Academy›Special courses: eight courses for day-to-day selling›AI for listing copy: facts first, then text
Special courses: eight courses for day-to-day sellingAI in everyday selling

AI for listing copy: facts first, then text

Course lesson 1/3 ⏱ ~15 Min. 🏅 up to 205 points By Enes Kurt Updated November 2026 Check your listing free →
What you'll take away

A language model writes five bullets in twenty seconds, and they read more smoothly than much of what is live in your category. That is exactly the risk: the copy sounds finished before it is correct. Language models fill gaps with whatever is most likely, and the most likely feature in a category is often your competitor's, not yours. This lesson shows a workflow that uses the AI's speed without shipping its mistakes: facts first, then text, then a sign-off that traces every claim back to a source. How the individual fields work is covered in the lessons on the title, bullets and description and backend. This one is about using AI for them properly.

1

The workflow: facts first, then text

Most weak AI listings come from a single step: product name and a few keywords in, finished listing out. The AI then has no choice but to guess. So separate collecting, structuring and drafting into five steps:

  1. Build a fact sheet. Everything that is true about the product, each with its source: dimensions and weight (measured yourself), material (supplier specification), contents (sample carton), test reports and certificates with their numbers, care and safety notes. What is not on the fact sheet will not be in the listing.
  2. Collect customer language. The search terms from your keyword research, the most frequent customer questions and typical phrases from reviews in the niche (niche validation). That is your material for use cases and wording.
  3. Fix the outline. Which bullet answers which buying question, which keyword goes into which field. You decide that, not the AI.
  4. Let it draft. Only now does the AI get its brief, with fact sheet, customer language and outline as material.
  5. Sign off. Every claim in the draft is checked against the fact sheet, lengths are counted, mandatory information is added. Only then does the copy go into Seller Central.

The effort shifts: drafting now costs almost nothing, while the fact sheet and the sign-off take time. That is how it should be, because that is where the expensive mistakes happen. You build a fact sheet once per product and update it whenever the supplier changes something. It then serves images, A+ and customer questions too.

In plain terms

The AI is a fast copywriter with no product knowledge. Give a new temp only the product name on day one and they will write an ad about some spice grinder they imagine. Give them the spec sheet, the sample carton and the customers' questions and they will write about yours. The difference is not the writer, it is the brief.

2

What belongs in a good prompt

A good prompt for a language model is a brief, not a magic spell. It has six building blocks:

  • Role and goal: which marketplace, which field, which audience. For example: bullets for amazon.de, buyers are home cooks looking for an electric grinder for coarse salt.
  • The fact sheet verbatim, with the explicit sentence: use only these facts, do not invent features, measurements, certificates or comparisons.
  • Keywords in order of priority: main keyword, two or three secondary keywords, use cases. With the note that each appears only where it fits naturally.
  • Format and limits: number of bullets, structure (a one- or two-word label, then the explanation), the field limits from lesson 16, language and tone.
  • Prohibitions: no filler like “premium”, “bestseller” or “No. 1”, no price, discount or shipping claims, no health, environmental or testing promises, no words in all caps, no emojis.
  • Gap rule: if information for a claim is missing, write [MISSING: …] in its place instead of assuming anything. This one sentence prevents more errors than any other line.

This is the core of the prompt for the AURELO spice grinder set (shortened, fact values as an example):

  • Task: 5 bullet points for amazon.de, factual and friendly, each with a one- or two-word label followed by one sentence of explanation, every bullet under 200 characters.
  • Facts (only these): 2 electric grinders; stainless steel housing; ceramic burr, grind adjustable steplessly from fine to coarse; each runs on 6 AAA batteries, batteries included; height 19.5 cm; capacity about 40 g of peppercorns per grinder; contents: 2 grinders, 12 batteries, 1 drip tray, manual.
  • Keywords: main: electric spice grinder; secondary: salt grinder, pepper mill set, ceramic burr; use cases: coarse sea salt, grilling, housewarming gift.
  • Prohibited: features outside the facts, superlatives, advertising and price claims, health and environmental promises.
  • Gaps: insert [MISSING: …], assume nothing.
Example

Two drafts for the AURELO set, same model, same day (numbers as an example). Draft A got only “AURELO spice grinder set electric” as input. Its five bullets contained 19 factual claims, seven of them with no basis at all: LED light, USB-C battery, “LFGB certified”, five fixed grind settings, 80 g capacity, “lifetime guarantee” and “test winner”. That is 7 of 19, or 37% invented claims, each a potential “not as described” return. Draft B got the fact sheet and the gap rule. It contained 21 factual claims, all backed, plus two markers: [MISSING: weight without batteries] and [MISSING: diameter of the filling opening]. Both gaps really were open on the fact sheet. The second draft is not more creative, it is more honest, and it tells you what you still need to measure.

Tip

Save the prompt as a template. For each new product you swap only the fact sheet and the keywords. Your listings become consistent, and the sign-off gets faster because you know where each fact is supposed to land.

3

The typical AI errors

Language models don't make random mistakes, they make predictable ones. Once you know the patterns, you find them in minutes:

ErrorHow it happensHow you spot it
Invented featuresThe AI adds what products in the category typically haveA claim with no line on the fact sheet
Wrong measurements and unitsInch to centimeter conversion, swapped values, estimated numbersCheck every number individually against the fact sheet
Prohibited advertising claimsTrained on ad copy, the AI likes “bestseller”, “premium”, “No. 1”Search for superlatives and ranking words
Legally risky promisesHealth, environmental, guarantee and test claims sound persuasiveEvery such claim needs proof or goes
Missing mandatory informationThe AI knows neither your registrations nor manufacturer data and warningsCompare with the mandatory-information list from lesson 15
Wrong lengthsLanguage models count characters unreliablyRecount with a counting tool
Translated instead of localized copyThe AI carries over words, not the target country's search termsCheck keywords against local search data (lesson 9)

The most dangerous error is the first one in the table, because it is plausible. Electric spice grinders with an LED light exist by the hundred. Anyone who proofreads only for whether the copy sounds good will miss it. The customer notices when unboxing. Amazon names this behavior openly in the notes on its own AI tools: when information is missing, the AI tries to fill in the blanks, such as whether batteries are required or which color and style apply.

Example

Suppose the invented LED light stays in the AURELO listing. Across 1,000 orders the return rate rises from 4% to 7% (assumption), which is 30 extra returns. Using the return costs from the bullet lesson (about €4.65 per return whose goods can be resold, €9.15 if not), a single invented feature costs €139.50 to €274.50 per 1,000 orders, at a €24.99 selling price and €4.50 purchase cost. On top come the 1-star reviews (“no light, description wrong”) that the listing carries for months. The check that would have caught it takes five minutes.

4

Prohibited and risky claims

Two sets of rules review your copy: Amazon's style rules and unfair-competition law. The AI knows both only roughly and follows neither without an explicit prohibition.

  • Amazon's style rules: no advertising or promotional claims in title and bullets (bestseller, special offer, free shipping, prices, discounts), no words entirely in capitals, no references to other sellers. The details per field are in the lessons on the title and bullets.
  • Advertising legal rights as a feature: a line like “2-year warranty included” presents a statutory right as something special about the offer. German unfair-competition law lists that among the practices that are always misleading (UWG annex no. 10). AI copy produces such lines readily because they appear in so many shop texts.
  • Test and certificate claims: “LFGB tested”, “TÜV tested”, “test winner” only with a report or test you can produce. The AI picks up such seals from competitor copy.
  • Health and environmental claims: “freshly ground pepper boosts your immune system” or “produced climate-neutrally” invite a cease-and-desist letter without proof. What applies to environmental claims since 27 September 2026 is covered in lesson 15.
Important

You are liable for the content, not the tool. Amazon states in the notes on its AI features that you are responsible for representing your products accurately and that by using the content you confirm you have reviewed it. A cease-and-desist letter works the same way: the sender doesn't ask who wrote the sentence. For concrete legal questions about individual wording, a specialist lawyer for competition law is the safe call when in doubt. A separate labelling duty for AI-written product copy does not exist under the EU AI Act, by contrast: for text, Article 50 covers only publications on matters of public interest, and not even those when a person holds editorial responsibility.

5

Amazon's own AI during listing creation

Since June 2024, sellers in Germany can have listing content generated directly in Seller Central. Under “Add products” a draft with title, bullets, description and attributes is created from a few keywords or a short description, from up to three product photos, from the URL of your own product page, or for many new products from a spreadsheet. On top come improvement suggestions for existing offers, AI features in the A+ editor for brand owners, and since 2026 an assistant in Seller Central that answers questions about policies and account data.

These tools have a real advantage: they know your category's attribute schema and fill fields you would have overlooked. They share the weakness of every language model: whatever is not in the input gets filled in. Here is how to use them well:

  • Fact sheet as input: describe the product with the verified facts instead of three buzzwords. The less the AI has to guess, the less it guesses.
  • Only your own sources: for the URL feature, Amazon refers to pages whose content you own or have a license for. A competitor's product page does not belong there.
  • Check attributes one by one: every suggested attribute is a statement about your product that filters, search and Rufus take literally. A wrongly pre-filled field “rechargeable: yes” contradicts your bullet “runs on batteries” and confuses both.
  • Never accept suggestions for live listings blindly: an improvement suggestion can remove a ranking keyword or describe a variation wrongly. Check first which terms bring you sales today (lesson 12).
Most common mistake

“But Amazon suggested it.” A suggestion from Amazon's AI is a draft, not a verified catalog value. Accept it unread and you accept every gap the model filled with an assumption, under your own responsibility. Precisely because the suggestion appears in the familiar form, it gets checked less often than copy from an outside tool.

6

The sign-off before upload

The general check for every listing is the 20-point sign-off. AI copy needs extra checkpoints aimed squarely at the typical errors.

Sign-off for AI copy
  • Every factual claim (material, measurement, quantity, function, certificate) has a line on the fact sheet.
  • Every number checked individually, units verified (centimeters not inches, grams not ounces).
  • No [MISSING] marker left in the copy: gap closed with a real value or the claim deleted.
  • Search for superlatives, guarantee, health, environmental and test words without proof: zero hits.
  • Lengths checked with a counting tool, backend search terms counted in bytes.
  • Mandatory information sits in the fields meant for it, not just in the copy (GPSR, EPR registrations, warnings).
  • Attribute fields and copy contradict each other nowhere.
  • Read aloud: sounds like someone who had the product in hand, without the same sentence pattern five times.
  • A second person has proofread, or you checked the copy a second time the next day.

The attribute point is the one skipped most often. Amazon's AI assistant reads copy and data fields together. If the bullet says “stepless from fine to coarse” and the attribute says “number of grind settings: 5”, your listing has two answers to the same customer question, and both look unsure.

7

Expert insight: claim inventory and measuring length

Experienced teams don't simply proofread AI copy, they take inventory. Each bullet is broken down into individual, checkable claims, and each claim gets one of three tags: backed (source on the fact sheet), derived (follows from backed facts but needs a short check of your own) or unbacked (delete or prove). That sounds laborious but goes fast, because the AI does the breakdown for you. Hand the finished copy back in a new session, without the original prompt, and ask it to list every verifiable factual claim separately. Extracting claims is far easier for a language model than not inventing anything while writing. You then match the list in a spreadsheet with the columns claim, source and status.

Example

Inventory of the five AURELO bullets: 21 claims. 17 are backed. 3 are derived, among them “suitable for coarse sea salt”: it follows from the stepless grind, but you still test it with three kinds of salt before it goes live. 1 claim is unbacked: “pleasantly quiet”. There is no measurement for it, so the word goes, or you measure with a sound level meter and write a number. Result: 20 of 21 claims stay, about 95%, and you know where each one comes from.

Lengths: language models process text in tokens, meaning word fragments, not individual letters. Their self-reported “74 characters” is therefore regularly off. Always count with a tool. For backend search terms Amazon counts bytes: ä, ö, ü and ß take 2 bytes each in German copy (lesson 16). A search term field with 240 characters and 12 umlauts has 252 bytes and sits above every limit in circulation, even though the AI reported “240 characters, fits”.

Important

Two runs of the same prompt produce different copy. For wording that is intended, for facts it is a warning sign. If one run says “40 g capacity” and the next says “50 g”, the AI guessed. Numbers that differ between two runs reliably show you where the fact sheet is missing information or phrases it unclearly.

Versions: for each listing version, file the fact sheet, prompt, draft and inventory. If the supplier changes the material, you update the fact sheet, regenerate only the affected bullets and can show at any time what was claimed when. That helps with questions from Amazon as much as with a cease-and-desist letter, where the state of a specific day matters.

✎ Practice

Your turn: 2 exercises

Solve them to unlock the quiz. Up to 30 points each when you get it right first time without a hint.

1
Spot the mistakes

Find the mistakes in the AI draft

Fact sheet for the electric AURELO spice grinder set: 2 grinders, stainless steel housing, ceramic burr stepless from fine to coarse, 6 AAA batteries each (12 included), height 19.5 cm, capacity about 40 g per grinder, 1 drip tray, manual.

A language model has written bullets from it. Click every part that has to go before upload.

Stepless adjustment: the ceramic burr grinds from fine to coarse. An LED light shows you exactly how much lands on the plate. Stainless steel housing, 19.5 cm tall, capacity about 40 g of peppercorns per grinder. The built-in USB-C battery lasts for weeks. Ready to go: 12 AAA batteries included, 6 per grinder. LFGB certified and test winner in its class. Comes with a drip tray to keep your counter clean. Freshly ground pepper strengthens your immune system. A lovely housewarming gift or treat for your next barbecue.

Tap every part that contains a mistake (4 in total).
2
Sort it

Take inventory: backed, derived or unbacked?

You are breaking the AI copy for the AURELO set down into individual claims. The fact sheet says: 2 electric grinders, stainless steel, ceramic burr stepless from fine to coarse, 6 AAA batteries each, 19.5 cm tall, about 40 g capacity, drip tray. Give each claim the right tag.

Stainless steel housing
Suitable for coarse sea salt
Pleasantly quiet
Drip tray included
Dishwasher safe
Salt and pepper can be ground separately
19.5 cm tall
Rechargeable battery instead of AAA cells

Check yourself

6 quick questions, one at a time, with instant feedback. With a free account your progress is saved.

Why does the AI get a fact sheet before it drafts bullets?
Language models fill missing information with whatever is most likely, and that is often a competitor's feature. The fact sheet defines what may be claimed at all.
Which sentence in a prompt prevents invented claims most effectively?
The gap rule gives the AI a permitted way out instead of forcing it to guess. At the same time it shows you which information is still missing from the fact sheet.
An AI draft contains the bullet “2-year warranty included”. What is the problem?
Presenting statutory rights as something special about the offer is on the German UWG list of practices that are always misleading (annex no. 10). AI copy often picks up such lines from shop texts.
What does Amazon say about content from its own AI listing tools?
The suggestions are optional and editable. Amazon makes clear that the seller is responsible for accurate representation and confirms the review by using the content.
The language model reports that your new title has 74 characters. What do you do?
Language models split text into word fragments and therefore count characters unreliably. You always check lengths with a tool, backend search terms in bytes.
Two runs of the same prompt state 40 g and 50 g capacity. What does that show?
Backed facts stay the same across runs. If numbers differ, the model filled a gap. The difference points you to exactly the spot on the fact sheet you need to complete.

Frequently asked

Can I write my Amazon listing with AI?

Yes. Amazon itself offers AI tools for creating listings in Seller Central. What matters is that every claim is true: you are responsible for the content, no matter who phrased it. There is no labelling duty for AI-written product copy under the EU AI Act.

Does Amazon rank AI-generated listings lower?

Amazon has not documented any downgrading of AI copy as such. What gets evaluated is relevance, complete data and customer behavior. AI copy that is accurate and answers buying questions has no disadvantage. Copy with invented features loses through returns and poor reviews.

What should a prompt for an Amazon listing include?

Role and goal, the fact sheet with a ban on inventing anything, keywords in priority order, format and field limits, a list of prohibited claims, and the gap rule, meaning the instruction to mark missing information instead of assuming it. Most of a good prompt is facts, not wording wishes.

← Previous lessonSponsored Brands video, Brand Story and more: the brand tools that really exist in 2026 Next lesson →AI images and video: product accuracy, labelling and rights
Enes Kurt
Founder of Listimo

Everything in this academy comes from more than ten years of hands-on work with Amazon listings, the same playbook behind Listimo: the tool that finds what holds an Amazon listing back and fixes it.