C1, Lesson 3: Product Pages and Generative Content: Writing 2,000 Descriptions Without Lying
Tags: Sourced / General / Our view / Illustrative. Reading time about 13 minutes.
1. The story: from "write it for me" to "I cannot fulfill that request"
On September 13, 2023, Amazon announced generative AI tools to help sellers write product titles, bullet points and descriptions: a seller gives a few words, an image or a link, and the tool drafts the listing. [Sourced: Amazon, "Amazon launches generative AI to help sellers write product descriptions," Sept 13, 2023; TechCrunch same day] Four months later, on January 12, 2024, Ars Technica reported that Amazon search results held products named things like "I'm sorry but I cannot fulfill this request it goes against OpenAI use policy." Items ranged from lawn chairs to office furniture. Some were taken down shortly after publication. [Sourced: Ars Technica, Jan 12, 2024] Careful point: the report describes sellers pasting raw output from a chatbot into the title field. It does not say these came from Amazon's own tool, and we do not claim that. [Our view] The lesson: the cost of making a description fell to near zero, and the cost of checking one stayed the same. Stores that skip checking get the refusal message on their product page.
2. The one idea
Generative AI makes first drafts cheap. It does not make facts true. A product page is a promise: size, material, what is included, what it works with. A draft that sounds right but states the wrong promise creates returns and legal risk. Cheap writing plus honest review is the win; cheap writing alone is a liability. [General; our view]
3. The kitchen-table version
You have 2,000 items to put on shelves with price tags and a short note. You hire a fast helper who writes beautiful notes from a photo. Some notes say "dishwasher safe" because the photo looked like a dishwasher-safe thing. Nobody checked. Three weeks later, three dozen bottles come back warped. The helper was fast. The checking was the job you skipped. [General]
4. What a product page is made of
- Title: what it is, key attribute, brand.
- Attributes: size, color, material, capacity, compatibility. Structured fields. These feed search filters (Lesson 2), shopping feeds and, soon, shopping agents (Lesson 7).
- Description and bullets: the human-readable promise.
- Images: how it looks, and sometimes a scale or use shot.
- Reviews and Q&A: third-party evidence. Treat as separate from your own copy. Google's product structured data (Merchant Center and search listings) asks for items like price, availability, shipping and returns, so attributes are not only for your site. [Sourced: Google Search Central, Product and merchant listing structured data docs] Amazon says its listing tools now generate more than 70% of required product attributes and that listings made with its AI tools show a 40% increase in "overall listing quality." Both are Amazon's own claims, from a seller marketing page dated November 3, 2025 and footnoted "Amazon internal data," and the quality measure is Amazon's (a composite of its own data-quality scores, audits, seller feedback and listing acceptance rates). [Sourced: Amazon Sell blog, Nov 3, 2025, page re-read Oct 7, 2026; company claim, not independently measured]
5. Where generative AI helps, and where it does not
| Task | Helps? | Why |
|---|---|---|
| Rewrite a supplier's dry spec into readable copy | Yes | Facts are in the input; the model reshapes them |
| Fill missing attributes from a photo | Maybe | Color yes; capacity or material often wrong |
| Write claims ("BPA free," "lasts 24 hours cold") | No, unless supplied | These are legal and safety claims; supply them, never invent |
| Translate listings | Yes, with spot checks | Cheap, still check units and sizes |
| Write reviews | Never | See Section 8 |
| [General; our view] |
6. Worked example: 2,000 descriptions (the $12 water bottle line)
All numbers [Illustrative], computed by script. Your store has 2,000 product descriptions to write or refresh.
- By hand: 10 minutes each, $30 an hour: 2,000 x 10 / 60 x $30 = $10,000.
- AI drafts with no review: generation cost, say $0.02 each = $40.
- AI drafts with a 2-minute human check each: $40 + 2,000 x 2 / 60 x $30 = $2,040. Now the cost of an unchecked error. Say 6% of AI drafts state a wrong fact (120 listings), and 10% of those cause a return costing $8.50 (the figure used in Commerce Module 22): 120 x 0.10 x $8.50 = $102 in return cost. That sounds small. It leaves out lost trust, chargebacks, complaints and legal risk from a safety claim. The point of the arithmetic: reviewing costs $2,000 and the visible return cost is $102, so review has to be justified by the risks, not by the returns alone. A sensible design: automatic checks on every listing, and a human check only on listings with claims, safety, health, or compatibility. [Our view] Also compare to hand-writing: $2,040 versus $10,000 is about 80% cheaper, even with review. That is the real prize.
7. Controls that work
- Give it the facts. Feed the model your own spec sheet, not just a photo. Tell it to use only those facts.
- Required fields and a facts table. Every claim in the copy must match a field (material, capacity, dishwasher-safe yes/no).
- Block lists. Words that need proof: "BPA free," "waterproof," "medical," "FDA," "organic."
- Catch the tell-tale text. Search your catalog for phrases like "as an AI," "I cannot," "I'm sorry," "OpenAI."
- Sample and score. Each week, read 20 random listings and score for facts.
- Keep human voice for the top 20 products. They earn most of the money.
8. The law is not optional
On August 14, 2024, the FTC announced a final rule banning fake reviews and testimonials, including AI-generated ones that claim to be from someone who does not exist, or who had no real experience with the product. The rule lets the agency seek civil penalties against knowing violators. [Sourced: FTC press release Aug 14, 2024; Federal Register notice of Aug 22, 2024 says effective Oct 21, 2024] The rule is in the Code of Federal Regulations (16 CFR Part 465), and on December 22, 2025 FTC staff sent warning letters to 10 companies about possible violations, saying the letters are not formal findings and that violations can bring civil penalties of up to $53,088 each. [Sourced: eCFR, Part 465, read Oct 7, 2026; FTC press release, Dec 22, 2025] Practical rule for a store: never generate reviews or Q&A. Use AI to summarize real reviews if you must, and label it. This is a US federal rule; other countries differ. This is not legal advice. [General]
9. What it costs and what it breaks
- Sameness: if everyone feeds the same tool the same kind of product, every page reads the same. Your page stops being a reason to buy. [Our view]
- Data debt: bad attributes (Commerce Module 1) get better prose, not better facts.
- Brand voice: a store known for dry humor ends up sounding like everyone.
- Ownership: check each tool's terms for who owns the output and what happens to your product data.
10. Monday checklist
- Search your catalog for AI tell-tale phrases.
- Pick your 20 best sellers. Read them aloud. Do they say anything a competitor could not?
- List claims on your pages that you cannot prove from a spec sheet or a test report.
- Decide which fields a human must check before publishing.
11. Try it yourself (no code, 30 minutes)
Take one product you know. Write the facts as a short table. Ask any AI tool for a description using only that table. Mark every sentence that adds something not in the table. Count them. That count is your invented-claims rate for one listing.
12. Self-check
- What mistake did the "I cannot fulfill that request" titles show? Raw chatbot output published without review.
- Which parts of a product page should AI never invent? Safety, health, compatibility and other claims; and reviews.
- 2,000 descriptions: hand-write at 10 minutes and $30 an hour costs? $10,000.
- Why is a $102 return cost an incomplete picture of the risk? It omits lost trust, complaints, chargebacks and legal exposure.
- What did the FTC's August 2024 rule ban that touches AI? Fake reviews and testimonials, including AI-generated ones from people who do not exist.
13. Where this is going
Shopping agents (Lesson 7) will read your attributes, not your adjectives. In that world, correct structured data is worth more than pretty copy. [Our view]
Sources
- Amazon, Sept 13, 2023: https://www.aboutamazon.com/news/small-business/amazon-sellers-generative-ai-tool ; TechCrunch: https://techcrunch.com/2023/09/13/amazon-debuts-generative-ai-tools-that-helps-sellers-write-product-descriptions/
- Ars Technica, Jan 12, 2024: https://arstechnica.com/ai/2024/01/lazy-use-of-ai-leads-to-amazon-products-called-i-cannot-fulfill-that-request/
- Amazon Sell blog (70% attributes, 40% listing quality, company claims): https://sell.amazon.com/blog/amazon-listing-ai
- FTC warning-letter press release, Dec 22, 2025: https://www.ftc.gov/news-events/news/press-releases/2025/12/ftc-warns-10-companies-about-possible-violations-agencys-new-consumer-review-rule ; eCFR: https://www.ecfr.gov/current/title-16/chapter-I/subchapter-D/part-465
- FTC press release: https://www.ftc.gov/news-events/news/press-releases/2024/08/federal-trade-commission-announces-final-rule-banning-fake-reviews-testimonials ; Federal Register: https://www.federalregister.gov/documents/2024/08/22/2024-18519/trade-regulation-rule-on-the-use-of-consumer-reviews-and-testimonials
- Google Search Central Product structured data: https://developers.google.com/search/docs/appearance/structured-data/product
Review flags
- The "I cannot fulfill" listings are not attributed to Amazon's own tool; the article does not say so.
- RESOLVED (Oct 7, 2026): Amazon's figures. The Sell page (dated Nov 3, 2025; read live) still says tools "now generate more than 70% of required product attributes" and sellers "see a 40% increase in overall listing quality," footnoted "Amazon internal data." Both stay labeled company claims. The separate Sept 19, 2024 Amazon post on "Enhance My Listing" and a 2026 Seller Assistant post were read; no comparable percentages found, so there is nothing to reconcile. The "500,000+ sellers" blog figure is not used.
- RESOLVED: FTC rule status. 16 CFR Part 465 is in the eCFR (read Oct 7, 2026; source 89 FR 68077, Aug 22, 2024). The FTC's Dec 22, 2025 release says staff sent warning letters to 10 companies and that civil penalties are up to $53,088 per violation. The lesson states only that. A docket-by-docket search for court challenges was not done; the lesson makes no claim either way. Not legal advice.
- Worked-example numbers (generation cost, 6% error, 10% return rate, $30 hourly) are illustrative; $8.50 per return is reused from our Module 22 illustration.
Claim ledger (Lesson 3)
| # | Claim | Source | Verified | Status | Re-check |
|---|---|---|---|---|---|
| 1 | Amazon announced gen AI listing tools Sept 13, 2023 | Amazon, TechCrunch | 2026-10-06 | Read | none |
| 2 | "I cannot fulfill" titles reported Jan 12, 2024, some taken down | Ars Technica | 2026-10-06 | Read | none |
| 3 | Amazon: >70% of required attributes generated; 40% quality increase | Amazon Sell blog, Nov 3, 2025 | 2026-10-07 | Re-read; company claim ("Amazon internal data") | 2027-01-06 |
| 4 | FTC final rule Aug 14, 2024 incl. AI-generated fake reviews; civil penalties | FTC release | 2026-10-06 | Read | 2027-01-06 |
| 4b | Rule in eCFR Part 465; Dec 22, 2025 warning letters to 10 companies; penalties up to $53,088 | eCFR; FTC release Dec 22, 2025 | 2026-10-07 | Read | 2027-01-06 |
| 5 | Rule effective Oct 21, 2024 | Federal Register | 2026-10-06 | Read | none |
| 6 | Google asks for price, availability, shipping, returns in product data | Google Search Central | 2026-10-06 | Read excerpts | 2027-01-06 |
| 7 | Cost arithmetic ($10,000; $40; $2,040; $102) | Our computation | 2026-10-06 | Illustrative, script-checked | none |