pressure" and "cocoa lowered blood pressure in this trial population" are different statements. Our grades are set for healthy adults, and results that come only from clinical or older groups cap a grade at a lower level, as our How we grade evidence page explains.

A statistically significant result means a difference this large would be unlikely if the supplement had no effect. It does not say whether the difference is large enough to notice. For that, look for the effect size and its confidence interval, the range of values the data are compatible with.

The cocoa review gives both. Across 40 treatment comparisons, systolic blood pressure was 1.76 mmHg lower than control, with a 95% confidence interval of -3.09 to -0.43 mmHg [9]. The authors describe this as a small (2 mmHg) effect in mainly healthy adults in the short term [9]. Many trials report a standardized mean difference instead, a unitless score that lets reviewers pool studies that used different questionnaires.

Size matters in another way too. In a study of 93 meta-analyses covering 735 randomized trials of treatments with yes-or-no outcomes, treatment effects in trials with fewer than 50 patients were, on average, 48% larger than in trials with 1,000 patients or more [3]. Two trials on this page, including the Ethical Naturals-funded L-theanine trial, enrolled 30 and 36 people and measured scores rather than yes-or-no outcomes [2][10]. The 48% figure does not carry over to them directly, but a striking result from a trial that size still deserves a second look.

A meta-analysis pools the results of several trials. That usually gives a more stable answer, but only if the trials were measuring the same thing in similar people.

Check whether the pooled trials agreed. A 2022 meta-analysis of 12 randomized trials of ashwagandha (1,002 participants aged 25 to 48) found lower stress scores than placebo, with a standardized mean difference of -1.75 (95% CI -2.29 to -1.22) [11]. The same analysis reported an I2 of 83.1% for stress, meaning the trials varied a great deal, and the authors rated the certainty of the evidence as low [11]. That variation between trials is called heterogeneity, and high heterogeneity is a reason for caution even when the pooled number looks impressive. More detail, including what sources report on safety, is on our ashwagandha page.

Pre-registration means the researchers publish their main outcome and analysis plan before the trial starts, so it is harder to pick whichever result looked best later, and readers can check the reported outcome against the registered one. Among 55 large heart-health trials of drugs or dietary supplements funded by the US National Heart, Lung, and Blood Institute, 17 of 30 (57%) published before 2000 reported a significant benefit on the primary outcome, compared with 2 of 25 (8%) published after 2000 [13]. Registration in clinicaltrials.gov was strongly associated with the shift toward null findings [13].

In that set of trials, registration went with less exciting results. In a pre-registered trial, a switch to a better-looking main outcome can be spotted by comparing the paper with the registration, and that is a point in its favor. Two pre-registered, double-blind crossover trials, in which each person tries every condition in turn, tested single doses of cocoa flavanols in young adults: one in 48 university students, using cocoa powders donated by Barry Callebaut, found faster visual search but no effect on temporal attention [14], and one in 36 healthy adults found no significant effect of 415 mg or 623 mg on cognitive control or response inhibition [10].

Next to each grade on our pages you may see one or more modifiers. The modifier itself does not change the letter, though our rubric withholds an A when no independent trial or meta-analysis supports the outcome. Modifiers tell you what kind of doubt still hangs over the evidence. Where a section shows a badge, it shows the modifiers that apply today, because grades are kept up to date as new studies arrive.

Industry-funded

This modifier appears when most of the trials behind a grade were funded or run by a manufacturer, an ingredient supplier or their employees, or used study product the company donated. In a Cochrane review of 75 studies, industry-sponsored drug and device studies more often had favorable efficacy results (risk ratio 1.27, 95% CI 1.17 to 1.37, pooled from 25 of the 75 papers) [1]. The authors' analyses suggested an industry bias that standard risk-of-bias assessments cannot explain [1], which is why we label funding separately from our risk of bias checks.

Industry-funded does not mean wrong. Funded trials can be well designed: the Ethical Naturals-funded L-theanine stress trial above was double-blind and registered (NCT05808595) [2]. It means a reader should look for independent trials that point the same way. Look for the funding and competing-interests statements near the end of a paper; our reference lists give the funding we found for each study, or say that none was stated. Our own commercial relationships are listed on How we make money, and they never change a grade.

Dose gap

This modifier appears when typical products contain less of an ingredient than the doses used in studies. A trial result applies to the dose tested, so a product with a smaller amount sits outside what the trial tested.

To check it yourself, find the dose in the methods section and compare it with the label. Doses can vary widely even within one ingredient: cocoa trials in the Cochrane review provided 30 to 1218 mg of flavanols [9]. Trials that test several doses can show a dose-response pattern, where the effect changes with the amount.

Short-term only

This modifier appears when the trials supporting a grade were all short, from a single dose up to a few months. The trials in the Cochrane review of cocoa and blood pressure ran 2 to 18 weeks, and the review authors called for long-term trials to see whether cocoa affects cardiovascular events and to check for adverse effects of long-term use [9]. Short trials cannot show whether an effect lasts, fades as the body adapts, or brings side effects that take months to appear. Single-dose crossover trials like the cocoa attention studies answer a narrower question again: what happens in the hours after one dose. The full picture, including the longer Mars Edge-supported COSMOS trials in older adults, is on our cocoa flavanols page.

Single research group

This modifier appears when the supporting trials all come from one laboratory or one overlapping set of authors. Results from a single group need independent replication. When the Open Science Collaboration repeated 100 psychology studies, 97% of the originals had reported statistically significant results, but only 36% of the replications did, and replication effects were half the size of the original effects [4].

The same pattern turns up in health research. In a Cochrane review of trials that compared placebo with no treatment, four similarly designed acupuncture trials by an overlapping group of authors reported large placebo effects, while three other pain trials reported low or no effect [15]. In our own records, both pre-registered cocoa crossover trials above came from the Experimental Psychology department at the University of Groningen and share an author [14][10], and the 2018 trial used cocoa powders donated by Barry Callebaut [14]. That is not a flaw in either trial, but agreement between them is not the same as confirmation by another team. Pre-registration helps, but it does not replace a second group.

Subjective outcome only

This modifier appears when the graded outcome is one people rate themselves, such as stress, mood or sleep quality questionnaires, and objective measures do not consistently confirm it. Self-ratings are real data, but they can be more open to expectation than objective measures.

The evidence on how much this matters is mixed. A combined analysis of 146 meta-analyses (1,346 trials) found that in trials with subjective outcomes, effect estimates were exaggerated when blinding was lacking, with little evidence of bias for objective outcomes [5]. A later meta-epidemiological study, MetaBLIND, found no evidence of an average difference between trials with and without blinding, and its authors suggested repeating the study and keeping blinding as a methodological safeguard [6]. A Cochrane review of trials that compared placebo with no treatment concluded that placebo can influence patient-reported outcomes in certain settings, though this is hard to separate from biased reporting [15].

L-theanine shows what this looks like in practice. A 2025 systematic review and meta-analysis (19 articles, 897 participants of all ages and health status, 18 pooled) found improvements in subjective sleep onset, daytime dysfunction and overall subjective sleep quality, with a standardized mean difference of 0.43 (95% CI 0.04 to 0.83) for sleep quality [16]. The authors also noted a lack of studies on pure L-theanine [16]. Our L-theanine page covers the individual sleep trials and who funded them, and the Sleep and Mood and stress hubs compare it with other options.

Some kinds of evidence cannot support a claim about what a supplement does in people, however often they are repeated:

• Cell and animal studies. They can suggest how something might work. On our pages they are labelled as mechanism and never count as evidence of an effect in people; a grade built only on them is a D.
• No comparison group. Case series and before-and-after studies, like the uridine series above, show change over time but not its cause [8].
• A press release or conference abstract with no full paper behind it. A claim with no published paper cannot be checked. Even the abstract of a published paper is a short summary and can leave out dropouts, secondary outcomes and funding details, so our reference lists say whether we read the full text or only the abstract, and say when funding was not stated in what we read.
• Conclusions that outrun the results. In the Ethical Naturals-funded L-theanine stress trial, the conclusion reported lower perceived stress with L-theanine, while the results showed almost the same fall with placebo [2]. Read the results section, not just the conclusion.
• Retracted papers. Papers can be withdrawn after publication. We check every paper we cite against PubMed and Crossref retraction records, and a retracted paper is never cited as support.

Every graded outcome on our pages is set by the rubric on How we grade evidence, which applies the questions above. Review authors often rate the certainty of a body of evidence with a separate system; one common one is GRADE, which its authors described in 2008 as increasingly being adopted by organizations worldwide [12]. The evidence matrix shows every compound and outcome side by side, with the same modifiers.

What is the difference between a placebo-controlled trial and an open-label study?

In a placebo-controlled trial, one group gets the supplement and another gets an identical-looking placebo, so the placebo group shows how much people change without the active ingredient. In an open-label study everyone knows what they are taking, so the supplement's effect cannot be separated from expectation. A study with no comparison group at all, like the uridine case series of seven adolescents with bipolar disorder (a clinical population), cannot separate it from ordinary change over time either [8].

Does industry funding mean a study is wrong?

No. Industry-funded trials can be well designed, but a Cochrane review that pooled 25 papers on this question found that industry-sponsored drug and device studies more often reported favorable efficacy results [1]. We name the funder where we first cite such a study in each section, and look for independent trials that agree.

Why does a small study count for less than a large one?

Small trials tend to report larger effects. Across 93 meta-analyses of treatment trials with yes-or-no outcomes, trials with fewer than 50 patients reported effects 48% larger on average than trials with 1,000 patients or more [3]. A small trial is a reason to look for a larger one, not proof on its own.

What does subjective outcome only mean on a grade?

It means the graded outcome is a self-rating, such as a stress or sleep-quality questionnaire, and objective measures do not consistently confirm it. In one large analysis, effects on subjective outcomes were exaggerated in trials without blinding [5], though a later analysis found no evidence of an average difference between trials with and without blinding [6]. The modifier tells readers that the result rests mainly on how people said they felt.