AI product sheets for supplements: GEO yes, claims no (and the AI Act)
“Can I have ChatGPT write my product sheets?” Yes, as a draft. No, as unfiltered publishing.
On a regulated catalog (supplements, but the pattern holds for anything else that sits under lock and key), the model does more than give you “nicer copy”. It moves your claims. It stretches them. It softens them. It invents benefits that sound natural in a chat and become a problem on the product page. I see it up close on mid-size stores I have worked with for years: the temptation is to feed the whole catalog to one prompt and call it GEO. It is not.
I work on Woo and Shopify, on feeds, product pages, and tracking. I am not selling a magic tool here. I am trying to pull apart three layers that most people blend: ranking and citations by AI models, compliance on claims, and the transparency duties in the AI Act. They are different things. Treat them as one “just use AI” and you sign up for risks you were not looking for.
What goes wrong (and why it is not a style issue)
The standard advice goes like this: “write as if you were talking to a customer, be empathetic, tell the benefits.” On a supplement, “benefit” is a loaded word. Left off the leash, the model slides into cures, treats, prevents, solves. Sometimes it does so openly. Sometimes it uses synonyms that slip past a quick read and do not survive a serious check.
This is not an SEO detail. It is the piece that exposes you to customers, marketplaces, ads, and, for ChatGPT ads, policies that keep health copy on a short leash. I already covered this in the piece on the three ChatGPT tickets: a model that paraphrases “cures” is not doing GEO for you. It is creating a problem.
A concrete example, no names. A catalog with dozens of product sheets. A generic prompt: “rewrite for SEO and to convince people to buy.” The output reads like a wellness brochure. No up-to-date structured attributes. Ingredients fine in the back office, wrong tone on the front. Feed and page drift apart. The agent reading your data picks up one thing, the human buying reads another. That is not positioning. It is debt.
GEO is not inflated copy
What does a model or an agent read when it has to cite you or pull something out of a catalog? A clean title, price, stock, shipping, attributes, ingredients, warnings, a product page consistent with the feed. Not the animated H1. Not the emotional paragraph three screens long.
The distinction I care about, compared to the LinkedIn advice: here GEO means making data readable and true. AI can help you write the narrative part after you have closed what can be said. Not before. If you flip the order, you optimize the violation.
On Woo or Shopify the dirty work is the same: structured fields, product schema, Google and Meta feeds that do not invent claims, an llms.txt that does not start from the cookie policy. AI prose is the last mile, not the first.
The process I use (no guru tools)
Opinion from projects, not from a white paper.
First: a list of allowed and banned claims. The prompt does not write it. It is signed by whoever holds scientific or legal responsibility for the catalog, or at least by a human source who knows what can go on the label and on the page. Without that list, any “AI copywriter” is a risk generator.
Then: the page skeleton. What you do not touch (name, ingredients, doses, warnings, link to the leaflet if there is one). What can be rewritten (kicker, short description, product FAQ within the claims). AI works only on the narrative piece, with explicit constraints: do not use these words, do not promise these outcomes, do not add ingredients that are not in the product record.
Then: the draft. One SKU or a small batch, not the whole catalog overnight. Human review line by line on the claims, not just spelling. If the review is “looks fine”, you did not review. You approved by gut feeling.
Then: alignment. What you publish on the page has to carry over to the feed and, if needed, to ad landing pages. I have seen mid-size stacks where the on-site page says one thing and Doofinder or the feed show another. AI does not fix that. It makes it worse if it generates a third variant.
Finally: top SKUs before the long tail. If the catalog is large, start where the traffic and the reputational risk are. Rewriting everything “because AI is free” is the fastest way to multiply errors.
AI Act: transparency without panic
This needs clarity, not alarm. This is not legal advice.
The AI Act (Art. 50) provides, among other things, transparency duties for certain content generated or manipulated by AI. For text published to inform the public on matters of public interest, the deployer has to disclose that the text is artificial if there is no substantial human review and no editorial control with clear responsibility. If a person reviews the substance, can change or block it, and answers for the publication, that disclosure typically does not kick in the same way.
Translated for a store: if you generate 400 product sheets and put them online without a competent human eye on the claims, you are playing two games at once (product compliance plus AI transparency). If you use AI as a stenographer and you (or your team) sign off every piece after a real check, you are in the “draft + editor” model, which makes sense even without the regulation.
Two points I see ignored.
Review is not Word’s spellchecker. Checking commas and leaving “promotes healing” in place is not editorial control. On regulated catalogs the substance is the claim.
Second: the machine-readable marking duty on model providers (whoever trains and offers the system) is not the same thing as the badge you put in your footer. Do not mix the layers. A soft note like “we use AI tools for drafts, we publish after review” can be a trust choice toward customers. It does not replace the review and, on its own, it is not the answer to the AI Act.
On photorealistic AI-generated images the deepfake rules are tighter. For text on product sheets the focus stays on process, claims, and responsibility. If you generate “almost real” photos of products or results, treat them separately. Here I stop at copy.
What not to delegate to the agent
Do not mass-publish from an agent with access to the catalog. Do not let it update claims on its own after an “optimize for SEO” instruction. Do not hook an MCP server or an n8n automation to the live product page without an approval gate: same logic as wiring a production checkout to a plugin downloaded on Tuesday.
If something goes wrong (an invented claim, a published page, an updated feed), you need to know who approved it. The agent does not sign. You do.
Checklist before “letting AI write everything”
- Do you have a list of allowed and banned claims signed by someone who is not the model?
- Have you decided which fields AI can touch and which it cannot?
- Is human review of the claims mandatory before publishing?
- Do the product page, the feed, and the on-site search tool say the same thing?
- If you use automations, is there a stop, a log, and a rollback?
If you answer no to any of these, you do not have a prompt problem. You have a process problem.
I have already counted the TCO of a mid-size stack elsewhere: licenses, infrastructure, support. Adding AI copy without a process does not lower that cost. It adds a line of risk that never shows up in the spreadsheet.
If you want an honest read of your product sheets, claims, and feed (without being sold “the prompt that ranks”), I can run a targeted audit. Write to me via michy.it: we start from the top SKUs, not from the whole catalog.
FAQ
Do I have to write “generated with AI” on every product sheet?
Not automatically. If there is substantial human review of the content and a clear editorial responsibility for publishing, the disclosure duty on AI text (AI Act, Art. 50) typically does not apply the way it does for text released without control. Review the claims, not just the spelling. This is not legal advice: for edge cases, talk to your own counsel.
Can AI help me get cited by models without putting my claims at risk?
Yes, if you use it after you have closed the data and the claims. Titles, attributes, warnings, consistent feeds: that is what gets you read. An “inspirational” paragraph generated without a banned list does little for GEO and a lot for risk.
Is it better to rewrite the whole catalog or start from the top SKUs?
Top SKUs. Less error surface, a review you can sustain, simpler feed alignment. The whole catalog “in one night” is the fastest way to multiply wrong claims.
