SEO/GEO Content for E-commerce — AI Article Generation Experiment
SEO/GEO Content for E-commerce
TL;DR: Testing whether AI-generated product articles can scale e-commerce content while maintaining quality for both Google SEO and AI search citation. Results from pigu.lt show the approach produces publish-ready content with minimal editing.
The Business Problem
E-commerce sites face a content scaling challenge:
| Challenge | Traditional Approach | Cost |
|---|---|---|
| 10,000+ products need descriptions | Hire content writers | €5-15 per article |
| Product specs change frequently | Manual updates | Ongoing labor |
| SEO requires unique, quality content | Can’t use manufacturer copy | Writing time |
| AI search engines need citable content | Requires specific structure | Expertise |
| Multiple languages (LT, RU, EN) | Multiple writers per language | 3x cost |
The hypothesis: AI can generate product articles that are:
- SEO-optimized (title, meta, headings, schema markup)
- GEO-optimized (citable by AI search engines like Perplexity, ChatGPT)
- Human-sounding (avoids AI tell-signs that hurt trust)
- Publish-ready with minimal editing
The Approach
We built a tools/product-article-generator skill that:
- Scrapes product data from the retailer’s existing product pages
- Analyzes context — category, target audience, search intent
- Generates article with enforced structure:
- GEO anchor intro (first sentence answers directly)
- Honest assessment section (real weaknesses = AI trust)
- Self-contained FAQ (quotable by AI engines)
- Complete schema markup (Product, Article, FAQ)
- Applies human writing rules — blocked phrases like “game-changer,” “dive into,” etc.
- Outputs pre-publish checklist — ensures nothing is missed
Key Design Decisions
Why “honest assessment”? AI search engines preferentially cite sources that acknowledge weaknesses. An article that says “this freezer is perfect” gets skipped. An article that says “the 40 dB noise level may be noticeable in open-plan kitchens” gets cited as trustworthy.
Why native language, not translation? Lithuanian readers detect translated content instantly. The skill generates directly in Lithuanian, matching local idioms and product terminology (e.g., “šaldymo dėžė” not “freezing box”).
Why schema markup? Both Google and AI search engines use structured data for understanding and citing content. Product schema helps with shopping results; FAQ schema enables direct answers.
Test Results: pigu.lt
Test subject: Hisense FC184D4AWLYE freezer article for pigu.lt (Lithuanian e-commerce)
Input
- Product URL from pigu.lt
- Target: Lithuanian homeowners shopping for freezers
- Output language: Lithuanian
Output Quality Assessment
| Dimension | Rating | Notes |
|---|---|---|
| Factual accuracy | 9/10 | All specs match source; price verified |
| SEO structure | 10/10 | Title 47 chars, meta 148 chars, proper H-structure |
| GEO optimization | 9/10 | Intro answers “what is this?” directly; FAQ self-contained |
| Human voice | 8/10 | Natural Lithuanian; minor rhythm improvements possible |
| Schema completeness | 10/10 | Product, Article, FAQ schemas present and valid |
| Publish-readiness | 8/10 | Needs: real customer quote, hero image, internal links |
Time comparison:
- Human writer (from scratch): ~2-3 hours
- AI generation + human review: ~20-30 minutes
- Speedup: ~5-6x
Generated Article Excerpts
The GEO Anchor (Intro)
The glossary/geo-anchor is the most critical section for AI citation. Here’s what the skill generated:
Original Lithuanian:
“Hisense FC184D4AWLYE šaldymo dėžė yra 142 litrų talpos laisvai pastatomas šaldiklis, skirtas šeimoms ir individualiems naudotojams, kuriems reikia papildomos šaldymo vietos be didelio biudžeto. Ji veikia vos 40 dB triukšmo lygiu, turi elektroninį valdymą ir vidaus apšvietimą.”
Translation:
“The Hisense FC184D4AWLYE freezer is a 142-liter freestanding chest freezer designed for families and individuals who need extra freezing space without a large budget. It operates at just 40 dB noise level, has electronic controls and interior lighting.”
Why this works for AI citation:
- Product name in sentence one (primary keyword)
- Exact capacity (142 liters) — not “large” or “spacious”
- Target audience explicitly named (families, individuals needing extra space)
- Value proposition clear (budget-friendly)
- Specific features (40 dB, electronic controls, lighting)
An AI answering “What is the Hisense FC184D4AWLYE?” can cite this sentence directly.
The Honest Assessment
The glossary/honest-assessment pattern increases AI trust signals:
Original Lithuanian:
“Stiprioji pusė — tai kainos ir kokybės santykis. Už ~184–240 € gaunat 142 litrų talpą, elektroninį valdymą, vidaus apšvietimą ir tikrai tylų veikimą.
Silpnoji pusė — energijos klasė E. Tai reiškia 183 kWh per metus, o tai nėra pats efektyviausias rezultatas šioje kategorijoje. Per metus tai kainuos apie 40–50 €, priklausomai nuo jūsų elektros tarifo.”
Translation:
“The strong side is price-to-value ratio. For ~€184–240 you get 142 liters, electronic controls, interior lighting, and genuinely quiet operation.
The weak side is the E energy rating. That means 183 kWh per year, which isn’t the most efficient result in this category. Per year that costs about €40–50, depending on your electricity rate.”
Why this works:
- Names a real limitation (E energy rating)
- Provides specific cost impact (€40-50/year, 183 kWh)
- Doesn’t fake balance (“the only downside is it’s too good”)
- Helps readers self-select
Self-Contained FAQ
Each FAQ answer is structured to be quotable without surrounding context:
Question: “Ar Hisense FC184D4AWLYE galima laikyti garaže?” (Can the Hisense be kept in a garage?)
Answer:
“Taip. Šaldiklio klimato klasė yra SN-N-ST-T, o tai reiškia, kad jis skirtas veikti aplinkos temperatūroje nuo +10 °C iki +43 °C. Garažas, sandėliukas ar rūsys tinka puikiai, jei žiemą temperatūra nenukrenta žemiau +10 °C.”
Translation:
“Yes. The freezer’s climate class is SN-N-ST-T, which means it’s designed to operate in ambient temperatures from +10°C to +43°C. A garage, storage room, or basement works well, as long as winter temperatures don’t drop below +10°C.”
Why this works for AI:
- Direct answer (“Yes”)
- Self-contained explanation (doesn’t need context)
- Specific operating range (+10°C to +43°C)
- Practical guidance (where it can be placed)
What Required Editing
- Hero image — uploaded and compressed to WebP
- Internal links — added to related products and category page
- Price verification — confirmed current at publish time
- Customer quotes — sourced from actual reviews when available
Total editing time: ~15 minutes
Business Impact
For pigu.lt (ongoing client)
| Metric | Before | After |
|---|---|---|
| Articles per week | 3-5 (agency-written) | 15-20 (AI + review) |
| Cost per article | €10-15 | €2-3 (review time only) |
| Schema markup coverage | Inconsistent | 100% |
| AI search citations | Unknown | Tracking started |
Scaling Considerations
What scales well:
- Single product reviews (Module A)
- Categories with standardized specs (appliances, electronics)
- Languages with good model coverage (Lithuanian, Russian, English)
What needs more work:
- Comparison articles (Module B) — require curator judgment on “best” picks
- Fashion/lifestyle products — specs less important than styling
- Niche categories — may lack model knowledge
Limitations Observed
- Price freshness — Article captures price at generation; needs update workflow
- Customer quotes — Must be sourced from real reviews; can’t fabricate
- Visual content — Images must be added separately
- Competitor context — Doesn’t know what else is on the market unless told
- Local availability — May not know regional stock status
Verdict
Does AI article generation solve the e-commerce content scaling problem?
Yes, with conditions:
- ✅ Works well for spec-heavy products (appliances, electronics, tools)
- ✅ Produces SEO/GEO-optimized content consistently
- ✅ Reduces cost per article by ~80%
- ✅ Human review still required (15-30 min vs. 2-3 hours)
- ⚠️ Not “set and forget” — needs price updates, real quotes, images
- ⚠️ Fashion/lifestyle categories need different approach
ROI calculation (pigu.lt scale):
- 100 articles/month × (€12 saved per article) = €1,200/month savings
- Plus: consistent schema markup, GEO optimization, faster time-to-publish
Key Takeaways
- GEO optimization requires honesty — AI engines cite sources that acknowledge limitations
- Native generation beats translation — write directly in target language
- Schema markup is table stakes — both Google and AI search use it
- Human review is non-negotiable — AI generates draft, humans verify and polish
- Specs-heavy products scale best — fashion/lifestyle needs different approach
Related
- cases/product-article-generator-pigu — Full case study on building this skill
- tools/product-article-generator — The skill documentation
- glossary/geo-anchor — The intro paragraph pattern
- glossary/honest-assessment — The AI trust signal pattern
- glossary/geo-aeo — The optimization discipline
- seo/ai-seo-content — Content strategy for AI search
- experiments/ai-visibility-ecommerce — Measuring AI search visibility
Sources
- Product Article Generator skill (Primores internal, April 2026)
- pigu.lt production deployment (ongoing client work)
- Hisense FC184D4AWLYE test article output