What AI tells Australian consumers about supplement brands — and what it leaves out
A five-week field study of 240 AI answers to the health questions your customers actually ask. The claims reached them. The warnings, mostly, did not.
The short version
Across 240 AI answers to common Australian health questions, supplement claims travelled to consumers but the specific warnings a product's label is required to carry usually did not. The answers were written largely from overseas pages, often from the model's memory rather than a live source, and frequently credited a stockist's copy rather than the brand. No brand chooses which AI a consumer opens, or what it says. The page that AI reads, though, is within the brand's control — and that is where the fix sits.
The study at a glance: 4 AI systems · 5 ingredients · 5 weeks running · 240 answers logged.
For years, a supplement brand's website was the destination. Every campaign drove people to it, and the compliance work went into making each page right for the person reading it. That has quietly changed. Euromonitor's 2026 survey found the share of Australians who say AI influenced a health decision rose from 8 to 14 per cent in a single year — the fastest growth of any of the sixteen information sources it tracks. When an answer carries a source link, roughly one visit in a hundred follows it.
So the website is no longer the destination. It is the raw material for an answer the consumer reads somewhere else, assembled by a system the brand does not control. To find out what that answer actually contains, I asked four AI systems the same questions every week for five weeks: five ingredients, three kinds of question — learning, buying and safety — logged across 240 answers. The pattern below held consistently enough to act on.
Finding 01 — The claim travels. The warning doesn't.
A claim moves through a long chain before it reaches a consumer. Above the line, every step checks the one before it: a trial finding becomes supplier evidence, becomes an indication the TGA permits. Below the line, everyone restates it — the brand, the stockist, the influencer — and the AI reads all of it and blends it into one answer with the brand's name on it. The claim survives that journey well. The mandatory warning does not.
General caution was common; the specific, required statement was rare. The moment a question shifted from learning to "which one should I buy", disclaimers roughly halved — from 19 answers to 10. The full andrographis safety statement, the allergy risk plus what to do about it, appeared in 4 of 20 learning answers and 6 of 20 buying answers. Mostly it surfaced only when someone asked directly about safety.
- A generic "consult your healthcare professional" is not the warning your product is required to carry. The two are not interchangeable.
- The warning is most likely to go missing at the exact moment it matters most — the point of purchase.
- When pages were opened, 56 of 68 claim-making pages that needed a warning had no version a machine could read near the claim.
Finding 02 — The answer is written from overseas, and often from memory
When an Australian consumer asks what an ingredient does, the answer is mostly built from American pages that have never seen the local TGA indication or the local warning. For one learning question about ashwagandha, the four systems drew on 93 sources over five weeks: just under half were research, about a quarter were US government agencies, the Australian regulator accounted for roughly one per cent, and Australian brands and retailers about two. Across all five ingredients, only six per cent of the links in learning answers pointed to Australian websites.
The systems also often answer from memory — training frozen at a fixed date — before looking anything up, if they look at all. In one test a system admitted it: its first answer came from background knowledge, and only then did it go and find sources. And when credit is given, AI search engines frequently attribute a copy rather than the original. In this industry, the copy is the stockist's or the influencer's version of the claim — which can quietly become "the source".
Finding 03 — A small club of pages, and a lag when the rules change
Brand Medicine International examined a single health topic — diabetes — across ten thousand searches and five thousand websites. The top forty sites, under one per cent, carried roughly three-quarters of the content, and the AI answers drew almost entirely from those forty. The practical reading is encouraging: a brand does not need to fix five thousand pages. It needs to know its own top forty and make sure the claim and the warning are right on them.
The bigger exposure is time. Andrographis came off the listed-medicines framework in September 2026, and the public noticed — search interest hit its highest point of the year. Yet asked to learn about it, 3 answers in 20 mentioned the removal; asked what to buy, 16 in 20; asked plainly whether it was safe — the question that should surface it — 4 in 20. One system did not mention the change once in four weeks, and only did so in the fifth. Consumers assume AI answers are current. Often they are not, and a brand's name can sit inside an answer that is out of date.
Finding 04 — What good looks like, and what you control
Governed, well-structured content can win here. Healthdirect estimates AI platforms cited it 562 million times last year; in the same period its own website sessions fell 18 per cent, which it attributes to AI. It lost traffic but kept its voice — the authority moved into the answer. On supplement questions, that space is open: in this testing, Healthdirect was not cited once.
To test whether structure was the difference, I built two pages for a product that does not exist — same claims, same six warnings. One was laid out like a typical product page. The other placed the warnings in a clearly labelled safety section and in the page description. Given each page, the structured version came back from the AI systems with every statement intact. The standard page did not. Same claims, same warnings, different structure, different answer.
This is the shift that matters: from visual compliance to structural compliance. You do not have to put warnings in your hero banner. You do have to put them where a machine can find them — in plain text, beside the claim, in a clearly labelled section, and in your page description. That is work most brands have not started, and it is entirely within their control.
The one-hour audit
- Ask (15 min). Put one learning and one buying question about your top three products to four AI systems, and ask each for its sources.
- Open (20 min). Read the pages they cite. Is the claim right? Is the warning there, near the claim, in plain text a machine can read?
- Fix (15 min). Close the biggest gap on your own page, and send current copy to the top site republishing you.
- Repeat (10 min). Book a re-run for next quarter, and after any TGA change to your ingredients.
- Remember. A generic caution is not your warning; the specific required statement is the one most likely to go missing.
How CompliGUARD fits
One claim becomes forty versions across pages, listings and posts, and checking every version by hand is where this breaks down. CompliGUARD is compliance infrastructure that reviews marketing content against the TGA · AHPRA · FSANZ · ACCC frameworks — returning a low, medium or high risk assessment against the permitted indication, a plain-English reason, a suggested compliant rewrite, and a dated record of the check. It does not replace the qualified person who decides whether you hold the evidence and takes accountability; it makes sure every version is measured against the same rules, and that a regulator's question is a lookup rather than a three-week reconstruction.
Watch the 15-minute talk → · Speaking & sessions →
Chris Sarandis · Founder, CompliGUARD
Chris spent seven years taking advertising through sign-off as a marketing and brand manager across a multi-brand VDS portfolio. This study is written from the marketer's chair, not the regulator's — it is a field observation of how AI represents health claims, not legal or regulatory advice.
Sources & references
- Original field study: 4 AI systems × 5 ingredients × 3 question types, weekly over 5 weeks (240 answers), 2026. CompliGUARD.
- Euromonitor International, consumer health survey, 2026 — AI influence on Australian health decisions (8% → 14%).
- Brand Medicine International — source-concentration analysis (diabetes topic; top-40 sites).
- Healthdirect Australia — reported AI citation volume and website-session change, 2025–26.
- Therapeutic Goods Advertising Code 2021 — mandatory statements and required warning statements; Permissible Indications Determination.
- FSANZ Food Standards Code — Standard 1.2.7 (nutrition, health and related claims).
- TGA — removal of andrographis from the listed-medicines framework, September 2026; andrographis allergy advisory statement.
CompliGUARD provides compliance and creative marketing insights based on TGA, FSANZ, ACCC and AHPRA guidance as current at the time of review. The Therapeutic Goods Advertising Code is currently under formal review by the TGA (2025–26). CompliGUARD assessments reflect the current Code; guidance may be subject to change. This output does not constitute legal advice.
CompliGUARD — Marketing Velocity. Compliance Certainty. · compliguard.com.au
