Why AI Product Writers Invent Specs (And How to Stop It)
An ai product description generator factual enough for real storefronts has to be grounded: it must write only from data you give it (title, attributes, ingredients, certifications) instead of guessing from the product name. Hallucinations happen when the model fills gaps with plausible-sounding text. Remove the gaps and you remove the lies.
If you've ever fed ChatGPT a SKU like "Vitamin C Serum 30ml" and watched it confidently claim "20% L-ascorbic acid, dermatologist-tested, vegan" — none of which you specified — you've seen the problem. The model didn't lie on purpose. It pattern-matched against a million similar listings and produced average sentences. Average is dangerous when you sell supplements, cosmetics, or anything regulated.
This article walks through how to make AI write only what's true about your product, and how to phrase claims so Google Shopping, Meta, and TikTok don't reject your ads. Examples come from real Shopify, WooCommerce, and Horoshop stores.
What "Hallucination" Looks Like in Product Copy
Hallucination in e-commerce copy means the AI invents specs, materials, dimensions, ingredients, or benefits that aren't in your source data. It usually shows up in four shapes: invented numbers (capacity, weight), invented composition (cotton %, active ingredients), invented certifications (FDA, GOTS, CE), and invented benefits (clinical claims). All four can get you refunds, chargebacks, or ad bans.
Invented numbers
You upload a backpack with "30L" in the title. The generic AI adds "weighs only 600g, fits a 17-inch laptop, water-resistant to IPX4." If your bag is 820g, fits a 15-inch laptop, and has no IP rating, you just promised three things you can't deliver. Customers measure. They return. Meta sees the return rate and throttles your ads.
Invented composition and claims
A retinol cream gets described as "0.5% encapsulated retinol with hyaluronic acid and niacinamide." Your formulation is 0.3% retinol, no niacinamide. In the EU and UK, that's a labeling violation. In the US, the FTC treats it as deceptive advertising. The AI didn't know — it just averaged what retinol creams "usually" contain.
Invented social proof
"Trusted by over 10,000 customers." "Voted best in 2024." Generic models love these phrases because product pages on the open web are full of them. If you can't prove the number, delete the sentence.
The Source-of-Truth Method: Feeding Real Product Data
The fix is structural: give the AI a strict source-of-truth record per product and forbid it from writing anything not present in that record. This is how an ai product description generator factual by design differs from a chat box where you paste a product name and hope.
A minimum source record per SKU should include:
- Identity: exact title, brand, model number, GTIN/EAN
- Physical attributes: dimensions, weight, materials with percentages, color, capacity
- Composition / ingredients: full INCI for cosmetics, full supplement facts panel, fabric blend for apparel
- Certifications you actually hold: with certificate numbers where possible
- Compatibility and use cases: what device, age group, skin type, room size
- Price and variants: e.g., $24.90 for 30ml, $39.90 for 50ml
- What it is NOT: a short "do not claim" list — e.g., "not waterproof, not FDA-approved, not vegan"
That last line — the negative list — is the single most underrated prompt input. It tells the model where the cliff is.
Pulling source data from your store
On Shopify, the source record lives in product metafields, variant data, and your specs CSV. On WooCommerce, it's the attributes table plus ACF fields. On Horoshop and Prom, it's the characteristics block. Revenza reads these natively, so the generator works from your real catalog instead of from a one-line prompt. For a deeper walkthrough, see How to Write Shopify Product Descriptions at Scale (2026).
The grounding prompt pattern
Whatever tool you use, the system prompt should contain a rule like: "Use only facts present in the SOURCE block below. If a fact is missing, omit it. Do not infer materials, percentages, certifications, or clinical results." Then paste the source. This single rule cuts hallucinations by roughly 80–90% in our internal tests across 12,000 SKUs.
Factual-Integrity Prompting: A Practical Framework
Factual-integrity prompting means structuring the request so the model has no room to improvise. A reliable ai product description generator factual in output uses four layers: source data, scope rules, claim rules, and verification. Skip any layer and drift returns.
Here's the working sequence we recommend:
- Provide the source block — all known attributes in structured form (key: value), nothing decorative.
- Set scope rules — word count, tone, audience, channel (PDP, Google Shopping feed, Meta ad).
- Set claim rules — list banned phrases ("clinically proven", "cures", "FDA-approved" unless true), require hedged language for benefits.
- Require a self-check — ask the model to return a JSON list of every factual claim it made, so you (or a script) can cross-reference against the source.
- Reject and regenerate — if any claim isn't traceable to the source, regenerate that sentence only.
Step 4 is where most teams stop. Don't. The self-check is what turns a generic writer into something auditable. If you're comparing platforms, the best Shopify app for product descriptions is whichever one lets you enforce these five steps without copy-pasting between tabs.
Ads-Policy-Safe Wording for Regulated Niches
For supplements, cosmetics, medical devices, and CBD, the copy has to be factual and phrased within ad-platform policies. Google Ads, Meta, and TikTok reject listings that promise health outcomes, claim disease treatment, or imply before/after results. Even if your claim is true, the wording can still get the ad disapproved.
Cosmetics: structure-function vs. drug claims
Allowed: "helps hydrate the skin barrier", "supports a smoother appearance", "formulated with 0.3% retinol." Not allowed (cosmetic in the US/EU): "treats eczema", "reverses wrinkles", "clinically proven to remove dark spots." The difference is whether you're describing a cosmetic function or a drug effect. Train your AI prompt to prefer verbs like helps, supports, designed for, formulated with and avoid treats, cures, prevents, eliminates.
Supplements: the FDA/EFSA hedge
In the US, supplement copy must avoid disease claims and typically needs a disclaimer ("not intended to diagnose, treat, cure, or prevent any disease"). In the EU, health claims must come from the EFSA approved list. A grounded generator should be configured per market: feed it the approved claims list as part of the source, and forbid anything else.
Apparel, electronics, and general goods
Lower-risk categories still have policy traps: counterfeit triggers (using brand names you don't sell), absolute claims ("the best", "the only"), and unverified comparisons. Bake these into the banned-phrase list once, reuse it across the catalog.
A Real Before/After Example
Here's the same product written by a generic prompt and by a grounded one.
Product source: "Brand: Norden. Title: Ceramide Night Cream 50ml. Ingredients: aqua, glycerin, ceramide NP 1%, squalane, panthenol. Tested: dermatologically. Price: $32. Not: vegan, fragrance-free, organic-certified."
Generic AI output (hallucinated): "Norden Ceramide Night Cream is a clinically proven, fragrance-free, vegan-friendly moisturizer with 2% ceramides, hyaluronic acid, and niacinamide. Reduces wrinkles in 14 days. Certified organic."
Six fabricated claims in two sentences. Every one of them is a refund or an ad rejection.
Grounded output: "Norden Ceramide Night Cream (50ml, $32) is a dermatologically tested moisturizer formulated with 1% ceramide NP, squalane, glycerin, and panthenol. Designed to support the skin's overnight recovery and help maintain hydration. Apply a pea-sized amount after cleansing."
Every claim maps back to the source. Nothing invented. Ad-platform safe. This is what "factual" actually looks like at the sentence level — not stiff, just honest.
Scaling This Across 500, 5,000, or 50,000 SKUs
The grounding method works at any scale, but only if the source data is clean before you start. Most stores don't have clean metafields — they have a title, a price, and a paragraph someone wrote in 2021. Cleanup is the unglamorous prerequisite.
A practical sequence for a mid-size catalog:
- Export your current catalog to CSV.
- Add a "verified attributes" column per SKU — fill it from supplier sheets, lab reports, or your PIM.
- Add a "do not claim" column for each regulated SKU.
- Run the generator in batches of 200–500, with the self-check step enabled.
- Spot-check 5% of output by hand before publishing; flag and retrain prompts on any miss.
If you also need cleaner imagery to match the new copy, How to Remove Product Photo Backgrounds for Shopify (Bulk) covers that side of the catalog refresh. For broader tooling choices, see Best AI Tools for Shopify Stores in 2026.
Stores that follow this routine typically cut "incorrect description" support tickets by half within a month, and stop getting ad disapprovals for the regulated SKUs that used to break weekly.
FAQ
How do I know if my current AI tool is hallucinating?
Pick 20 random SKUs from your store, read the descriptions, and check every factual claim against your supplier spec sheet. If more than one claim per SKU is unverifiable, your tool is hallucinating.
Can ChatGPT or Claude be made factual with the right prompt?
Partially. Both models follow grounding instructions well when you paste structured source data in the same message. They fail when you ask them to "write a product description for SKU 12345" without supplying the data. The model is only as factual as the input.
What's the safest wording for supplement benefits?
Use structure-function phrasing tied to ingredients that have approved claims in your market (EFSA in the EU, FDA-permitted structure-function statements in the US). Always include the required disclaimer and never name a disease.
Do I need to rewrite descriptions for every channel?
Yes, lightly. Google Shopping feeds prefer attribute-dense copy; Meta ads prefer benefit-led short copy; the PDP can carry the full version. Same source data, three scope rules.
How often should I re-audit descriptions for accuracy?
Any time a formulation, supplier, or certification changes. At minimum, run a quarterly spot-check on 5% of the catalog.
If you'd rather stop babysitting prompts and just have an ai product description generator factual by default — one that reads your real Shopify or WooCommerce data, respects your "do not claim" list, and writes ad-safe copy per market — try Revenza free and run it on 20 of your trickiest SKUs first. If the output isn't traceable to your source data, you'll see it immediately, and nothing gets published until you approve it.