How to Write Product Descriptions That Sell Using ChatGPT

Why AI Product Copy Falls Flat

Two failures account for almost every disappointing listing. Both are caused by what you left out of the prompt, not by the model.

The feature-dump trap

Ask for a description of a “20oz double-wall vacuum-insulated stainless steel bottle” and you will get those exact facts, rearranged and padded. Facts are not persuasion. A buyer standing in front of a spec sheet has to do the translation work themselves — and most of them will not bother.

The missing instruction is the one that says what to do with each fact. Every specification has a consequence, and the consequence is the thing worth paying for. Double-wall vacuum insulation is a fact. Coffee that is still hot at eleven o’clock is a reason to buy.

The buzzword tell

The second failure is vocabulary. Left unconstrained, ChatGPT gravitates toward a narrow band of promotional language — elevate, nestled, must-have, unleash, crafted. Readers have learned to recognise that register, and once they do, the listing stops sounding like a shop and starts sounding like a template.

Same product. Same specs. The difference is entirely in what the prompt was told to do with them.

How to Write Product Descriptions That Sell Using ChatGPT

Four passes, run in order, in one chat. Keep them separate — the single biggest mistake is cramming all four instructions into one paragraph and hoping the model weighs them correctly. It will not. It will follow the last thing you said loudest. Each pass is a small, complete instruction in its own right, which is the same discipline behind the Logic System of role, goal, context and format.

Open a new chat and give ChatGPT the raw material: measurements, materials, what is in the box, who buys it. Then add the one instruction most guides skip — a rule that forbids invention. Without it, the model will cheerfully assign your bottle a warranty length nobody promised.

You are an experienced e-commerce copywriter. I am going to build a product description with you over several steps. Do not write the description yet.

PRODUCT: [product name]
SPECIFICATIONS:
[paste your spec list — dimensions, materials, contents, care, compatibility]
WHO BUYS IT: [describe the buyer in one or two sentences, including what they are trying to get done]
WHERE IT WILL BE SOLD: [Shopify / Etsy / Amazon / your own site]

STRICT FACT RULE: Use only the facts I have given you. Do not assume or invent measurements, materials, certifications, warranties, origin, or performance claims. If something important is missing, ask me for it instead of filling the gap.

For now, reply only with a list of anything you think is missing.

That last line is doing real work. It turns the first exchange into a gap check, and the gaps it finds are usually the exact details a customer would have emailed you about.

This is the pass that separates a listing from a spec sheet. You are asking for a specific transformation: feature, then what it means practically, then the moment in the buyer’s day when it matters. Ask for it in a table so you can see the ones that came out weak.

Using only the specifications above, build a three-column table.

Column 1 — FEATURE: the specification, stated plainly.
Column 2 — ADVANTAGE: what that specification does in practical terms.
Column 3 — BENEFIT: a specific moment in this buyer's day when the advantage matters. Name the moment. Do not write an abstract quality like "convenience" or "peace of mind".

One row per distinct feature. If a feature has no meaningful benefit for this buyer, write "cosmetic only" in column 3 rather than inventing one.

Then tell me which three rows are the strongest and why.

The “cosmetic only” escape hatch matters more than it looks. Without it, ChatGPT will manufacture a benefit for the colourway, and manufactured benefits are where listings start sounding desperate.

Now the tone. Two instructions do most of the work here: a description of how you sound, and an explicit list of words the model may not use. That second one is the technique we cover in full in our guide to negative constraints in prompting — telling the model what not to do is consistently more reliable than telling it what to do.

If you already have listings you are happy with, paste two of them in as examples. Showing the model your voice works better than describing it, which is the whole idea behind few-shot prompting, and it is what keeps a catalogue sounding like one shop rather than forty.

Now write the description using the three strongest rows from the table.

LENGTH: [120–180] words.
VOICE: [describe how you sound — e.g. plain, warm, slightly dry, never hard-sell]. Short sentences. Active voice. Second person.
VOICE EXAMPLES: [paste one or two of your existing descriptions you are happy with]

BANNED WORDS — do not use any of these, or close variants: elevate, unleash, nestled, must-have, game-changer, revolutionary, premium, crafted, perfect companion, take your [anything] to the next level, whether you're a.

ALSO BANNED: opening with a rhetorical question; listing three adjectives in a row; any claim not present in the specifications.

Open with the strongest benefit moment, not the product name.

The last pass handles keywords and layout together, because doing them separately is how you end up with keywords wedged into finished sentences. Give the model your target phrases and let it build the structure around them from the start — the same principle behind writing an SEO blog post with ChatGPT, applied to a much shorter page.

One caution worth repeating to the model explicitly: a keyword that will not sit naturally should be left out. A listing that reads badly converts badly, and conversion is the thing you are actually optimising.

Rewrite the description to work as a live product page.

PRIMARY SEARCH PHRASE: [phrase]
SECONDARY PHRASES: [two or three]

RULES:
- Put the primary phrase in the first sentence only if it reads naturally there. If it does not, put it in the first paragraph.
- Use each secondary phrase at most once.
- If a phrase cannot be used without the sentence sounding forced, leave it out and tell me which one you dropped.

STRUCTURE:
- One benefit-led opening line
- One short paragraph of context
- Four bullets, each starting with the outcome, not the feature
- One closing line covering the most common hesitation about this product

Then write a meta description under 155 characters.

Shopify, Etsy, and Amazon Want Different Copy

Learning to write product descriptions that sell using ChatGPT is only most of the job — a description that converts on your own store can underperform badly on a marketplace, and the reason is structural rather than stylistic. Each platform trains its buyers to read listings a particular way.

Rather than rewriting three times, ask for all three from the finished Pass 4 copy. Marketplace formatting rules change without much notice, so check the current character limits in your seller dashboard before you paste — do not trust a limit ChatGPT states from memory.

Take the finished description and produce three versions of it.

1. STORE VERSION — benefit-led opening line, one short paragraph, four outcome-first bullets.
2. ETSY VERSION — first person, warm, leads with how it is made or why it exists, practical details underneath.
3. MARKETPLACE VERSION — five bullets only, no paragraph. Each bullet opens with the outcome in two or three words, then the supporting fact.

Same facts in all three. Do not add anything that is not already in the specifications. Label each version clearly.

Scaling Past a Handful of Listings

The four passes take about ten minutes the first time and two or three minutes per product after that, because Passes 1 and 3 barely change between products in the same range. Past a certain catalogue size, though, retyping stops being sensible.

At the top end, the answer is to stop pasting instructions altogether and bake them in. Your voice description, your banned-words list and your fact rule all sit permanently inside a Custom GPT built for a recurring task, and you feed it nothing but specs. Building one needs a paid ChatGPT plan; free and low-cost accounts can use GPTs other people have published but cannot open the builder.

Two more prompts earn their keep at any catalogue size. The first mines the language your customers already use. The second rescues supplier copy, which is where most drop-shippers and resellers start from anyway.

Below are customer reviews for [product name].

Identify the three things customers mention most often as reasons they are glad they bought it, and quote the exact phrasing they use. Then flag any recurring complaint or hesitation.

Write three bullets for the product page using the customers' own framing — not marketing language. Do not include anything only one reviewer said.

REVIEWS:
[paste reviews — remove customer names and email addresses first]

Below is a manufacturer's description. It was written for retailers, not shoppers.

1. Extract every verifiable specification into a plain list.
2. Discard all promotional language, superlatives and unverifiable claims.
3. Flag anything that reads like a claim I would need evidence for before republishing.

Do not write a new description yet — return the list and the flags only.

SUPPLIER COPY:
[paste]

Keeping voice consistent across a growing catalogue is its own discipline, and worth reading about separately if you are producing at volume — our guide to scaling content production without losing your voice covers the drift problem in detail.

Will Google Penalise AI Product Copy?

Not for being AI-written, no. According to Google Search Central’s guidance on AI-generated content, the company’s ranking systems assess the quality and usefulness of content rather than how it was produced. What does cause problems is content generated at scale with the primary purpose of manipulating rankings, which falls under Google’s spam policies regardless of whether a person or a model wrote it.

For a store, the practical dividing line is whether each listing tells a shopper something specific and true about that product. Which gives you a short checklist:

• Every listing is materially different. Forty descriptions built from one template with the colour swapped is the pattern that gets flagged.
• Every claim traces back to a spec you supplied. This is the Pass 1 fact rule doing double duty.
• A human read it before it went live. Oversight is what separates assisted writing from bulk generation.
• The page answers a real hesitation. Sizing, compatibility, care — the things buyers email you about.

Disclosure is a separate question, and a commercial one rather than a search one. Most stores do not label individual listings, and there is no search-side requirement to.

What ChatGPT Cannot Do Here

Four things stay yours. Skipping them is how stores end up with returns caused by their own copy — and the first is the one worth understanding properly, because a model filling a gap with a confident guess is how ChatGPT hallucinations happen.

Product data is also business data, which matters if you are working in a personal ChatGPT account. Anything you paste into a consumer account may be used to improve the models unless you turn that setting off yourself. OpenAI’s business plan documentation states that Business and Enterprise workspaces do not train on your data by default — a contractual default rather than a toggle you have to remember. That difference tracks account type, not price.

If your catalogue includes anything unreleased or under NDA, it is worth reading our fuller guide on whether ChatGPT is safe to use at work before you paste a launch calendar into a chat window.

When the Output Goes Wrong

Almost every failure traces back to a missing instruction, and almost every fix is a line you add rather than a rewrite you do. The most common ones, with the specific repair:

The invented-spec failure is the expensive one, because it survives proofreading — a fabricated warranty length reads perfectly well. If you write nothing else into your template, write the fact rule.

Start with one product you already know well, run all four passes, and compare the result against your current listing. The four passes are the whole system, and once you have run them once you will write product descriptions that sell using ChatGPT faster than you can currently write a first draft by hand. Then save Passes 1 and 3 somewhere you can find them — that saved block is what turns a good afternoon into a repeatable process.