Five checks separate health-product claims from usable evidence

An April 2026 BMJ Open audit rated 49.6 percent of 250 responses from five consumer chatbots as problematic when tested with misinformation-prone health prompts. Consumers can reduce that risk by checking the exact claim against regulator records, matching it to human evidence and opening every cited source.

A usable check has five parts: capture what the seller actually promises, identify the regulator for the market where the product is sold, trace the evidence to its full source, search for missing or contrary information, and verify any leads supplied by artificial intelligence (AI). No approval mark, product register or advertising rule applies worldwide, so the country and product category belong at the start of the check.

Urgent symptoms come before claim checking

The United Kingdom National Health Service lists sudden swelling of the lips, mouth, throat or tongue, trouble breathing or swallowing, sudden confusion and fainting as possible signs of anaphylaxis. If any of these symptoms begin after a food, supplement or other health product, call the local emergency service immediately instead of spending time investigating the advertisement.

Pregnancy and breastfeeding

The US National Center for Complementary and Integrative Health says many supplements have not been tested in pregnant or breastfeeding people. Evidence from other adults cannot establish safety in pregnancy or breastfeeding, so a clinician should assess the exact product before use.

Children

The same US guidance says many supplements have not been tested in children. Adult studies and label instructions cannot establish pediatric safety, so a clinician or pharmacist should assess the exact ingredients and intended use.

Chronic conditions, surgery and regular medicines

The US center says supplements may interact with medicines, pose added risks with some medical conditions or affect surgery. People in these groups should give a clinician or pharmacist the full product label and a list of regular medicines before use.

Step one: write down the claim before checking it

Save the exact words, images and context rather than reducing an advertisement to a vague phrase such as “good for immunity.” The US Federal Trade Commission's 2022 guidance says an advertisement's overall message can include express statements, product names, pictures, graphs and implied claims. That is a rule for US advertising, but the recording method prevents a seller's visual or indirect promise from disappearing during a fact-check.

Record the product name, manufacturer, full formulation, route, amount, advertised population, promised outcome, time frame, country of sale and date viewed. A study cannot support a claim until its tested product and outcome match those details.

Disease words also change the regulatory question. The US Food and Drug Administration says a product intended to treat, diagnose, cure or prevent disease is a drug even if its label calls it a dietary supplement, while the agency does not approve dietary supplements for safety and effectiveness before they reach the US market. This distinction is specific to the United States; it is not a shortcut for deciding how another country classifies the same product.

Step two: find the regulator and read what its record means

Start with the authority responsible for that product category in the place where it is sold. Search the exact product or claim, then compare the manufacturer or sponsor, ingredients, record number, status and date with the package. A number printed on a label is not verified until those details match the official record.

Registers answer different questions. The European Commission says its public register lists permitted nutrition claims and both authorized and non-authorized health claims for foods in the European Union. It is a claims register, not proof that every food bearing an allowed phrase has been tested as a finished product.

Australia's Therapeutic Goods Administration says all medicines supplied in Australia must appear in the Australian Register of Therapeutic Goods, but AUST L medicines are listed without a premarket efficacy assessment while AUST L(A) and AUST R products receive one. A register entry therefore needs to be read with its category and assessment route rather than treated as a universal efficacy certificate.

The US, European Union and Australian examples are not a complete global directory. If the seller does not identify the country where a registration or claim is valid, ask for the authority's full name and a direct record link. APPI News could not verify the corresponding databases and claim rules in every market at the time of writing.

Step three: trace “research-backed” to the full source

The US National Institutes of Health Office of Dietary Supplements advises checking who runs and funds a health website, its purpose, its original sources, the credentials of its reviewers and when the information was updated. The same guide says medical facts should carry retrievable references and that testimonials are not evidence-based.

A citation should lead to the full paper or an official report, not only a press release, sales page, influencer post or abstract. Confirm the title, authors, journal, year and digital object identifier, then check whether the linked source makes the claim attributed to it. Read the funding and conflict-of-interest declarations before weighing the result.

The evidence must match the finished product, route, amount, population, comparison, duration and outcome in the advertisement. A cell experiment cannot establish a benefit in people, an animal experiment cannot establish a consumer outcome, and a study of one ingredient cannot automatically support a mixture with different ingredients.

The US Federal Trade Commission generally treats randomized controlled human trials as the most reliable support for health-benefit claims, says independent replication adds weight and warns that research quality matters more than the number of weak studies. Its guidance also distinguishes an observational association from evidence of causation and calls for the whole body of relevant evidence to be considered.

Step four: look for evidence a sales page leaves out

A favorable paper is one part of a research record. Search for systematic reviews, later trials, corrections, retractions, adverse-event reports and regulator warnings using the same ingredient and product terms. Look for null or harmful findings, not only papers cited by the seller.

Cochrane describes systematic reviews as a structured examination of all research that meets predefined criteria for a question, making them more reliable than an individual study for assessing whether an intervention or test works. A review still needs to match the product and outcome, and uncertainty in its findings cannot be converted into a positive marketing claim.

For a clinical trial, search a registry by title, investigator, intervention and registration number. The World Health Organization says prospective registration helps reveal publication bias and selective reporting and directs users to primary registries in its registry network. A registry record shows what researchers planned; it does not show that a trial was completed or that the result favored the product.

Compare the registered primary outcome, enrollment target and time frame with the published report. A changed outcome, smaller sample or missing result needs an explanation, while a matching record improves transparency without proving that the study was unbiased.

Step five: use AI for leads, then verify every source

The BMJ Open audit tested free versions of Gemini, DeepSeek, Meta AI, ChatGPT and Grok in February 2025. Researchers used 50 prompts across cancer, vaccines, stem cells, nutrition and athletic performance for each chatbot, producing 250 responses. The prompts were designed to push the systems toward misinformation or contraindicated advice, so the study tested a demanding scenario rather than every routine health question.

The researchers rated 30 percent of responses somewhat problematic and 19.6 percent highly problematic, found a median reference-completeness score of 40 percent and reported that fabricated citations prevented every chatbot from producing a fully accurate reference list. Those results apply to the named versions and test design; they do not rank current models, which can change after release.

AI can still organize the exact claim, generate neutral search terms and suggest a regulator, trial registry or paper to inspect. Open each suggested page independently and confirm its title, authors, date, identifier and conclusion. A plausible citation that cannot be found, or a real paper that does not support the sentence, should be treated as a failed check.

A short record makes the check reproducible

Keep the original claim and screenshot, the market and product category, the regulator record, the strongest matching human evidence and any safety warning or unresolved gap. Add the date of the check because advertisements, database entries, warnings and AI systems can change.

The result does not have to be a simple “true” or “false.” “Authorized wording in one market,” “supported only for a different formulation,” “one small trial with no independent replication” and “no matching source found” are more precise findings, and they preserve the limits of what was actually verified.

Frequently asked questions

Does a regulator number prove that a health product works?
No. Its meaning depends on the jurisdiction, product category and assessment route. Check the official record for what was reviewed and whether it covers efficacy, legal supply, manufacturing, an ingredient or only specified claim wording.

Does publication in a journal prove a marketing claim?
No. Publication does not by itself establish that the study was well designed or that it tested the same formulation, route, population and outcome. Read the full methods and results, then look for independent studies and systematic reviews.

Are cell and animal studies useless?
No. They can identify mechanisms and support further research, but they do not establish that a consumer product produces a health benefit in people. A human-benefit claim needs evidence measuring that outcome in an appropriate human population.

Can an AI chatbot complete the five checks?
It can propose search terms and possible sources, but every citation and claim still needs manual verification. The 2026 audit found poor reference completeness and fabricated citations even when answers were written confidently.

Who can help when the evidence or safety question remains unclear?
The regulator responsible for the product in the country of sale can clarify records, warnings and reporting routes. A clinician or pharmacist can assess personal factors such as pregnancy, childhood, chronic conditions, surgery and medicine interactions.