Product Article Generator — AI Content at E-commerce Scale

Product Article Generator — AI Content at E-commerce Scale

TL;DR: A Claude skill that transforms product URLs into publish-ready articles optimized for both Google (SEO) and AI search engines (GEO). Deployed at pigu.lt, producing 15-20 articles/week at ~80% cost reduction vs. human writers. The key insight: AI engines cite content that admits weaknesses.

The Problem

E-commerce content has a scaling paradox:

Requirement	Challenge
10,000+ products need unique content	Human writers can’t scale
SEO requires quality, not just quantity	Template-based content ranks poorly
AI search engines need specific structure	Traditional SEO doesn’t work for GEO
Multiple languages (LT, RU, EN)	Each language multiplies cost
Specs change, prices update	Content goes stale quickly

The old math: €10-15 per article × 10,000 products = €100,000-150,000 just for initial content, not counting updates.

The hypothesis: AI can generate content that satisfies SEO, GEO, and human readers simultaneously, at a fraction of the cost.

Why This Skill Matters

This isn’t just another “AI writes content” story. Three insights make it critical:

1. SEO + GEO Convergence

Traditional SEO optimizes for Google’s ranking algorithm. But AI search engines (Perplexity, ChatGPT, Gemini) work differently — they extract and cite sentences, not rank pages.

The skill implements both:

• SEO signals: Title tags, meta descriptions, heading hierarchy, schema markup
• GEO signals: glossary/geo-anchor intros, glossary/honest-assessment sections, self-contained FAQ answers

Content optimized only for Google won’t get cited by AI. Content optimized for AI may not rank on Google. The skill does both.

2. The Honest Assessment Pattern

Counter-intuitive insight from AI engine behavior: admitting weaknesses increases citations.

AI engines are trained on review sites (Wirecutter, Consumer Reports) that include balanced analysis. Content that only praises products gets flagged as promotional and skipped.

The skill enforces an “Honest Assessment” section in every article:

• What the product does well (specific, with evidence)
• One real limitation (specific, with cost/impact)
• Who will be disappointed (the anti-persona)

This single pattern increases both AI citations AND reader trust.

3. Native Language Generation

Translation produces awkward content. The skill generates directly in the target language:

• Detects product page language automatically
• Uses native idioms and product terminology
• Matches regional conventions (Lithuanian formal address, etc.)

Lithuanian readers detect translated content instantly. Native generation doesn’t feel like AI.

How It Works

Input

• Product URL (or multiple URLs for comparisons)
• Target language (auto-detected from page)

Workflow

URL → Scrape → Analyze → Generate → QA → Output

Step 1: Scrape & Analyze

• Product name, category, price, variants
• Key specs with actual numbers
• Target audience (inferred from page copy)
• Existing description (to avoid repetition)
• Customer reviews if visible
• Brand tone (premium, casual, clinical, playful)

Step 2: Keyword Strategy (Silent)

• Primary keyword: product name + benefit
• 3-5 secondary keywords: long-tail, question-based
• Semantic entities: related concepts AI associates with this
• Search intent: discovery / comparison / ready-to-buy

Step 3: Generate Article Structure

Section	Purpose	GEO Signal
Meta (title, description)	SEO ranking	Character-limited for SERP
H1	SEO + scanning	Keyword + outcome
Intro (GEO Anchor)	AI citation	Direct answer, sentence 1
What It Is	Context	Plain-language explanation
Key Benefits	Value proposition	Benefit → outcome, with specifics
Who It’s For	Targeting	2-3 specific buyer profiles
How to Use	Utility	Numbered steps + tips
Honest Assessment	Trust signal	Real weakness named
Social Proof	Conversion	Customer quotes or placeholder
Verdict	Decision	Clear recommendation
FAQ	GEO citation	Self-contained answers

Step 4: Schema Markup

• Product schema (sku, brand, offers, aggregateRating)
• Article schema (author ≠ brand — critical for trust)
• FAQ schema (matches visible content exactly)

Step 5: QA Checklist

• Character counts (title ≤60, meta ≤155)
• Schema integrity
• AI visibility signals
• Internal link placeholders

Output

Complete markdown article + JSON-LD schemas + publishing checklist

Design Decisions

Why “Human Writing Rules” Are Mandatory

AI content has tell-signs that hurt trust. The skill blocks these phrases:

Blocked	Why
”Dive into”, “delve into”	AI clichés
”Game-changer”, “revolutionary”	Empty hype
”In today’s fast-paced world”	Obvious opener
”Seamlessly”, “effortlessly”	Unverifiable
”Comprehensive”, “robust”	Filler words

The skill also enforces:

• Varied sentence length (rhythm)
• Direct address (“you” not “one”)
• One mild opinion per section (human signal)
• No padding (stop when done)

Why Author ≠ Brand in Schema

A subtle but critical rule:

// ❌ WRONG
"author": { "@type": "Organization", "name": "Hisense" }

// ✅ CORRECT
"author": { "@type": "Person", "name": "Redakcija" }

Setting author: "Hisense" on an article about Hisense products creates a trust conflict. Google cross-references brand entities. The author must be the shop editor, not the product manufacturer.

Why Two Modules

Module A: Single Product (800-1200 words)

• Input: 1 product URL
• Output: Complete review article

Module B: Comparison (1200-2000 words)

• Input: 2-10 product URLs
• Output: Ranked comparison with Quick Verdict box
• Key rule: Name the winner in the intro (AI engines reward this)

Comparison articles have higher GEO value — they answer “which is best” questions directly.

Results at pigu.lt

Efficiency

Metric	Before	After
Articles per week	3-5	15-20
Cost per article	€10-15	€2-3
Time per article	2-3 hours	20-30 min
Schema coverage	Inconsistent	100%

Speedup: ~5-6x Cost reduction: ~80%

Quality Assessment (Hisense Freezer Article)

Dimension	Rating	Notes
Factual accuracy	9/10	All specs verified
SEO structure	10/10	Title 47 chars, proper hierarchy
GEO optimization	9/10	Strong anchor, self-contained FAQ
Human voice	8/10	Natural Lithuanian
Schema completeness	10/10	All three schemas present
Publish-readiness	8/10	Needs images, real customer quote

What Still Requires Humans

1. Customer quotes — must be sourced from real reviews
2. Images — must be uploaded and compressed
3. Price verification — scraped at generation time, needs check at publish
4. Internal linking — specific links added by editors

Total human time: ~15-30 minutes per article (vs. 2-3 hours writing from scratch)

Key Takeaways

1. SEO + GEO must be solved together — optimizing for only one loses the other
2. Honesty increases citations — AI engines trust balanced content
3. Native generation beats translation — readers and AI both detect awkward phrasing
4. Schema markup is table stakes — both Google and AI engines use it
5. Human review remains essential — AI generates drafts, humans verify and polish
6. Author ≠ brand in schema — subtle but critical for E-E-A-T

Limitations

• Spec-heavy products work best — appliances, electronics, tools
• Fashion/lifestyle needs different approach — specs matter less than styling
• Price freshness requires workflow — need update process for changing prices
• Customer quotes can’t be fabricated — must wait for real reviews
• Visual content still manual — images must be added separately

Related

• tools/product-article-generator — The skill documentation
• experiments/seo-geo-content-ecommerce — Experiment validating this approach
• glossary/geo-anchor — The intro pattern
• glossary/honest-assessment — The trust signal pattern
• glossary/geo-aeo — The GEO/AEO discipline
• seo/ai-seo-content — Content strategy for AI search
• automation/advisor-strategy — The cost structure pattern

Sources

• Product Article Generator skill (Primores internal, 2026)
• pigu.lt production deployment (ongoing)
• Hisense FC184D4AWLYE test article output