Start with one workflow that costs the most manual time. Prove value there before expanding.
The problem is not AI. The problem is the prompt. Generating a description by sending the product name and a few bullet points to an AI model produces output calibrated to the average of everything the model has ever seen. Average descriptions do not convert.
A description that converts does three specific things. It names the problem the customer has before they find the product. It explains how the product addresses that problem in concrete terms. And it addresses the most common objection a potential buyer would have. These three elements require information about your customer that is not in the product name or the spec sheet.
Building the right prompt
Before writing a prompt, gather four pieces of information for each product: who buys it (not demographics, but the specific situation they are in when they decide to purchase), what problem they are solving, what they are most likely to hesitate about, and what a satisfied customer says about it (from reviews, if they exist).
The prompt then takes this form: you are writing a product description for [audience in situation]. The product is [name]. It solves [problem] by [mechanism]. The most common customer concern is [objection]. A customer who loves this product says [review excerpt]. Write a description of [length] that speaks directly to the customer’s situation, addresses their concern, and ends with a clear reason to add to cart.
A prompt with this level of specificity produces output that requires much less editing than a generic prompt. It also produces output that sounds like your brand rather than like a template.
A prompt with this level of specificity produces output that requires much less editing than a generic prompt.
Working on something similar?
Let's talk →Handling a large catalog
Working on something similar?
Let's talk →For a Shopify store with hundreds of products, the manual information-gathering step does not scale. The practical solution is to build product archetypes. Group your products into five to ten categories where the audience, problem, and objection are similar. Write the prompt information once per archetype. Run the AI against all products in each archetype.
The output for products within an archetype will be structurally similar, which is acceptable. Individual differences come from the product-specific details in the spec sheet. Review the output for each archetype with one or two sample products to calibrate the prompt before running it across the full group.
What to test after publishing
AI-generated descriptions need the same testing any copy change does. In Shopify, product page conversion rate (add-to-cart rate) is the primary metric. If you update descriptions across a category, compare the add-to-cart rate for that category in the period before and after, controlling for traffic changes.
eCommerce conversion is affected by more than description copy, images, price, reviews, and page load time all matter. A description change that shows no improvement might be fine copy being dragged down by other factors. Isolate variables before drawing conclusions.
