9 AI Product Description Mistakes and What Each One Costs a Seller

TL;DR

Nine errors AI puts into product copy, each tied to the refund right, marketplace policy or advertising rule it breaks.

An AI-written product description fails in nine predictable ways, and each one has a documented cost. Three make an item "not as described", which UK guidance, EU law and eBay policy turn into a refund, a repair or a return at the seller's expense. Five break a specific written rule on fibre content, origin, environmental claims, health claims or fake reviews. One gets a Google Shopping listing disapproved. None is a style problem. All are facts the model had no source for.

Grouped below by what the mistake costs, with the rule quoted for each.

Key Takeaways

"Not as described" means a refund. UK guidance says a seller "must offer a full refund if an item is faulty, not as described or does not do what it's supposed to" (GOV.UK, retrieved 2026-09-17).

On eBay, the seller pays return shipping when an item does not match the listing (eBay Money Back Guarantee policy, retrieved 2026-09-17).

Google Merchant Center requires AI-written descriptions to be declared in their own attribute (Merchant Center description specification, retrieved 2026-09-17).

From 27 September 2026 the EU prohibits generic environmental claims a trader cannot back with recognised excellent environmental performance (European Commission, retrieved 2026-09-17).

A review from a reviewer who does not exist breaks the FTC's reviews rule, whatever wrote the text (16 CFR 465.2, retrieved 2026-09-17).

Why product descriptions go wrong in a particular way

The input is a short list of attributes; the output is expected to be longer and more persuasive than the input. Everything the model adds to get there is restatement or invention. A 2024 paper on LLM-enriched product listings defines hallucination as "the generation of content that is unfaithful to the source input" and says it "poses significant risks in customer-facing applications" (Hallucination Detection in LLM-enriched Product Listings, ECNLP 2024, retrieved 2026-09-17, SOURCED). The problem predates current models: an EMNLP 2019 paper called it "of vital importance to generate faithful descriptions that stick to the product attribute information" (Chan et al., arXiv:2503.08454, retrieved 2026-09-17, SOURCED).

The first three turn into refunds.

Mistake 1 — Specifications the input never contained

You supply a name, a material and a price. The model returns a weight, a battery life and a dimension, because good product descriptions contain those.

UK guidance for businesses: "You must offer a full refund if an item is faulty, not as described or does not do what it's supposed to" (GOV.UK, Accepting returns and giving refunds: the law, retrieved 2026-09-17, SOURCED). In the EU, if goods "do not look or work as advertised, the seller must repair or replace them at no cost", with a full or partial refund where that fails (Your Europe, Guarantees, retrieved 2026-09-17, SOURCED).

Catch it: every number in the output must appear in the input. A figure with no source line is an invention until proven otherwise.

Mistake 2 — Details bled in from a sibling product

Generate fifty variants in one session and attributes travel. The navy version inherits "charcoal grey" from the previous row; the 500ml bottle inherits 750ml. Nothing is invented — real data lands on the wrong product (ANALYSIS).

The buyer receives the same mismatch either way. eBay's policy: "Sellers are required to deliver the item as it was described in the listing," and for those returns "the seller is responsible for return shipping" (eBay Money Back Guarantee policy, retrieved 2026-09-17, SOURCED). Google Merchant Center asks you to "ensure information provided here is consistent with other details in the data source (for example, Title) and on your landing pages" (Merchant Center description specification, retrieved 2026-09-17, SOURCED).

Catch it: one product per prompt, or check each output's colour, size and capacity against its own row.

Mistake 3 — Compatibility and fit claims

"Fits all standard models." "Compatible with every major phone." The model fills a gap the buyer cares about with the most reassuring sentence available.

A compatibility claim states what the product does, and GOV.UK names that case separately: a refund is owed when an item "does not do what it's supposed to" (linked above, SOURCED). A charger that does not charge the promised phone is the plainest example.

Catch it: cut every compatibility claim not backed by a manufacturer specification you hold. "Compatible with" needs a list, not an adjective.

The next five break a written rule whether or not anyone sends the item back.

Mistake 4 — Upgraded fibres and materials

"Cotton" becomes "Egyptian cotton"; "cotton blend" becomes "100% cotton". Models reach for premium vocabulary.

In the US that is regulated copy. The FTC's textile guidance says written advertising "includes internet advertising", and if an ad "makes any statement about a fiber, or implies the presence of a fiber, the fiber content information that's required on the label must appear in the ad, minus the percentages." On premium cottons: "make sure it doesn't convey that the product is made only of the premium cotton, unless that's true" (FTC, Threading Your Way Through the Labeling Requirements, retrieved 2026-09-17, SOURCED).

Catch it: copy fibre names from the care label, in its order. Never let a model name a material.

Mistake 5 — Origin claims nobody checked

"Handcrafted in Italy." "Proudly made in the USA." High-probability phrases for some categories, written by a model that cannot know where your product was made.

For textiles sold online, the FTC says "the description must include a clear and conspicuous statement that the item was either 'made in U.S.A.,' 'imported' or 'made in U.S.A. and [or] imported'" (FTC textile guidance, linked above, SOURCED). The Made in USA Labeling Rule bars an unqualified US-origin label unless "all or virtually all ingredients or components of the product are made and sourced in the United States", and extends to mail order material, defined to include material "disseminated in print or by electronic means", that carries such a label (16 CFR Part 323, retrieved 2026-09-17, SOURCED).

Catch it: delete every origin statement the model wrote, then add back the one your supplier documentation supports.

Mistake 6 — "Eco-friendly", "sustainable", "green"

The commonest filler adjectives in AI product copy, and the category with the newest rule.

The FTC's Green Guides say "marketers should not make unqualified general environmental benefit claims," and use "Eco-friendly" as a worked example of a deceptive brand name (16 CFR 260.4, retrieved 2026-09-17, SOURCED).

In the EU, Directive (EU) 2024/825 enters into application on 27 September 2026 (European Commission, linked above, SOURCED). It adds to the practices prohibited in all circumstances: "Making a generic environmental claim for which a trader is not able to demonstrate recognised excellent environmental performance relevant to the claim" (European Commission, Questions & Answers on Directive (EU) 2024/825, June 2026, retrieved 2026-09-17, SOURCED).

Catch it: flag "eco", "green", "sustainable" and "planet" in every output. Keep a provable fact — "made with 60% recycled polyester" — and cut the adjective.

Mistake 7 — Health, safety and efficacy claims

"Soothes irritated skin." "Supports healthy sleep." A model writing a cosmetics or supplement listing produces these unprompted.

The FTC says claims about "the health benefits or safety of foods, dietary supplements, drugs, and other health-related products require substantiation in the form of competent and reliable scientific evidence," and that "as a general matter, substantiation of health-related benefits will need to be in the form of randomized, controlled human clinical testing" (FTC, Health Products Compliance Guidance, retrieved 2026-09-17, SOURCED).

Catch it: no health claim ships without the specific study behind it on file.

Mistake 8 — Invented reviews, ratings and testimonials

Ask for "social proof" and a model writes a quote from a happy customer who does not exist.

The FTC's rule makes it "an unfair or deceptive act or practice" for a business "to write, create, or sell a consumer review, consumer testimonial, or celebrity testimonial that materially misrepresents, expressly or by implication: (1) That the reviewer or testimonialist exists" (16 CFR 465.2, retrieved 2026-09-17, SOURCED). The test is whether the reviewer exists, not what wrote the words (ANALYSIS).

Catch it: never prompt for social proof. Search outputs for quotation marks, star characters and "customers".

The last one is the easiest to prevent.

Mistake 9 — Price, shipping and delivery promises inside the description

"Free next-day delivery." "Now 20% off." Written into a description once, these outlive the promotion.

Google Merchant Center: "Don't include promotional text such as price, sale price, sale dates, shipping, delivery date, other time-related information, or your company's name," and not meeting its minimum requirements "can lead to disapprovals or limit the serving of your ads" (Merchant Center specification, linked above, SOURCED). The FTC's mail order rule adds that "you must have a reasonable basis for stating or implying that you can ship within a certain time" (FTC, Business Guide to the Mail, Internet, or Telephone Order Merchandise Rule, retrieved 2026-09-17, SOURCED).

Catch it: prices, promotions and shipping live in their own structured fields. The description describes the product.

The disclosure rule most sellers have not read

Merchant Center's specification states: "All descriptions created using generative AI must be provided using the structured description [structured_description] attribute instead of the description [description] attribute," with the digital source type set to "trained_algorithmic_media" — and "if you provide both the structured description [structured_description] and description [description] attributes we'll only use the description [description] attribute" (linked above, SOURCED). If your feed pushes AI text through plain description, check it now.

What each mistake costs

#	Mistake	Rule	Cost
1	Invented specifications	GOV.UK refund guidance; EU legal guarantee	Refund, repair or replacement
2	Sibling-product details	eBay Money Back Guarantee; Merchant Center	Return at seller's cost
3	Compatibility and fit	GOV.UK refund guidance	Refund
4	Upgraded fibres	FTC textile rules	Non-compliant ad copy
5	Origin claims	FTC textile rules; 16 CFR 323	Non-compliant origin claim
6	Generic green claims	16 CFR 260.4; Directive (EU) 2024/825	Prohibited in the EU from 27 Sep 2026
7	Health and efficacy	FTC Health Products Compliance Guidance	Needs clinical substantiation
8	Invented reviews	16 CFR 465.2	Unfair or deceptive practice
9	Price and shipping in text	Merchant Center; FTC mail order rule	Disapproval; shipping-time liability

These are the error types from AI content errors: the 14 mistakes that reach published pages — invented numbers, stale prices, unverifiable claims. Product copy attaches a refund to them.

How This Guide Was Sourced

Written and maintained by the LogicBalls editorial team (logicballs.com).

AI involvement. This post was AI-assisted. Every quoted line was fetched from its source on 2026-09-17 and matched against the text above, and every link was resolved before publication.

Disclosure: LogicBalls builds AI writing tools, including an AI product description generator whose page describes a clarification-first process. All nine mistakes apply to our tool's output as much as any other model's, and all nine checks are editing, not software.

Sources. GOV.UK and Your Europe for refund and guarantee rights; eBay's Money Back Guarantee policy and Google Merchant Center's description specification for marketplace rules; FTC business guidance and 16 CFR Parts 260, 323 and 465 via eCFR for US rules; the European Commission's page and June 2026 Q&A for Directive (EU) 2024/825; the ECNLP 2024 and EMNLP 2019 papers for the research. All retrieved 2026-09-17.

What could not be fetched. The UK Consumer Rights Act 2015 on legislation.gov.uk, and the Sale of Goods Directive and Directive 2024/825 on EUR-Lex, returned bot challenges to automated requests. We quote only the official guidance summarising them. None of this is legal advice.

Why the title changed. This post was planned as "9 AI Product Description Mistakes That Cause Returns". We found no citable source measuring how often inaccurate descriptions cause returns, and five of the nine are rule breaches rather than return triggers. The title promises what the sources support.

What is not claimed. No return rate, no frequency ranking of the nine, no estimate of how often models make each. No LogicBalls telemetry is used in this guide. Every figure above is external and linked.

Frequently Asked Questions

Do AI-written product descriptions have to be disclosed?

In Google Merchant Center feeds, yes: generative-AI descriptions go in structured_description, marked "trained_algorithmic_media". Other platforms set their own rules.

Is a wrong spec really grounds for a refund?

In the UK, GOV.UK guidance says a full refund is owed when an item is "not as described". In the EU, repair or replacement comes first, then a refund. On eBay, the seller pays return shipping.

Can I say "eco-friendly" if the product uses some recycled material?

In the EU from 27 September 2026, only with recognised excellent environmental performance. In the US, the Green Guides say to qualify it with the specific benefit. State the specific fact instead.

Does a better model fix this?

It reduces some errors and changes none of the rules. An invented origin claim is non-compliant whichever model wrote it.

Conclusion

Invented specifications, sibling-product bleed and compatibility claims become refunds. Upgraded fibres, origin claims, green adjectives, health claims and invented reviews break written rules. Price and shipping text gets listings disapproved. All nine are caught the same way: compare the output with the input and delete what has no source.