AI Product Descriptions at Scale: Prompt Templates for E-commerce

Why AI for Product Descriptions?

Mid-sized stores commonly carry 2,000-5,000 SKUs. Manually writing 150-200 word descriptions means 15-20 SKUs per day per copywriter — 6 months for 3,000 SKUs. AI with well-designed templates can draft 3,000 in 2 days; human editors approve on day 3. What matters is the loop: prompt chain + human review + SEO check.

Core Prompt Architecture

1. Role: "You are an e-commerce copywriter…"
2. Context: Brand tone, target audience, positioning
3. Input: Product specs (JSON or bullet list)
4. Constraints: Char limits, keywords, banned words
5. Output format: Title, summary, features, SEO meta

Template 1: Fashion & Apparel

You are an e-commerce copywriter. Tone: friendly, young, trend-focused.
Product: {product_name}
Features: {list}
Audience: 22-35, mid-to-upper segment women
Write:
- 60-char SEO title
- 155-char meta description
- 3-paragraph product story
- Styling suggestion
Avoid: "best", "amazing", hype adjectives

Template 2: Electronics & Technical

Role: technical product writer. Tone: objective.
Product: {product}
Specs (JSON): {specs}
Write:
- Use case (who, when, why)
- Comparison to previous model
- Spec table
- 5-question FAQ (for FAQPage schema)

Template 3: Food & Cosmetics (Regulated)

For health-claim products always add constraints: "do not use cures, guarantees, or regulated claims". Human approval is mandatory for compliance.

Quality Control Checklist

1. Hallucination check: Never make up specs — always inject source JSON
2. Monotony: Check first 100 descriptions for varied sentence structure
3. Keyword density: 1.5-2.5% optimal; > 4% is stuffing
4. Readability (Flesch-Kincaid): 60-70 for broad audience
5. Compliance: Health claims, restricted phrases (drug regulations, food labeling)

Automation Pipeline

1. Product data (CSV/JSON) → prompt template
2. OpenAI/Anthropic API batch generation (Batch API saves 50%)
3. Human editor reviews 1 in 10 products, refines prompt
4. SEO scan: keyword density via Yoast/similar
5. A/B test: AI copy vs old copy — conversion delta

Cost Comparison

Common Mistakes

1. Treating AI as a copy-paste tool without human editor
2. Ignoring keyword strategy
3. Same opening template for all 500 products (duplicate signal)
4. No product specs in prompt (hallucinations)

FAQs

Will Google penalize AI text?

No. Helpful Content criteria focuses on user value, not source. Poor quality control causes penalties, not AI itself.

Which LLM is best for multilingual?

Claude Sonnet and GPT-4o are excellent for multiple languages. For cost efficiency, Claude Haiku 4.5 leads.

Should I AI-generate product photos?

Catalog photos must show the actual product (consumer protection). AI for background/lifestyle shots can add value.

Next Step

Build your AI product-copy pipeline — schedule a consultation.