Use AI to write product descriptions: 6 prompts and the field limits for Amazon, Etsy, eBay and Google

What should you give AI before it writes a product description?

Give it four things: the spec sheet, the buyer, the platform's limits and a list of claims it may not make. Anything a model writes beyond the specs is a guess. On a product page, a guess is a return. Paste this into ChatGPT, Claude or Gemini:

```

Write a product description for [platform: Amazon, Etsy, eBay, Shopify or Google Shopping] of about [word count, e.g. 120] words.

Specs: [paste the spec sheet: materials, size, weight, capacity, colors, what's in the box, care, country of origin]

Buyer and the job it does for them: [one sentence]

The two things a buyer should notice first: [features]

Limits: [the title and description limits from the table below]

Use only the specs. If a buyer would expect a detail I didn't give, write [NEEDED: detail] instead of guessing. No certifications, awards, health claims, prices, shipping promises or "customers love it". After the description, list every claim next to the spec line it came from.

```

The last line is the one that saves you. A model that has to cite a spec for each claim can't slip in "BPA free" or "fits most adults" without a blank showing up next to it. Then hand it the limits, because each marketplace counts differently, and the number a field allows isn't the number a shopper sees.

We read each platform's own rules on 25 October 2026. Etsy and eBay don't publish a character target below their limits, so they sit in the table only:

One rule in that table is easy to miss. Google's description help page says: "For descriptions created using generative AI, use the structured description [structured_description] attribute." The title page says the same for titles, with the digital source type set to trained_algorithmic_media. Send both the plain and the structured attribute and Google uses only the plain one, so a feed that carries both never shows the flag.

Can I use AI to write product descriptions?

Yes, and the biggest platforms build it in. Amazon starts a listing from "a few words, an image, or a URL" (Sell on Amazon, 3 November 2025). Shopify's description generator needs "a product title and at least 2 product features or keyword items" and works in every language Shopify supports. eBay announced a tool on 7 September 2023 that writes titles and descriptions from a photo.

Sellers use them heavily. Amazon reports that more than 900,000 selling partners have used its generative AI listing tools, with sellers "accepting AI-generated content with little to no edits approximately 90% of the time" (post of 19 September 2024, updated 8 May 2025).

Every platform leaves the accuracy with you. Shopify's help page puts it plainly: "You're responsible for the accuracy of all of the content that you publish to your store." Amazon says the same about listings it drafts. Google Search has said since 8 February 2023 that "Appropriate use of AI or automation is not against our guidelines" (Google Search Central). Its line is content made mainly to manipulate rankings.

So the built-in tool is the right start for one listing at a time. A chat model earns its place when you want control over the inputs: a supplier's spec sheet, your brand voice, 20 rows at once, or a second model checking the first.

Which ChatGPT prompts write the rest of the listing?

Five more prompts cover the rest of a listing: titles, bullets, bulk rows, brand voice and a fact check. Each one fills one field, which is why there are six and not forty. Start with the title, since it's the field with the hardest limit.

```

Write 5 product titles for [platform]. The limit is [limit] characters with spaces, and the target is [shorter target].

Product: [what it is, material, size, color]

Words buyers search for: [keywords, or "none yet"]

Put the words a buyer types first. Use each word at most twice. No ! $ ? _ { } ^ symbols, no "sale", no "best", no gift phrases. Show the character count after each title, and mark any title within 5 characters of the limit.

```

Ask for the count, then check it yourself. Models read text in tokens, not letters, so their character counts drift. Paste each title into a character counter before it goes live.

Bullets work best when they answer what buyers already ask. Amazon's listing guide (29 April 2026) allows "up to five bullets", so feed the model your real questions:

```

Here are the questions buyers ask about this product, from messages, reviews and Q&A: [paste]

Here are the specs: [paste]

Write up to 5 bullet points, one per most asked question. Open each bullet with the answer and back it with a spec. Where the specs don't answer a question, skip it and list it under "Questions I need to answer".

```

For a catalog, paste rows instead of one product. Keep each batch small enough to read line by line, because you'll read every row before it goes live:

```

Below are [number] products as CSV rows with the columns [name, material, size, color, notes]. For each row, write a title under [limit] characters and a description of about [word count] words, using only that row.

Return CSV with the columns handle, title, description, needed. Put [NEEDED: detail] in the needed column whenever a row lacks something a buyer would expect. Leave prices, sale dates and shipping out of the description.

```

That last line isn't style. Google's description rules say "Don't include promotional text such as price, sale price, sale dates, shipping, delivery date", and they ban comparisons with other products too.

Brand voice comes from your own copy, not from adjectives like "playful". Give the model three descriptions you like:

```

Here are 3 product descriptions I wrote and like: [paste]

Describe their voice in 5 rules: sentence length, person, words I use, words I avoid, how I open. Then rewrite the description below so it follows those rules, without adding any fact: [paste]

```

Save the five rules. In a chat app with saved instructions they ride along on every new chat, so you don't paste them forty times. Last, check the copy in a fresh chat, on a different model if you have one:

```

Here is a spec sheet and a product description written from it: [paste both]

List every claim in the description. Mark each one SUPPORTED, with the spec line, or UNSUPPORTED. Then list anything a buyer could hold me to in a return or a complaint.

```

A fresh chat has no stake in the copy it's checking. For ready versions of these and five more store prompts, see the ecommerce prompt pack.

How do you stop AI from inventing product details?

You stop it by leaving it nowhere to guess: every claim points to a spec line, and every gap becomes [NEEDED]. Three habits do most of the work.

Ban the claims that cost money by name. Certifications, safety ratings, health benefits and "eco" words are where an invented detail turns into a complaint, so list them in the prompt as banned unless the spec sheet says them.

Keep promotion out of the fields that forbid it. Amazon's title rules ban the same word more than twice and the symbols ! $ ? _ { } ^ ¬ ¦ unless they're part of the brand name (Amazon Global Selling notice). Google bans price, shipping and comparisons in a description. Put those rules in the prompt, then check for them in the output.

Then read the output against the box, not the screen. The color in the photo, the size on the tag and the count in the box are the last check, and only you have the box.

Should you use your platform's AI or a chat model?

Use your platform's AI for single listings and a chat model for everything that needs your inputs. The built-in tools are free with the store and write straight into the field. A chat model writes from your spec sheet, holds a voice across a catalog and lets a second model check the first.

Whizi, which publishes this page, sits in the second group, and it loses the first comparison: it doesn't publish anything to Shopify, Amazon or Etsy, so you paste the copy into the field yourself. What it adds is GPT, Claude, Gemini and 280+ other models in one chat. Its projects hold up to 32,000 characters of instructions and up to 10 pinned files, CSV and .xlsx among them, so the voice rules and the spec sheet ride along on every chat. On Pro, side by side model chats put two models' drafts next to each other.