Natural products, synthetic advice: What does AI say about your supplement brand?

The website used to be the destination. Now it's raw material.

When I worked brand-side, everything pointed to the website. People landed, read the page, and made up their own minds. In 2026 that's changed: your buyer asks an AI model first, and often never arrives at your page at all.

Euromonitor's 2026 survey found the share of Australians who say generative AI influenced a health decision rose from 8% to 13.9% in a single year — a 74% jump, the fastest-growing of the sixteen sources they track, now ahead of social media and nearly three times as influential as influencers. Trust is climbing with it: 37% now rate AI a trustworthy source of health information, closing in on brand advertising at 41%. And when people see an AI summary, Pew found only 1% click a link inside it to check the source.

So your customers are trusting AI's version of your brand almost as much as your own advertising — and they rarely verify it against your page.

How a claim drifts

Up to the label, evidence is held carefully: research becomes supplier evidence, which becomes a TGA-approved indication. Past the label, everyone rephrases it — the brand, the stockist, the practitioner, the wellness influencer — each bound by the same Advertising Code, each a little further from the original. Then AI reads and blends those versions and tells the consumer "this product helps with anxiety." No source. No warning.

The parts least likely to survive that process are exactly the ones that protect the consumer: the citation, the disclaimer, the ingredient warning, the line to see a health professional. Published research found disclaimer presence in AI health outputs fell from 26.3% in 2022 to under 1% in 2025.

So I tested it

I asked four AI systems — ChatGPT, Claude, Gemini and Perplexity — the same questions every week for five weeks, across five ingredients chosen to span the regulatory spectrum, with three kinds of question: learning, buying and safety. A few things stood out:

• Ashwagandha ("does it help with stress?") — all four said yes, but most sources were overseas, a quarter from US government sites alone. The Australian regulator appeared in one answer out of twenty. These answers aren't written for Australia.
• Andrographis — the TGA removed it from the listed-medicines framework in September over severe allergic-reaction risk. Asked bluntly "is it safe?", only 20% of answers mentioned the removal. Gemini didn't mention it at all, across every question, for four weeks. People assume AI answers are current. Often they aren't.
• Who gets quoted — Brand Medicine International found that in a single health category, 0.8% of publishers produced 75% of the content, and AI cites from the top of that ranking. You don't need to police 5,000 websites. You need to know your top 5, 10 or 20 — and make sure your claims and warnings are right on them.

What you can actually do about it

You can't control which AI model your customer uses. You can control what those models find when they go looking. That's the whole point: the quality of what our industry publishes is what these systems increasingly distribute.

The homework I gave the room: ask four AI systems, logged out, a learning and a buying question about your top three products. Note which pages they cite — yours versus third-party stockists — and check whether the warning is visible, current, and sitting next to the claim.

That's exactly what our one-hour AI visibility audit does with you, end to end.