AI Product Descriptions Without Invented Facts: A QA Checklist

To rewrite product descriptions with AI without inventing facts, give the model only your own product data, forbid it from adding any material, measurement, certification or claim that is not in that data, and check every draft against the original before it goes live. Keep the previous version so you can roll back. A description that sounds better but states one wrong spec costs more than the old one did.

This checklist covers the rules to set, the checks to run and the mistakes to watch for.

Why do AI rewrites invent facts?

A language model writes what is likely, not what is true. Ask it to make a description “more detailed” and it fills the gaps with details that are common for that kind of product: a cotton t-shirt becomes “100% organic cotton”, a backpack gains “water resistant” fabric, a candle gets “a 40 hour burn time”. Each of those sounds right. None of them came from you.

That matters more now than it used to. AI shopping assistants quote product descriptions to shoppers, word for word, as facts about your product. A made up spec in your description becomes a made up spec in an assistant’s answer, and then a return.

What should the AI never add?

Write these into your instructions, every time:

Never add	Why
Materials and fabric percentages	The most common invention, and a legal claim in many markets.
Measurements, weights, capacities	Shoppers buy on them; a wrong one means a return.
Certifications and standards	“Organic”, “vegan”, “BPA free”, “CE” are claims you must be able to prove.
Performance numbers	Battery life, burn time, waterproof ratings.
Origin and manufacturing	“Handmade”, “made in Italy”, “ethically sourced”.
Comparisons with other brands	Unverifiable and easy to get wrong.
Customer quotes or ratings	Never generate social proof.

The rule underneath all of them: every factual claim in the new text must already exist in the old text, the product fields or your own documents. The model may reorder, clarify and restructure. It may not add.

What should the AI be allowed to do?

Plenty, as long as it stays inside your facts:

• Lead with who the product is for and what it is used for, if your listing says so.
• Turn a wall of text into short paragraphs.
• Move specs that are buried in a paragraph into plain sentences an assistant can quote.
• Remove filler (“perfect for any occasion”, “you’ll love it”).
• Keep the language of your listing, not translate it into English.

How do you check a draft before it goes live?

Run these five checks on every AI written description:

1. Claim match. List every factual claim in the draft. Find each one in your original listing or product data. Anything you cannot find comes out.
2. Number match. Every number in the draft (size, weight, percentage, count) must match the source exactly, including units.
3. Category field match. If you fill category fields (material, colour, size) from the text, each value must be supported by the text, not guessed from the product type.
4. Language check. The draft is in the same language as the original listing.
5. Formatting check. If your old description had tables, images or custom HTML, decide whether you are willing to lose them before replacing it.

If a draft fails the claim check more than once, it is a sign your source listing is too thin. Add the real facts yourself first; then ask for a rewrite.

How do you roll back a bad rewrite?

Before you replace a description, keep the previous version somewhere you can restore it from in one step. Shopify’s admin has no one click way to bring back an earlier description, so a rewrite you saved by mistake is hard to undo unless you kept the old text. Keep the old version for at least a few weeks, long enough to notice a drop in conversion or a jump in returns.

How does Octolift handle this?

We built Improve with AI in Octolift around these rules, because a tool that writes to merchants’ stores cannot afford invented specs. Its suggestions keep every factual claim already in the listing and add no materials, measurements, certifications or other facts that are not there; category values are only picked when the title or description clearly supports them, from the values Shopify’s taxonomy allows.

Each draft is shown next to your current text, and nothing is written until you press Save. Every change is recorded with its previous value, and title and description changes undo in one click for 30 days. The details, including exactly which fields are written, are in Improve with AI.

It is still a language model. It can phrase a true fact awkwardly or stress the wrong thing, so read every suggestion as you would copy from a new writer.

What does a good rewrite look like?

Here is an example, built only from facts in the original:

• Before: “Our best base layer. Super soft and warm, perfect for the outdoors! Merino, 200g. Sizes XS-XXL. Slim fit, great for skiing and hiking.”
• After: “A slim fit base layer in 200 g/m² merino wool, made for cold weather hiking and skiing. It layers under a fleece or shell and comes in sizes XS to XXL.”

Every fact in the second version is in the first. Nothing about washing, odour or origin was added, because the original said nothing about them. If you want those in, they have to come from you.

Summary

Give the AI only your facts, forbid it from adding materials, numbers, certifications and claims, check every draft claim by claim, and keep the old version so you can roll back. A shorter description that is true beats a longer one that guesses, for shoppers and for the AI assistants that quote you. To see which of your listings are too thin to rewrite well, check how assistants read them with our guide to finding the questions shoppers ask AI, or try Improve with AI free on 5 products with Octolift.