Give the AI only facts you can document (ingredients, origin, serving size, how the product is used) and forbid any description of effects on the body. Then run each draft through a claim-check prompt that sorts every sentence by type, and have a person read the flagged lines against your market's written rules before posting.
Unsupported claims mostly enter at the first draft. Ask ChatGPT for "the benefits of" a product and it writes fluent, confident, unsourced ones, because that is what a great deal of food and supplement copy online looks like. So the brief matters more than the edit. One caveat runs through everything below: rules differ by market and change over time, and the AI cannot tell you what yours allow. It can compare wording with rule text you paste in, and a person decides. None of this is legal advice; for anything borderline, ask a qualified adviser or the regulator.
What a benefit-led prompt produces, line by line
Take the prompt a busy owner types at 8 pm: "Write an Instagram post for our golden turmeric latte mix and mention the health benefits." A typical draft looks like this (illustrative):
Boost your immunity and calm inflammation with our golden turmeric latte mix! Packed with antioxidants to detox your body and support healthy joints. Customers say it cured their bloating. Try it today for a healthier you. #immunity #detox #guthealth
That is eight problems in four sentences and a hashtag line. "Boost your immunity" and "calm inflammation" tell the reader what the product does to their body. "Packed with antioxidants" is a nutrient claim with no figure behind it. "Detox" is a word that published advertising rulings have treated as a general health claim, one that can't stand without approved wording beside it. "Support healthy joints" is a function claim, which many markets allow only in approved wording. "Customers say it cured their bloating" is a medical claim dressed as a testimonial, and the AI made it up: ChatGPT has no customers, it is imitating the genre. "A healthier you" is a promise, and the hashtags repeat the claims in a form people forget to check.
Nothing in that draft came from the shop. Every claim arrived because the prompt asked for benefits, which is why the fix starts with the brief rather than with editing afterwards.
A fact sheet and a ban list: the brief that keeps claims out
Build one fact sheet per product, once, from the pack and the supplier's specification. Below is a filled-in sheet for the same turmeric mix (illustrative). The right-hand column is the discipline: a fact with no source doesn't go in.
| Field | Example entry | Where it comes from |
|---|---|---|
| Product and size | Turmeric latte mix, 120 g tub | The pack |
| Ingredients | Turmeric, ginger, cinnamon, black pepper | Label, copied in the same order |
| Allergen line | Copied exactly as printed | The pack |
| Pack claims | Vegan (printed on the front) | The pack |
| Suggested use | One teaspoon in a mug of warm milk or plant milk | Label |
| Taste and texture | Warm, earthy, slightly peppery; fine powder | Staff tasting |
| Supplier line | One sentence from the supplier's own catalogue, copied, with the date | Supplier catalogue |
| Price and shelf | $9.50, second shelf by the till | The till |
The brief then does two jobs. It confines the AI to the sheet, and it bans the vocabulary of benefits, so even a plausible-sounding sentence has nowhere to come from.
You write social posts for an independent health food shop.
Use ONLY the product fact sheet below.
Hard rules:
1. Describe what the product is, what is in it (ingredients as listed),
how it tastes, how it is used (as on the label), its size and price,
and why we stock it.
2. Never say or imply what the product does to the body or for health.
Do not use: health, healthy, wellness, immune, immunity, energy,
digestion, gut, detox, cleanse, inflammation, joints, sleep, mood,
stress, hormones, weight, skin, boost, support, heal, remedy, natural.
3. Do not invent customer comments, reviews, awards, studies or endorsements.
4. Repeat a pack claim (such as vegan or gluten-free) only if the fact sheet
lists it under Pack claims, word for word.
5. If you are unsure whether a sentence breaks a rule, leave it out and
list it at the end under QUERIES.
Write 3 post options of 40-60 words. End each with one practical line
(size, price, shelf). Hashtags: only #healthfoodshop and #refillstore.
FACT SHEET:
[paste the product fact sheet here]The same product, drafted under that brief (illustrative):
Golden turmeric latte mix is back on the shelf: a 120 g tub of turmeric, ginger, cinnamon and black pepper. Stir a teaspoon into warm milk or plant milk for a warm, earthy, slightly peppery mug. Vegan. $9.50, second shelf by the till.
It is plainer, and it says everything a customer can verify by turning the tub over. "Vegan" appears because the fact sheet lists it under pack claims. The QUERIES line is the part owners tend to undervalue: a model told to leave doubtful sentences out and list them will often show you where its habits lie ("a cosy, comforting ritual" is a typical one), and those lines become new entries on your ban list.
Five bands and one flag for sorting every sentence
The check works better when the AI has categories to sort into, rather than an instruction to "be compliant". The five bands below are my working model, not a legal classification. Names and thresholds differ by market, so your regulator's page overrides this table wherever they disagree.
| Band | What it covers | Example | What you do |
|---|---|---|---|
| A. Plain description | What it is, what is in it, size, price, taste, origin, label use | "120 g tub of turmeric, ginger and black pepper" | Check the facts against the pack. No rule lookup needed. |
| B. Pack claims | Organic, vegan, gluten-free, fairtrade and similar marks | "Organic" | Repeat only what the pack or the supplier's certificate states, and check the certificate is in date. |
| C. Nutrient content claims | Statements about how much of a nutrient is in the food | "High in fibre", "low sugar" | In many markets allowed only when the product meets a defined threshold. Compare the nutrition panel with your regulator's table. |
| D. Function or health claims | What a nutrient or ingredient does in the body | "Supports normal immune function" | In many markets only approved wording, used under stated conditions. Copy the approved wording exactly or leave it out. Never paraphrase. |
| E. Disease or medical claims | Treating, preventing, curing or relieving a condition or symptom | "Anti-inflammatory", "helps IBS", "balances hormones" | Not for a food or supplement seller to make. Delete it. No disclaimer repairs it. |
| Flag I. Implied claim | The overall impression suggests an effect even though no sentence states one | "Packed with goodness", a stethoscope in the photo | Treat it like the band it implies. |
Bands D and E are where the trouble starts, and a draft written under the fact-sheet brief should contain none of either. The implied flag matters because regulators in many markets judge the overall impression of an advert, not just its individual sentences, so images and hashtags are part of the same exercise.
The claim-check prompt and what it gives back
The check is a second, separate prompt, ideally in a new chat, so the model isn't marking its own homework in the same breath. Paste the rule text, the fact sheet and the draft.
You are a claims checker for a health food shop. You do not decide what is
allowed. You compare wording with the rule text I paste, and you flag
anything you cannot match.
RULE TEXT (copied from the regulator's page, with page title and date):
[paste]
FACT SHEET:
[paste]
POST DRAFT:
[paste]
1. Split the draft into sentences, hashtags and image descriptions.
2. Give each one a band: A plain description, B pack claim, C nutrient
content claim, D function or health claim, E disease or medical claim,
or I implied claim (the overall impression suggests an effect).
3. For B to I, quote the line of RULE TEXT that covers it. If no line
covers it, write NOT COVERED. A pack claim (B) must also appear in the
FACT SHEET; if it does not, write NOT ON FACT SHEET.
Never write that something is allowed, compliant or safe.
4. Suggest a rewrite that is band A only: no effect on the body.
5. Finish with a list of items that need a human decision.
Output a table: Sentence | Band | Rule line quoted | Rewrite.Here is the kind of result you get on a draft for a muesli (illustrative), including one thing it got wrong:
| Sentence | Band | Rule line quoted | Suggested rewrite |
|---|---|---|---|
| "Our muesli is high in fibre." | C | Rule 3: thresholds for fibre claims | "Oats, seeds and dried apple. See the nutrition panel for fibre per 100 g." |
| "It keeps you regular." | D | NOT COVERED | "Supports regularity" (rejected, see below) |
| "Gluten-free." | B | Rule 5: free-from claims must match the pack | Keep only if the pack says so |
| "Packed with goodness." | A | None | None |
| "Perfect for a healthy gut." | D | NOT COVERED | Delete |
| "#guthealth" | I | NOT COVERED | "#healthfoodshop" |
I'd fix three things before trusting this output. First, "Packed with goodness" was filed as plain description, but it implies an effect, so it belongs under the implied flag; the prompt needs a line saying vague words such as goodness, clean and nourishing count as I. Second, the suggested rewrite "supports regularity" still carries the claim: the model broke its own rule 4. Never accept a rewrite you haven't put back through the check. Third, "Gluten-free" was quoted against a rule line but not against the fact sheet, which in this run hadn't been pasted. The prompt above now includes it, with the NOT ON FACT SHEET instruction.
Pointing the check at your own market's rule text
This is the step that makes the check mean something, and it is the one most shops skip. The AI has no way of knowing which rules apply to you, so you give it the text. Here's the routine I'd follow:
- Find the regulator's own pages for your market on food and supplement advertising and on nutrition and health claims, including any list of permitted wording. Use the regulator's site, not a consultancy's summary of it.
- Copy the relevant sections into one document. Put the page title, address and the date you checked at the top. Cut it to what applies to a shop: advertising and labelling rules, plus the permitted-wording list if there is one. If it runs past a few pages, save it as a file for the Project rather than pasting it each time.
- Create a ChatGPT Project (a Claude Project works the same way) with the rule file, the stop-word list and the claim-check prompt in its instructions. Every check then starts with the rules in front of the model.
- Re-read the source pages every quarter and whenever the regulator announces a change. Put it in the calendar; rules move, and a file that was right in January can be wrong by autumn.
- Name the person who decides. "The AI said it was fine" is not a decision.
Two safeguards. Ask the model to quote the line it relied on, then search your rule file for that line. If you can't find it, the quote was invented, and everything built on it goes in the bin. And if you also sell online to customers in other markets, their rules can apply to them; that is a question for an adviser, not for ChatGPT.
Stop words, and what to write instead
Published advertising rulings have treated "detox" and "superfood" as general health claims that need an approved claim alongside them, and "boosts immunity" is the kind of function claim many markets limit to approved wording. Those rulings date from 2016 and 2018, so check what applies in your market now. Meanwhile, treat the list below as stop words: the AI flags them, and a person decides.
| Wording | Why it is risky | Plain alternative (band A) |
|---|---|---|
| Boosts immunity, immune support | Function claim, often read as preventing illness | "Elderberry and honey syrup, 250 ml. See the label for ingredients and use." |
| Detox, cleanse, flush out toxins | General health claim about the body | "Caffeine-free herbal infusion, 20 bags" (only if the pack says caffeine-free) |
| Anti-inflammatory, heals, soothes | Medical claim | "Turmeric root powder, 100 g. Stir into soups and rice dishes." |
| Superfood | Can itself read as a health claim | Name the food and what to cook with it |
| Balances hormones, fixes your gut, cures | Medical claim | Delete the sentence |
| Clinically proven, studies show | Needs evidence you hold and approved wording | Delete unless your adviser says otherwise |
| Good for your heart, bones or joints | Function claim, approved wording only | Copy the approved wording exactly or leave it out |
| Helps you sleep, relax or focus | Function claim about the nervous system | "Evening herbal tea, caffeine-free" (if on the pack) |
| Natural remedy, alternative to medication | Medical claim, and potentially harmful advice | Delete the sentence |
| Doctor recommended, nutritionist approved | Endorsement you must be able to document | Delete unless you hold the written endorsement and your rules allow it |
A good habit is to ask the AI to run this table against every supplier brochure before you borrow from it. Brochures are written for the supplier's market or for trade readers, and a line that is fine in one place can be a problem in yours. Supplier text is not pre-cleared just because it is on a data sheet. If you also sell supplements on your own website, the tutorial on what to cut from AI-written supplement copy covers product pages, which carry more text, and more room for a claim, than a post.
Testimonials, hashtags, photos and shelf tags make claims too
Most shops clean up the main sentence of a post and forget the rest of what the post says.
- Customer comments. In many markets, a customer's result that you repost, quote or pin is treated as your own claim, and a testimonial can't say a product treats or prevents a condition any more than you can. Taste, service and value comments are fine with the customer's permission. If the best line in a review is a medical result, skip the review. And never ask the AI for "a customer quote": it will invent one.
- Hashtags. Instagram has capped hashtags at five per post since December 2025, which helps: choose five that describe the product or the shop (#refillstore, #loosetea) and none that describe an effect (#immunity, #guthealth).
- Photos. Body shots, before-and-after images, white coats and stethoscopes create an impression of medical backing. A bowl of porridge on a table doesn't.
- Shelf tags and window posters. These are usually the riskiest copy in the building, because enthusiastic staff have written them over years. Photograph every tag and poster once, ask ChatGPT to read the text from the photos into a list, and run the list through the claim check. Expect to find a few you'd rather not have found.
The same routine applies to email. A newsletter is a longer post with more room for a claim to hide in, so the customer newsletter routine works well here if the fact-sheet brief and the check sit in front of it.
Eight post types that need no health claim at all
A health food shop doesn't need benefit language to be interesting. These eight formats work from facts you hold, so they pass the check with little or no editing. Give the AI the product fact sheet, the format, and the ban list each time.
- Ingredient of the week. Where it comes from (from the supplier's own material), what it tastes like, three ways a staff member cooks it. Prompt: "Using only this fact sheet and these three recipes, write a 60-word post and a 100-word email paragraph. Cooking times: say 'follow the pack'."
- Label-reading walkthrough. Pick a product on your shelf and show how to read the ingredient list and the allergen box. You are teaching people to read labels, not telling them what to eat for a condition.
- Refill days. How the refill station works: bring a clean container, we weigh it, you pay per 100 g. Take the price and hygiene rules from your own written policy.
- [Staff first name]'s pick, and how I use it. A short post with the staff member's recipe or serving idea. The sentence "it keeps me going all afternoon" is a claim in disguise, so the check will flag it; keep to how it's used.
- New on the shelf. Three new lines with size, price and one fact each, plus the shelf number. It is the easiest post to write and the easiest to keep honest.
- Supplier spotlight. Who makes the product, as the supplier describes it, with their permission to use any photos. Don't copy their benefit lines.
- Meal-based shelves. "Soup season: lentils, barley, stock cubes, shelf three." You are selling a meal, not an effect.
- The labelled free-from shelf. A tour of products whose packs carry free-from marks, with one standing line: "We can't advise on conditions, but we can show you the label."
Here is how a supplier-spotlight draft changes under the check (illustrative). The supplier's brochure said: "Our small-batch ashwagandha tea calms the nervous system and promotes restful sleep." The first-pass post kept half of it. After the check, the post said: "Our caffeine-free herbal tea is blended in small batches by the supplier; the box holds 20 bags. Steep for the time printed on the box. $6.80." The "calms" and "promotes" lines went; "caffeine-free" stayed because it is printed on the box and is on the fact sheet.
One month of posts: what the check caught
Take a three-person health food shop that plans 12 posts and one email for the month: three posts a week across Instagram, Facebook and a Google Business Profile, plus a monthly email. The owner tries two briefs on the same 12 products and runs both sets of drafts through the claim check. These figures are illustrative; what matters is the pattern.
| Measure | Benefit-led brief | Fact-sheet brief with ban list |
|---|---|---|
| Posts drafted | 12 | 12 |
| Sentences in total | 96 | 90 |
| Flagged by the check | 21 | 5 |
| Of those, band D (function or health) | 13 | 1 |
| Of those, band E (disease or medical) | 8 | 0 |
| Of those, bands B or C (pack or nutrient claims) | 0 | 4 |
| Share of sentences flagged | 22% | 6% |
On the first brief, 21 of 96 sentences were flagged, 13 function claims and 8 medical ones, and every one needed rewriting. On the second, five of 90: four pack or nutrient lines and one function line. The owner reads the five. Three pack claims ("vegan", "organic", "no added sugar") turn out to be printed on the packs, so they stay. One line said "high in protein" and the nutrition panel didn't meet the regulator's threshold, so it becomes "12 g of protein per 100 g, see the panel" (the figure is copied from the pack). The last line, "a tasty way to support your week", is the function line, and it goes.
Time, in this illustration: about 50 minutes to build the fact sheets for the 12 products (a one-off per product), 10 minutes to draft, 15 to run the checks and 15 to read the flagged lines. Call it 90 minutes for the month, and about 40 once the sheets exist. Time your own first month before you trust any of these numbers.
What the check cannot catch is a fact that has gone stale. If the supplier changes the recipe and the new pack now says "may contain nuts", a post written from last year's sheet is wrong even though it is claim-free. Put a "checked on" date on every sheet and re-check the sheet whenever a delivery arrives in different packaging. The claims register for AI-written marketing covers the evidence side if you want to go further, with a filled-in example.
Answering "what should I take for...?" messages without advising
Direct messages and counter questions are where staff and AI both drift into advice. A customer writes: "What's best for my anxiety?" The answer isn't a product. It is a short, kind reply that declines, explains why and offers what you can legitimately give: label information.
You draft replies to customer messages for a health food shop.
Rules:
- Never recommend a product for a health condition, symptom, medicine
interaction, pregnancy or child. Never diagnose.
- Say that we can't advise on health conditions and that a pharmacist or
doctor can.
- Offer what we can do: tell them what is in a product, what the label
says about use, and where it is on the shelf.
- Do not ask for health details. If they share some, don't repeat them.
- Warm, brief, 60-90 words. Sign off with [your name].
Draft a reply to: [paste message]A typical reply (illustrative): "Hi [first name], thanks for asking. We can't advise on health conditions or on what to take for them, and it wouldn't be fair to guess. A pharmacist or your doctor can, especially if you take medication, are pregnant or have a condition. If it helps, I can tell you what's in any product on our shelves and what the label says about using it. Pop in and we'll show you the labels. [your name]"
Notice what is missing: no "I'm sure it'll help", no product name, no reassurance that could be read as a promise. The same limits apply at the counter, which is why the limits on nutrition advice for gym staff are worth reading as a parallel: the rule for a gym and a health food shop is the same shape.
A ten-minute weekly check that the system still works
A claim-check routine decays quietly. People start pasting drafts without the rule text; scheduled posts go out without being checked; new stop words never get added. A short weekly review catches that.
- Pick three posts that went out this week and re-run them through the claim check with the current rule file. Flags on published posts mean the earlier check was skipped or the rules changed.
- Open the regulator pages you copied and compare the date or version against your file. If it moved, update the file and re-run last month's posts.
- Scan comments and messages for customers making their own claims ("this cured my..."). A customer's comment is theirs, but liking, pinning or replying to it can make it yours in some markets. The cautious default is not to engage with it, and to ask your adviser whether to hide it.
- Add any new risky phrase to the stop list, and keep the list in the same Project as the drafts, so nobody can skip it by using a different chat.
- Write one line in a log: date, posts checked, flags found, what changed.
Two process rules keep it honest. Check before scheduling, not after: a scheduling tool makes a bad post look finished. And give one named person the authority to say no, including to the owner. The pre-publish checklist for AI-drafted posts is a useful companion for the non-claim items (links, images, tone). Posts on your Google Business Profile go through the same claim check before they are scheduled, like everything else.
Questions health food shops ask before posting AI drafts
Can ChatGPT tell me whether a health claim is allowed?
No, and it shouldn't be asked to. It has no reliable, current knowledge of your market's rules and will answer confidently either way. What it can do is compare a draft with rule text you paste in, quote the line it relied on and flag anything not covered. Treat 'not covered' as a stop sign and let a named person decide.
Does a disclaimer make a health claim safe to post?
Not on its own. Some markets require a standard disclaimer beside permitted wording, but a disclaimer does not turn a disease or medical claim into an acceptable one, and small print does not cancel the headline above it. Write the post so it needs no disclaimer, and check your own rules for any wording that must accompany a permitted claim.
Can I share a study or article about an ingredient?
Be careful. In many markets a study or article placed next to a product can be read as your own claim about that product. Keep general nutrition information on a separate page or post, avoid linking it to a specific product, and run it through the same claim check. If you are unsure, ask a qualified adviser before publishing.
Do shelf tags and window posters need the same check?
Yes. Printed tags, window posters and counter signs are advertising too, and they are often the oldest and least reviewed copy in the shop. Photograph them, have ChatGPT read the text from the photos, and run it through the claim check once. After that, add the check to the routine for anything new that goes up.
Further reads
- Can AI Write a Tattoo Aftercare Card Without Medical Claims? — Same discipline: keep AI wording inside what you can support.
- AI Marketing Best Practices for a Local Business: A Pre-Post Checklist — A general pre-post checklist for any local business.
- AI Tools for a Grocery Store: Supplier Invoices and Shelf Prices — Supplier invoices and shelf prices for a food retailer.
- Personal Trainer Social Media Posts: A Month of Content With AI — A month of posts in another health-adjacent trade.
- Tea Shop Marketing Ideas: Tasting Notes and Gift Sets Written With AI — Turn your own cupping notes into shelf cards and product pages, then price subscription tiers and gift sets by margin before AI writes a word of copy.
- Baby Store Marketing Ideas: Gift Lists and New-Parent Offers With AI — A gift-list page, a four-email welcome series and milestone offers for an independent baby shop, with prompts and a claims check for safety wording.
- AI in Food Retail: Jobs a Deli or Farm Shop Can Hand Over First — Six jobs a deli or farm shop can hand to AI, the hours each saves, the check that stays with staff, and the order to start in.
- Can a Personal Trainer Edit Testimonials With AI? What Is Allowed — What AI may and may not do to a client's testimonial: consent, result claims, light edits and a four-minute check before you post.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: OpenAI and Anthropic help pages on Projects and reference files; Instagram's published limit of five hashtags per post; trade-press reports and published advertising rulings on 'detox' and 'superfood' wording (from 2016 and 2018, so check current rules with your own regulator).