When Results Become Rhetoric: Evidence, Authority and Commercial Incentives in Online Health Communication

The modern health influencer rarely presents as merely an advertiser. The more successful model is epistemic: the person sells a programme, supplement, diet or coaching service while simultaneously presenting a theory of why it works. The sales proposition and the scientific proposition become intertwined.

When this model is challenged, one answer appears repeatedly: the method produces results. There are transformations, satisfied clients, improved laboratory values, a large audience and a profitable business. Those facts may all be true. They still do not answer the scientific question being asked.

The central distinction is elementary but routinely lost:

an observed outcome is not the same thing as a causal effect, and a causal effect is not the same thing as a correct mechanistic explanation.

A client may improve. The intervention may genuinely have contributed. The explanation used to market the intervention may nevertheless be wrong.

That is not a semantic objection. It is the reason causal inference exists.

From observation to causal effect

Suppose a person joins a weight-management programme and loses 12 kg over six months. The before-and-after difference is an observation. It describes what happened after the person entered the programme.

The causal question is different. In the potential-outcomes framework, the individual treatment effect is

where (Y_i(1)) is the outcome for person (i) under the intervention and (Y_i(0)) is the outcome that would have occurred over the same period without it.

Only one of those outcomes can be observed. The other is counterfactual.

This is the fundamental problem of causal inference, not an inconvenience invented by academic methodology. If the untreated outcome is unknown, improvement cannot automatically be attributed to the intervention. The person may have improved because of the programme, because of changes made alongside it, because the underlying condition was already changing, or because several causes acted together.

Randomisation is one powerful way to approximate the missing counterfactual at the group level. Carefully designed observational studies use other strategies. Testimonials do neither. They document selected histories.

Hernán and Robins make this distinction explicit throughout Causal Inference: What If: causal questions require a comparison between well-defined interventions, not merely a temporal sequence of exposure followed by outcome.

A programme can work while its explanation is wrong

This problem becomes sharper when interventions contain many components.

Consider a typical online fitness programme. A participant may simultaneously begin resistance training, increase daily walking, eat more protein, reduce alcohol, prepare more meals at home, sleep more consistently, monitor body weight, receive weekly accountability and purchase a supplement recommended by the programme.

If health improves, the programme may deserve some credit as an organisational intervention. It created a structure in which several useful behaviours changed at once.

It does not follow that every element was necessary, nor that the programme's preferred biological explanation was correct.

If the programme also teaches that eliminating one ingredient was the decisive mechanism, the outcome cannot identify that mechanism because the ingredient was not manipulated independently of the other changes. This is an identification problem. The intervention is bundled.

A client can therefore say, truthfully, "this programme worked for me", while the programme owner makes an unsupported leap by saying, "this proves my theory about insulin, seed oils, cortisol, inflammation or detoxification."

Those are different claims.

This distinction matters because commercial health communication often treats successful implementation as retrospective validation of the theory used to sell it. In science, implementation success and mechanism identification require different evidence.

Testimonials are selected observations

The evidential problem with testimonials is not that they are necessarily false. The stronger problem is that their selection process is usually unknown.

A company or influencer may have hundreds or thousands of clients and publish the most visually impressive transformations. Every one of those cases can be authentic. What remains hidden is the distribution from which they were selected.

The relevant probability is not

but something closer to

Those quantities need not be similar.

To estimate the second, one would want to know how many people started, how many completed, how many were lost to follow-up, how outcomes were defined, how many achieved the target, how results varied and whether adverse outcomes occurred. A collection of selected successes contains almost none of that information.

This is not merely a statistical preference. Advertising regulators have recognised the same problem. The U.S. Federal Trade Commission's Endorsement Guides state that exceptional testimonials can be misleading when they imply outcomes that consumers do not generally achieve, and that advertisers should provide information about the results consumers can normally expect.

The general principle is broader than one jurisdiction: truthful anecdotes can still create a false impression of typical effectiveness when the denominator is concealed.

Attrition is part of the outcome

Commercial programmes often report the people who remain.

Scientifically, those who leave matter just as much.

Completion is rarely random. People who remain in a demanding programme may differ systematically from people who stop. They may have more time, more money, fewer adverse effects, greater early success, stronger social support or higher initial motivation. If outcome reporting is restricted to completers, the programme is being evaluated in a selected subgroup partly defined by its ability to tolerate or benefit from the programme.

This is not unique to influencer businesses. Attrition is a central problem in clinical trials, cohort studies and digital-health interventions. The difference is that formal research is expected to report it.

A serious evidence claim should therefore include the fate of the unsuccessful and the missing, not merely the visible successes.

This point also exposes a common asymmetry in persuasive health systems. When someone succeeds, the method receives credit. When someone fails, the explanation shifts to adherence, motivation or individual biology. Sometimes those explanations are valid. But a system in which success is attributed to the intervention and failure is systematically attributed to the participant is protected from disconfirmation by construction.

That is not evidence of robustness. It is a problem of falsifiability.

Improvement can occur without treatment efficacy

Another reason before-and-after results are difficult to interpret is that people tend to seek interventions at unusual moments.

Pain is severe. Weight has reached a personal maximum. Sleep is especially poor. A biomarker has produced an alarming value. Symptoms have flared.

Extreme measurements are often followed by less extreme measurements even without an effective intervention. This is regression to the mean. Barnett, van der Pols and Dobson described it as a ubiquitous problem in repeated measurements, particularly when participants are selected because of extreme baseline values.

Natural history creates a related problem. Many symptoms fluctuate. Musculoskeletal pain changes over time. Skin disorders flare and remit. Functional gastrointestinal symptoms vary. Minor injuries heal. Sleep responds to work schedules, illness and stress.

If treatment starts near a local maximum of symptom severity, subsequent improvement is psychologically compelling. The improvement is real. The attribution remains uncertain.

This is one reason "I tried it and then I got better" cannot carry the same evidential weight as a controlled comparison, even when the speaker is completely sincere.

More anecdotes do not remove selection bias

A frequent response is that the evidence is not one anecdote but hundreds of clients.

Large numbers help with random error. They do not automatically remove systematic error.

A biased sampling mechanism can produce a very large dataset and a very precise estimate of the wrong quantity. If only satisfied clients provide testimonials, increasing the number of testimonials tells us more about satisfied clients, not necessarily about all clients.

This distinction between variance and bias is basic statistical practice but deeply counterintuitive in public discourse. People often treat quantity of testimony as if it transformed the data-generating process.

It does not.

A thousand selected observations remain selected observations.

Popularity is a social signal, not an epistemic one

The second substitution occurs when audience size becomes evidence of correctness.

Popularity can indicate many things: communication skill, consistency, entertainment value, visual presentation, controversy, timing, advertising expenditure, community building, platform optimisation and genuine practical usefulness. Scientific accuracy may contribute. It is only one possible determinant.

A systematic review by Powell and Pring found that social-media influencers can affect health behaviours and outcomes, both positively and negatively. That is evidence that influence matters. It is not evidence that influence tracks scientific validity.

This distinction is important because social platforms create a feedback loop between persuasion and perceived authority. Large audiences generate social proof. Social proof makes claims appear more credible. Greater perceived credibility generates more engagement and further enlarges the audience.

None of those steps performs an experiment on the underlying medical or nutritional claim.

Expertise should also be treated as domain-specific. A successful strength coach may have substantial practical knowledge of training adherence and programme design. That expertise does not automatically extend to oncology, toxicology, endocrinology or causal inference. Competence is multidimensional, and professional success in one coordinate cannot be silently transferred to another.

The same is true inside academia. A professorship is not a universal credential. The relevant question is always: expertise in what?

Commercial success answers a market question

Revenue, sales and client demand are sometimes presented as a harder form of validation. The reasoning is that ineffective programmes would fail commercially.

Markets do not work that way.

Commercial success demonstrates demand under a particular set of incentives. Demand may be generated by efficacy, but also by branding, identity, convenience, community, price, novelty, aspiration and persuasive communication. A useful product can be profitable. A useless product can also be profitable. A harmful product can remain profitable for a long time.

The commercial variable and the biological variable are therefore different objects.

This matters especially when revenue is used defensively. A claim about physiology is challenged, and the answer refers to business growth, audience size or client volume. None of those quantities evaluates the disputed mechanism.

At that point commercial authority is being substituted for scientific argument.

Conflict of interest is not proof of dishonesty

The commercial dimension needs careful treatment because the opposite simplification is also common: if someone sells something, their claims are assumed to be false.

That is equally poor reasoning.

Clinicians are paid. Researchers compete for grants. Universities sell education. Pharmaceutical companies fund trials. Consultants charge for expertise. Financial interests do not determine the truth value of a proposition.

A conflict of interest is better understood as a condition that can alter incentives and therefore the risk of bias.

That distinction is why disclosure matters.

A systematic review by Helou and colleagues examined conflict-of-interest and funding disclosure in health communication on social media and found that reporting was often inadequate. The concern is not that every undisclosed commercial relationship produces a false claim. It is that audiences cannot properly evaluate the information environment when material incentives remain invisible.

The ethical obligation becomes stronger when the same piece of content both educates and sells. A post about a biological mechanism may simultaneously lead to an affiliate code, supplement, laboratory test, coaching programme or subscription. The commercial relationship does not invalidate the evidence, but it changes what should be disclosed and how carefully uncertainty should be communicated.

The ethics of selective evidence

There is a deeper ethical issue than formal disclosure: information asymmetry.

A provider usually knows much more about programme performance than the audience. They know, or could know, how many people enrolled, how many left, how many had no meaningful improvement, how many complained, how outcomes were measured and which cases were selected for promotion.

The audience often sees only the successes.

A communication can therefore be literally truthful while materially misleading. The question is not only whether each testimonial is genuine, but whether the selection gives a reasonable impression of the full outcome distribution.

This is where ethics and statistics meet.

If an organisation wishes to make empirical claims about its results, the natural evidence object is not the best transformation. It is the cohort.

At minimum, serious outcome reporting would include the number enrolled, follow-up period, attrition, outcome definition, baseline distribution, summary of results, variation, missingness and adverse events. If the intervention changed over time, that should also be stated.

The result may be less impressive than a collage of before-and-after photographs. It would be much more informative.

"Science-based" is a methodological claim

The phrase science-based has become a branding category in fitness and wellness. That is not necessarily a problem. It becomes one when the word science is used primarily as an authority signal.

Scientific communication is not defined by the presence of references. A post can contain twenty citations and still misrepresent the literature. The cited studies may involve animals rather than humans, mechanisms rather than outcomes, different doses, different populations or endpoints unrelated to the public claim. A single favourable study may be presented while stronger contradictory evidence is ignored.

The important property is not citation density but auditability.

A claim presented as science-based should make it possible to reconstruct the reasoning:

• What exactly is the claim?
• What population does it concern?
• What outcome is being predicted?
• Which studies bear directly on that outcome?
• What is the magnitude and uncertainty of the effect?
• What evidence points in the opposite direction?
• What assumptions are needed to generalise the result?
• What observation would cause the communicator to revise the claim?

These questions are deliberately uncomfortable. Scientific reasoning is structured around the possibility of being wrong.

Brand authority is often structured around the opposite incentive.

Once a claim becomes associated with a public identity, product line or distinctive philosophy, revision becomes costly. The scientific virtue of updating one's position can conflict directly with the commercial virtue of consistency.

That tension cannot be removed. It can only be managed transparently.

Mechanistic language creates an illusion of explanatory depth

Health communication becomes particularly persuasive when ordinary recommendations are embedded in sophisticated biological vocabulary.

Insulin, cortisol, inflammation, mitochondrial function, autophagy, gut permeability, dopamine and neuroplasticity all describe real phenomena. Mentioning them does not establish that they are the relevant causal pathway for a particular intervention.

Mechanistic plausibility is evidence, but it sits at a different level from demonstrated clinical effect.

A pathway can exist and contribute only minimally to the outcome of interest. Compensatory systems may counteract it. The relevant dose may never be reached in humans. A biomarker may change without altering disease risk. A mechanism observed in vitro may not dominate in a living organism.

The rhetorical danger is that mechanistic detail can make causal uncertainty feel resolved.

It is not.

Scientific authority should be impersonal

The most reliable test of a scientific argument is whether it survives removal of the speaker.

If a claim remains persuasive only because the communicator is famous, visibly fit, commercially successful or surrounded by testimonials, the audience is evaluating a person rather than the proposition.

Science does not eliminate authority. It attempts to make authority contestable.

Methods are described so others can inspect them. Data can be reanalysed. Studies can fail to replicate. Reviews can be challenged. Guidelines can be revised. None of these mechanisms works perfectly, but all are designed around the assumption that confidence must remain answerable to evidence.

That standard does not require an academic title.

A coach, journalist, physician, researcher or influencer can communicate scientifically if their claims remain proportionate to the evidence and open to correction.

Likewise, none of those titles guarantees scientific behaviour.

The useful divide is not academic versus influencer.

It is auditable reasoning versus authority-dependent reasoning.

A more defensible standard for real-world results

Real-world outcomes can contribute meaningfully to evidence. The alternative to testimonial marketing is not to ignore practice.

A coaching business or health programme can learn a great deal from its own data if it treats those data seriously. Complete cohort reporting can identify feasibility, adherence, heterogeneity of response, adverse events and long-term retention. Predefined outcomes can reduce selective reporting. Transparent denominators make claims interpretable. Comparison groups, where feasible, improve causal inference.

Such evidence will still have limitations. Participants self-select. Confounding remains. Measurement may be imperfect. The programme may evolve over time.

But those limitations can be stated.

That is the crucial difference.

Science is not the absence of uncertainty. It is a disciplined way of making uncertainty visible.

The responsibility created by reach

Large audiences do not confer scientific authority. They do create consequences.

An inaccurate claim made privately affects few people. The same claim repeated to hundreds of thousands can alter diets, supplement use, medication decisions and attitudes toward clinical care.

This does not mean every creator should communicate like a regulatory agency. It means confidence should scale with evidence, particularly when claims move from exercise technique into disease, medication, cancer, hormones, mental health or toxicology.

The more consequential the claim, the less defensible it becomes to substitute personal success for evidence.

Reach should increase responsibility, not certainty.

Conclusion

There is nothing trivial about real-world results. They can demonstrate that change is possible, that a programme is feasible, that people value a service and that an intervention deserves further study.

They do not, without additional design and analysis, establish causal efficacy.

Testimonials do not reveal the denominator. Completion data can conceal attrition. Before-and-after change can reflect regression to the mean or natural history. Multi-component programmes do not identify which component caused the outcome. Popularity measures influence. Revenue measures demand. Neither measures biological truth.

Commercial interests are not evidence of dishonesty, but they create incentives that should be disclosed and scrutinised. Expertise deserves respect within its domain, not automatic expansion beyond it.

The most important distinction is therefore not between professionals and influencers, or between academia and business. It is between two standards of justification.

One says: look at who I am, how many people follow me, and the results I can show you.

The other says: here is the claim, here is the denominator, here is the comparison, here are the limitations, and here is what would make me change my mind.

Only the second deserves to be called scientific.

---

References

1. Hernán MA, Robins JM. Causal Inference: What If. Boca Raton: Chapman & Hall/CRC; 2020. https://miguelhernan.org/whatifbook

2. Barnett AG, van der Pols JC, Dobson AJ. Regression to the mean: what it is and how to deal with it. International Journal of Epidemiology. 2005;34(1):215–220. https://doi.org/10.1093/ije/dyh299

3. Powell J, Pring T. The impact of social media influencers on health outcomes: Systematic review. Social Science & Medicine. 2024;340:116472. https://doi.org/10.1016/j.socscimed.2023.116472

4. Helou V, Mouzahem F, Makarem A, et al. Conflict of interest and funding in health communication on social media: a systematic review. BMJ Open. 2023;13:e072258. https://doi.org/10.1136/bmjopen-2023-072258

5. Federal Trade Commission. Advertisement Endorsements: The FTC's Endorsement Guides. https://www.ftc.gov/news-events/topics/truth-advertising/advertisement-endorsements

6. Federal Trade Commission. The FTC's Endorsement Guides: What People Are Asking. https://consumer.ftc.gov/business-guidance/resources/ftcs-endorsement-guides-what-people-are-asking

7. Federal Trade Commission. Health Products Compliance Guidance. https://www.ftc.gov/business-guidance/resources/health-products-compliance-guidance

---

This essay concerns standards of evidence and public health communication. It does not assess any named individual, infer motives, or treat commercial activity as evidence of dishonesty. The question throughout is narrower: what conclusions can legitimately be drawn from the evidence being presented?