A claim review fails for one of two reasons, and most teams treat them as the same reason. Either the claim is defensible but unproven in the form you wrote it, so it needs evidence attached before it can run. Or the claim may not be made at all: not to this audience, not about this condition, not in any wording, and not with any evidence behind it.
Two failure modes, one word apart
A claim review fails for one of two reasons, and most teams treat them as the same reason. Either the claim is defensible but unproven in the form you wrote it, so it needs evidence attached before it can run. Or the claim may not be made at all: not to this audience, not about this condition, not in any wording, and not with any evidence behind it.
Only the first has a fix that looks like work. Route it to medical affairs, attach the study, narrow the wording to what the study actually measured, and the line runs. The second has no fix of that shape. A stronger trial does not make a prescription-only medicine advertisable to the public. A larger sample does not make 'cures diabetes' publishable. When the bar is audience or subject matter rather than proof, evidence is the wrong tool.
So the first question on a draft is not whether it is true. It is which of the two problems you are looking at. That decides whether the note back to the writer is 'get me the data' or 'write a different sentence'.
Which claims are prohibited outright?
These are the lines no amount of substantiation rescues. They sit in statute, chiefly the schedules that make certain medicines prescription-only and the Drugs and Magic Remedies (Objectionable Advertisements) Act 1954.
Notice what they have in common. Not one of them turns on how strong your evidence is. They turn on who is reading, which condition is named, and whether an approval actually exists.
Prescription-only medicines advertised to the public. Schedule H, H1 and X products may not be advertised to the public at all. Content about them can address registered medical practitioners, and that is the only audience it can address.
Cure, prevention or treatment claims for a scheduled condition. The Act lists conditions that may not be claimed as treatable in advertising to the public, including cancer, diabetes, heart disease, HIV, kidney and liver disorders, infertility and obesity.
Absolute safety or efficacy. 'No side effects', '100% safe', 'completely safe', 'works for everyone'. No dataset makes these accurate, which is why they are a prohibition and not an evidence request.
Approval that was not granted for this product and this indication. 'CDSCO approved', 'FDA approved' and 'clinically approved' worn as a general badge. A real approval is narrower than a badge: it covers one product for one stated use.
Content positioned as a replacement for professional advice, diagnosis or prescription.
A dosage, course or regimen for a prescription medicine stated in public-facing copy, including inside an FAQ answer that reads as helpful.
Which claims only need evidence?
The second category is the ordinary work of a claims process. The claim can run once the proof is attached and the wording matches the proof.
The process rule underneath all of them is simple to state and unpopular to enforce: every efficacy, safety, comparative or regulatory-status claim carries a source attachment. If nobody can name the source on request, the claim is not ready, whatever the launch date says.
Comparative and superiority claims. 'More effective than', 'fastest acting', 'most effective', 'best in class'. These need published, comparable clinical evidence. Comparable is the load-bearing word: two trials with different populations and different endpoints do not compare, however neatly their headline numbers sit side by side.
Any cited clinical result. Give the study, the population, the sample size and the endpoint rather than the headline figure alone. A figure with no study around it is decoration, not evidence.
Patient testimonials and before-and-after imagery. Permitted, but they must not imply a typical or expected outcome. One person's result presented as what the reader should expect is the failure mode.
Medical-device claims. These must match the approved intended use. A cleared indication is a boundary, not a starting point to extend from in marketing copy.
Urgency and fear framing on a health decision. 'Silent killer', 'before it is too late', 'limited stock'. This is the copy that turns a defensible product page into a complaint.
The audience is part of the claim
This is the rule that catches experienced marketers, because it has no equivalent in most other categories. Elsewhere a claim is either substantiated or it is not, and the reader does not change the answer. Here the same sentence can be lawful addressed to a prescriber and unlawful addressed to a patient. The words did not change. The audience did.
That makes audience a drafting decision, not a distribution decision. Settle it before the brief rather than after the copy exists. Practitioner content belongs behind a declared gate and should say on the page who it is for. Consumer content should never be a route to obtaining a prescription-only product, however indirectly it is phrased.
It also means an approval that stops at 'is this accurate' has skipped a gate. Two reviewers can read the same page and disagree, not because they weigh the evidence differently, but because they assumed different readers.
What happens when a model rewrites your sentence?
Everything above governs what you publish. An AI answer engine does something you do not control: it reads your page, compresses it, merges it with other sources, and writes a new sentence with your brand attached to it.
Qualifiers go first, because they are the least quotable part of any sentence. A carefully hedged line about supporting a metabolic process can come back much closer to a treatment claim. Sources merge, so your page, an old release and a forum thread can be summarised into one paragraph that reads as your position. Audience flattens too: a page written for prescribers, if a crawler can reach it, can be read into an answer a patient asked for.
There are now two questions, and they need different evidence. Whether your copy is clean is answered by a claims process and a legal review. Whether the answer about you is clean can only be answered by asking the engines and reading what comes back. A brand can pass the first and fail the second, and the second is the one the patient sees.
You cannot edit the answer. You can change the sources it is built from, and you can find out what it says today, which is the part most regulated teams have never measured.
The compliant version is usually the more citable one
The additions these rules demand are the same additions that make a page worth citing. State recognised side effects, contraindications and risks alongside the benefit rather than the benefit alone. Name the generic, the INN, next to the brand. Give the study, the population, the sample size and the endpoint instead of a headline figure. Carry a line telling the reader to consult a qualified practitioner.
Read that list again as an engine would. Every item makes the page more specific, more attributable, more extractable. A model choosing a source on a condition is picking between a page that says 'clinically proven' and a page that names the trial, the population and the endpoint. The second survives paraphrase, because there is something concrete to carry across.
This is why Answer Engine Optimization and a claims process collide less often than teams expect. Vague superlatives are both the least defensible copy and the least quotable copy. Specific, sourced, hedged-where-it-should-be writing is what a reviewer approves and what a model repeats.
Keep a claim register, not a banned-words list
Most teams police claims with a list of forbidden phrases. It catches 'cures diabetes' and misses 'finally free of your insulin', which means the same thing to a reader and to a model.
A register works better. Three columns: the claim as it will appear in copy, the evidence behind it, and the audience it is allowed to reach. Prohibited claims never enter the register, which is the point of having one. Evidence claims enter with their source attached, or wait outside until they have one.
Then check it against reality on a cadence. Your pages are governed by the register. The paraphrase in an AI answer is not, and it moves when the sources around you move. Legal review stays a separate track and stays required; a register makes that review faster, it does not stand in for it.