AI product descriptions: how to scale content without losing quality

A store with 5,000 SKUs and no product descriptions is leaving money on the table. Google cannot rank pages with no content. Customers cannot make purchase decisions without details. But writing 5,000 unique descriptions manually takes months. AI makes this feasible in days. The challenge is not generating text. It is generating text that is accurate, SEO-friendly, brand-consistent and actually useful to the customer. This guide covers the practical workflow from raw product data to published descriptions.

Why descriptions matter more than you think

Product descriptions serve three audiences simultaneously. Search engines: unique, keyword-rich descriptions are the primary ranking factor for product pages. Duplicate or missing descriptions mean Google has no reason to rank your page over competitors. Customers: descriptions answer questions that product images cannot. Material composition, sizing details, compatibility, use cases. Good descriptions reduce returns by setting accurate expectations. Conversion rate: A/B tests consistently show that detailed descriptions improve add-to-cart rates by 10-30%. The effect is strongest for products that are not self-explanatory (supplements, technical products, fashion with specific materials). Yet most stores have either no descriptions, manufacturer copy (duplicated across every retailer) or thin 2-sentence descriptions that help nobody. This is the gap AI fills.

Tools: Claude API, Sanity AI Assist and Shopify Magic

Claude API: the most flexible option for bulk generation. You control the prompt, output format and integration. Supports structured output (JSON) so you can generate title, short description, long description, meta description and bullet points in a single API call. Pricing is per token: approximately $0.003-0.015 per description depending on length and model. Best for stores with custom CMS or headless architectures. Sanity AI Assist: built into Sanity Studio. Content editors can generate and refine descriptions directly in the CMS interface. The AI has context from other document fields (title, category, tags) which improves relevance. It supports custom instructions per field, so you can define different generation rules for meta descriptions versus long descriptions. Ideal for teams already using Sanity who want AI in their existing editorial workflow. Shopify Magic: free, built into Shopify admin. Enter product features and tone, get a description. Fast and frictionless for individual products. Limitations: no bulk generation, limited control over output structure, inconsistent brand voice across large catalogs. Best for: small stores adding products one at a time. GPT-4 and other models: similar capabilities to Claude but different strengths. Claude tends to produce more structured, factual output. GPT-4 can be more creative. Test both with your product data and choose based on output quality for your specific catalog.

The bulk generation workflow

Step 1 - Export product data: extract from your platform (VTEX Catalog API, Shopify Admin API, or CMS) a structured export with: SKU ID, product name, category, brand, attributes (size, color, material, weight), existing images and current description (if any). CSV or JSON format. Step 2 - Design the prompt template: a good prompt includes: role (You are a product copywriter for [brand name], a [brand description] store), brand voice guidelines (professional but approachable, avoid jargon, use active voice), SEO instructions (include primary keyword naturally in first paragraph, target 150-300 words), output structure (return JSON with fields: short_description, long_description, meta_description, bullet_points), and example input/output pair. Step 3 - Process in batches: use the Claude API to process 50-100 products per batch. Include error handling for rate limits and failed generations. Store results in a staging database (Supabase works well for this). Step 4 - Quality review: sample 10-20% of generated descriptions for manual review. Check for: factual accuracy (does the description match the product attributes?), brand voice consistency, SEO keyword inclusion, and any hallucinated features. Step 5 - Import to platform: use VTEX Catalog API, Shopify Admin API or Sanity's mutation API to update product descriptions in bulk. Run in staging/preview first. Typical timeline: 5,000 product descriptions generated in 2-3 hours of API processing, 2-3 days of sampling review, 1 day of import and verification. Total cost: $50-250 in API fees.

Prompt engineering for e-commerce descriptions

The prompt determines 80% of output quality. Generic prompts produce generic descriptions. Specific prompts produce useful content. Bad prompt: Write a product description for this item. Good prompt: You are the product copywriter for [Brand], a premium Brazilian fashion brand targeting women aged 25-40. Write a product description for the following item. Tone: confident, modern, inclusive. Structure: start with the occasion or use case (when would someone wear this?), then materials and construction details, then fit and sizing guidance. Include the primary keyword [keyword] naturally in the first sentence. Length: 150-200 words. Do not mention competitors. Do not use superlatives without justification. Return as JSON with fields short_description (max 50 words) and long_description. Advanced techniques: few-shot prompting (include 2-3 examples of ideal descriptions in the prompt), attribute-conditional instructions (if the product is clothing, include fit guidance; if electronics, include compatibility info), and negative instructions (never claim the product is the best or number one; never invent features not in the input data). Test your prompt with 20 diverse products before running the full catalog. Iterate on the prompt based on the worst outputs, not the best ones.

Integrating with your CMS and platform

VTEX integration: use the Catalog API endpoint PUT /api/catalog/pvt/product/{productId} to update product descriptions. The ProductDescription field accepts HTML. For bulk updates, batch API calls with rate limiting (VTEX allows approximately 40 requests per second on the Catalog API). Consider using VTEX's Data Import/Export spreadsheet for very large catalogs. Sanity integration: use the Sanity Client library to create or patch documents. For articles and product pages managed in Sanity, the mutation API lets you update specific fields without overwriting the entire document. Workflow: generate descriptions via Claude API, store in a staging collection in Sanity, editors review and approve, approved descriptions are published. This keeps AI generation in a review pipeline rather than publishing directly. Shopify integration: the Admin REST API or GraphQL API updates product descriptions via the body_html field. For Shopify Plus, use the Bulk Operations API for large catalogs. Headless with Next.js: if your frontend is Next.js with ISR or on-demand revalidation, updating descriptions in your CMS automatically triggers page rebuilds. No additional deployment needed. The key principle: AI generates, humans review, automation publishes. Never connect AI output directly to production without a review step.

Pitfalls: what goes wrong and how to avoid it

Hallucinated features: AI sometimes invents product attributes. A cotton shirt becomes a premium Egyptian cotton blend. Always validate generated descriptions against the source product data. Include only factual attributes in the prompt input. Generic voice: without strong brand guidelines in the prompt, AI descriptions sound identical across products and brands. Invest time in defining your brand voice with specific examples, not vague adjectives. SEO keyword stuffing: AI can over-optimize if instructed to include keywords. Set a maximum keyword density and instruct the model to prioritize readability over keyword placement. Duplicate patterns: when processing thousands of products, AI tends to reuse sentence structures. Vary your prompts slightly (rotate between 3-4 prompt templates) and use different opening structures. Legal compliance: in Brazil, product descriptions must comply with Codigo de Defesa do Consumidor. Health claims, performance guarantees and comparative claims have legal implications. AI does not know your legal constraints. Add explicit negative instructions (do not make health claims, do not guarantee results, do not compare to competitors by name). Ignoring existing content: if you already have descriptions for some products, do not overwrite them blindly. Compare AI-generated quality against existing descriptions. Sometimes the existing human-written description is better. Use AI to fill gaps, not replace good content.