How to generate ecommerce product descriptions with AI

What this automation does

Most catalogs have the same problem: copy pasted from the manufacturer, half the specs missing, and a backlog of new SKUs waiting for someone to write them. This workflow separates the facts from the writing, so the AI writes well without making things up.

1. Gather source data. Supplier spec sheets (PDF, Excel, CSV), your ERP or PIM export, existing product data in the store and product images.
2. Build an attribute sheet. The model extracts each product’s facts into a schema for its product type: materials, dimensions, weight, capacity, compatibility, care, what is in the box and certifications. Missing attributes are listed as missing, never filled with a guess.
3. Write every field. From the attribute sheet and your voice guide, the model writes the product title, description, bullet points, a specs table, SEO title, meta description and alt text for each image. Variant differences (color, size) are written per variant.
4. Validate in code. Every number in the copy must appear in the attribute sheet. Length limits are enforced per field and channel. A restricted-claims list blocks medical, environmental, origin and superlative claims unless the attribute sheet records support for them.
5. Review. A merchandiser approves products in a review sheet. Higher-risk categories, such as supplements, cosmetics, children’s products and electrical items, always get a person’s eyes.
6. Publish. Approved copy is written to Shopify, WooCommerce or BigCommerce through their APIs, as drafts or direct updates, with metafields for structured specs.
7. Refresh. When supplier data changes or return reasons point to a problem (“smaller than expected”), the affected products are regenerated and re-reviewed.

When it is worth building

Build this when you have hundreds of SKUs or more, a steady flow of new arrivals, product pages still running manufacturer copy, or return reasons that say the product was not as described.

Skip it for a store with a few dozen products. Shopify’s built-in AI text generator, or a chat assistant with a good prompt, will do for occasional products. Also skip it if your copy is the product, such as a small brand whose founder’s voice is the reason people buy. A custom workflow pays off on volume, structured specs, multiple channels and the claim checks, not on the writing alone.

Architecture

How to build it

1. Pick one product type with clear attributes and decent volume, for example a single apparel category or one line of tools.
2. Define the attribute schema for that type with the merchandising team, including which fields are required to publish.
3. Extract attributes from supplier files and existing data, and have a person fix the attribute sheet for the first batch. This is where most of the quality comes from.
4. Write the voice guide and examples. Pull your best existing descriptions, list banned words and phrases, and set field lengths per channel.
5. Build generation and validation. Generate all fields in one call with structured output, then run the number, length and claims checks. Failures go back for regeneration once, then to a person.
6. Set up review. A sheet or simple queue with before and after, the attribute sheet alongside, and approve or edit per product.
7. Create new products as drafts and hold changes to live products for approval, then switch to direct updates once edits per product drop. Tag any hand-written hero products so the workflow never overwrites them.
8. Measure against a control collection that keeps its old copy for a few weeks.

Lumeor’s Shopify AI copilot project worked with the same store data and APIs, and the Shopify tool page covers its API limits. Real estate teams use the same confirmed-facts approach in our listing description guide.

What breaks in production

• Invented specs. The model adds “waterproof” because similar products are. The number and claims checks catch most of this, and attribute-only prompting catches the rest.
• Variant mix-ups. Copy mentions a color that is sold out or does not exist. Generate variant text from the variant data, not the parent product.
• Overwritten hero copy. A bulk run replaces a description someone spent a day on. Lock fields by tag before any bulk update.
• Missing metafield definitions. Writes fail or land as untyped data. Create the metafield definitions in the store before the first run.
• Rate limits during bulk updates. Large catalogs hit API limits. Use bulk operations and batch with backoff.
• Sameness at scale. Two thousand descriptions with the same structure read like a template. Vary openings and structure by product type, and sample-check every batch.
• Translations. Machine-translated copy for other markets needs a native reviewer for at least the first batch.

Timeline and cost

Most catalogs are running within one to three weeks. Lumeor builds this in the range of $3,000 to $10,000. The low end is one store, one or two product types and descriptions plus SEO fields. The high end adds supplier file extraction, many product types, multiple channels including Amazon flat files, metafields and image alt text. Model usage is usually cents per product, so even a full catalog rewrite is a small line item. See what workflow automation costs for how we price these.

How to measure success

• Time from product receipt to live product page
• Share of SKUs with complete required attributes
• Conversion rate on rewritten collections versus the control
• Returns with “not as described” or sizing reasons
• Organic search impressions and clicks on product pages
• Merchandiser edits per description, which should fall over the first few batches