ChatGPT Prompts for Product Descriptions That Sell

Product descriptions are the copy marketers most want to hand to AI, because the volume is brutal - hundreds of SKUs, four channels, two languages - and the format looks simple. They’re also the copy where a model’s bad habits cost actual money. ChatGPT under-informed doesn’t write a thin description. It writes a confident one, inventing the water resistance rating, upgrading “cotton blend” to “premium organic cotton” and promising whatever makes the paragraph flow.

That’s not a style problem. In Salsify’s 2025 consumer research, a survey of 1,910 US and UK shoppers, 71% had returned items over inaccurate descriptions, and 54% had dropped purchases because product information conflicted between channels. The description that oversells collects the sale and then pays it back as a return, with shipping, and the description that contradicts your Amazon listing kills the sale outright. So I built every prompt below on one discipline: the model writes only from a fact sheet you gave it, and scaling happens by reusing the system, never by letting ChatGPT fill gaps.

The base prompt

I set this up once per store or product line, on top of the brand voice system prompt if one exists:

You are an ecommerce copywriter for [store - what it sells, to
whom, price positioning in one sentence].
Voice: [2-3 adjectives; one thing we never sound like].

Rules for every description you write:
- use ONLY facts from the fact sheet I provide. If a claim isn't
  on the sheet, you may not make it. If the sheet is missing
  something a buyer needs (size, material, compatibility, care),
  list it as a question instead of writing around it.
- lead with what the product does for the buyer, follow with how,
  end with a scannable specs block.
- no superlatives without a fact attached. "Premium", "best",
  "revolutionary" are banned unless the sheet proves them.
- write for a shopper on a phone: short paragraphs, front-loaded
  sentences, nothing that needs a second read.

Wait for the fact sheet.

Why it works: the fact-sheet rule converts hallucination from a silent failure into a visible question list. The banned-superlatives rule attacks the other failure mode - adjective soup that says nothing and reads like every AI store on the internet. What’s left for the model is the part it’s genuinely good at: turning verified attributes into fluent, benefit-led copy at volume.

The seven prompts

1. The fact sheet builder

Here is everything I have on this product: [paste - supplier
specs, old description, internal notes, top customer reviews].
Build a fact sheet: verified attributes (with the source line
for each), benefits each attribute supports, who this product
is for and not for, and a list of gaps I need to fill before
we write. Do not resolve contradictions yourself - flag them.

Why: this is the prompt that makes the whole system safe, and I never skip it even when I’m in a hurry. Supplier data is wrong often enough that “flag contradictions” earns its line - when the spec PDF says 400g and the old listing says 350g, I want that surfaced before it ships in copy, because that discrepancy is precisely the Salsify 54% problem in miniature.

2. The core description

Fact sheet: [paste from prompt 1, gaps filled].
Write the product description: a 1-sentence hook naming the
buyer's outcome, 2-3 short benefit paragraphs each anchored to
a fact, and a specs block. Then a 40-word short version for
category and search pages. Both must stand alone.

Why: the two lengths come from how descriptions actually get consumed - the full version on the product page, the short one everywhere else the product appears. Asking for both in one pass keeps them consistent, which is the cheap insurance against the conflicting-information problem. The “anchored to a fact” phrasing keeps benefit paragraphs from detaching into vibes.

3. The SKU scaler

These [number] products share a template but differ on:
[the differentiating attributes - color, size, capacity, model].
Master fact sheet: [paste]. Per-SKU differences: [paste table].
Write descriptions for each. Vary sentence structure and opening
between SKUs so they don't read as find-and-replace copies, but
never vary a fact. Flag any SKU where the differences are too
small to justify unique copy.

Why: this is where AI genuinely changes the economics of description writing - a 200-SKU catalog stops being a quarter’s work. The anti-find-and-replace instruction matters for both shoppers and search engines, and the final flag is the honest out: some SKUs should share copy with a variant selector, and pretending otherwise generates near-duplicate pages you’ll regret.

4. The objection pass

Here are this product's most useful customer reviews, including
critical ones: [paste 5-10].
List the hesitations and objections buyers actually have. Then
revise the description to answer the top 2 inside the copy -
factually, not defensively. If a review reveals the description
promises something the product doesn't deliver, flag that as a
product-content problem, not a copy problem.

Why: reviews are the only place buyers tell you why they almost didn’t buy, and mining them is the same move I use in my competitor analysis prompts - evidence in, claims out. The last line is there because sometimes the fix isn’t better copy, it’s an honest walk-back of an oversold claim, and that call belongs to you, visibly.

5. The search fields

From the fact sheet and final description, write: a title tag
(under 60 characters, leading with [target term]), a meta
description (under 155 characters, with the product's strongest
specific), and alt text for [number] product images I'll
describe: [describe what each image shows].
Target term: [from your keyword research]. Do not stuff variants.

Why: these fields get generated last because they compress the description, not the other way round. The target term comes from you, not the model - ChatGPT invents search volumes when asked for keywords, so the research step stays outside this workflow. Alt text runs on described images because the model can’t see your photos, and guessing produces alt text that’s confidently wrong.

6. The channel adapter

Final approved description: [paste].
Adapt it for [channel: e.g. Amazon - 5 benefit bullets + a
200-word paragraph, no promotional language; or Instagram Shop -
2 sentences, voice-forward].
Hard rule: every fact must match the original exactly. Change
form, never substance. List what you cut and why.

Why: channel adaptations are where conflicting information is born - someone rewrites for Amazon’s format, shortens a spec, and now the listings disagree, which is the exact failure the Salsify data prices at 54% of shoppers dropping a purchase. Anchoring every adaptation to one approved source description, with a stated no-drift rule, is the process fix, and the cut list gives you a ten-second review instead of a line-by-line comparison.

7. The claims check

Description to check: [paste final].
Fact sheet: [paste].
Audit line by line: flag every claim not supported by the sheet,
every superlative without evidence, and anything a regulator or
marketplace could read as a measurable promise (durability,
health, environmental, "guaranteed"). Output a table: claim /
status / suggested fix.

Why: I run this even on descriptions humans wrote, because oversold claims predate AI - the model just industrialized them. Environmental and health claims get special attention in the prompt because they’re the categories regulators actually move on, and “eco-friendly” with no substantiation is the kind of thing that survives four rounds of internal review and then costs you a marketplace suspension.

Your descriptions are read by machines now too

A growing share of product research happens inside AI assistants, and when ChatGPT or Perplexity answers “best [your category] under $100”, the material it works from includes your product pages. The same discipline that serves human shoppers serves this audience: specific facts, plain claims, specs the model can quote without guessing. I’ve written up how ChatGPT search picks what it cites and what earns citations across AI engines - the short version for product pages is that a description with real numbers in it is quotable, and adjective soup is not. Fact-sheet copy wins twice.

Measuring whether any of this worked

The backlog of description rewrites should be tested like any other change, and my working metrics are the pair the Salsify data points at: conversion rate on the rewritten pages against their own baseline, and the return rate with “not as described” reasons on those SKUs a quarter later. The first number can go up while the second goes up too - that’s the overselling signature, and it means the copy won the click and lied. A rewrite has actually worked when conversion holds or rises and description-related returns fall. If you can only watch one number, watch the returns reasons. It’s the one shoppers fill in honestly.

For where description work sits inside the rest of the toolkit, the prompt anatomy behind everything here is in the prompt library, and the bigger workflow map is the complete guide to AI marketing.

FAQ

Can I just paste supplier descriptions and ask ChatGPT to rewrite them? It works and it’s the most common workflow, but you inherit every error in the supplier copy plus whatever the paraphrase invents, with no fact sheet to audit against. If the catalog is big, run prompt 1 across it in batches first - it’s the difference between scaling copy and scaling mistakes.

How long should a product description be? As long as the fact sheet is interesting. A commodity phone case earns 60 words and a specs block. A $400 technical jacket earns every objection-answering paragraph you can support. The failure at both ends is the same: length decided by a template instead of by what a buyer needs to stop hesitating.

Do AI-written descriptions hurt SEO? Generic ones do, the same way generic human ones always did - especially at SKU scale, where near-duplicates compete with each other. The system above is built against exactly that: unique facts per SKU, varied structure, and the prompt-3 flag for products that shouldn’t have separate pages at all.