Marketing AI Skill

Personalized Recommendations

Generate AI-powered product and content recommendations for cross-sell, upsell, and personalized experiences across email, web, app, and social channels. Use when building recommendation engines, personalizing product suggestions, creating cross-sell and upsell strategies, implementing collaborative filtering, or optimizing recommendation performance. Triggers on phrases like "product recommendations", "personalization", "cross-sell", "upsell", "recommendation engine", "collaborative filtering", "personalized content", "AI recommendations", "recommendation algorithm", "personalization strategy".

Personalized Product Recommendations

AI-powered recommendation engines for cross-sell, upsell, and personalized customer experiences.

Workflow

Collect and unify customer data: Purchase history, browsing behavior, demographic profile, engagement data.
Select recommendation algorithms: Collaborative filtering, content-based, hybrid approach.
Define recommendation contexts: Product page, cart, email, homepage, post-purchase, abandoned cart.
Build recommendation logic: Rules-based + AI-driven scoring for each context.
Implement recommendation widgets: On-site modules, email personalization blocks, in-app carousels.
Test and calibrate: A/B test recommendation algorithms, measure click-through and conversion rates.
Monitor performance: CTR, conversion rate, revenue lift, recommendation diversity.
Continuously optimize: Algorithm tuning, fresh data pipeline, seasonal adjustments.

Recommendation Algorithms

RECOMMENDATION ENGINE ARCHITECTURE
====================================

ALGORITHM TYPES:

  1. COLLABORATIVE FILTERING (User-Based):
     → "Customers who bought this also bought..."
     → Method: Find similar users based on behavior patterns
     → Strengths: Surprising discoveries, leverages collective wisdom
     → Weaknesses: Cold start problem (new users/products), data sparsity
     → Data needed: 1,000+ user-item interactions minimum
     → Best for: Mature catalogs with substantial transaction data
     → Implementation: Matrix factorization, k-NN, neural collaborative filtering

  2. COLLABORATIVE FILTERING (Item-Based):
     → "People who viewed X also viewed Y"
     → Method: Find similar items based on co-occurrence patterns
     → Strengths: Stable (item similarities change slowly), explainable
     → Weaknesses: Doesn't capture user preferences directly
     → Data needed: 500+ co-view/co-purchase pairs per item
     → Best for: Product recommendations on product detail pages
     → Implementation: Cosine similarity, adjusted cosine, association rules

  3. CONTENT-BASED FILTERING:
     → "Because you viewed [category], you might like..."
     → Method: Recommend items similar to what user liked (by attributes)
     → Strengths: Works for new users (no history needed), transparent
     → Weaknesses: Limited to user's existing preferences, no serendipity
     → Data needed: Product attributes and user interaction history
     → Best for: New users, niche products, content recommendations
     → Implementation: TF-IDF + cosine similarity, word embeddings

  4. HYBRID APPROACH (recommended):
     → Combine collaborative filtering + content-based + rules
     → Method: Weighted combination of multiple algorithm outputs
     → Strengths: Mitigates individual algorithm weaknesses
     → Weaknesses: Complex to implement and maintain
     → Best for: Production recommendation engines
     → Typical weights: 50% collaborative + 30% content + 20% rules

  5. DEEP LEARNING RECOMMENDATIONS:
     → Neural networks for recommendation (wide & deep, two-tower)
     → Method: Learn complex patterns from user-item interaction data
     → Strengths: Captures non-linear relationships, scales well
     → Weaknesses: Requires significant data and compute, black-box
     → Data needed: 10,000+ interactions, ML infrastructure
     → Best for: Large-scale platforms (Amazon-level catalogs)

RECOMMENDATION CONTEXTS AND STRATEGIES:

  PRODUCT DETAIL PAGE (PDP):
    → "Customers also viewed": Item-based collaborative filtering
    → "Frequently bought together": Association rules (market basket analysis)
    → "Similar products": Content-based (same category, similar attributes)
    → Display: 4-8 products in horizontal scroll or grid
    → Position: Below product description, above reviews

  SHOPPING CART:
    → "Complete your look": Complementary products (cross-sell)
    → "Upgrade option": Premium version of items in cart (upsell)
    → "Don't forget": Frequently co-purchased accessories
    → Display: 2-4 products (avoid overwhelming at checkout)
    → Position: Cart sidebar or below cart items
    → Incentive: Bundle discount for adding recommended item

  EMAIL RECOMMENDATIONS:
    → "Recommended for you": Personalized product picks based on browsing/purchase
    → "New in [category you love]": New arrivals in user's preferred categories
    → "Because you bought [product]": Complementary or replacement products
    → Display: 4-6 products in email block
    → Timing: Post-purchase (Day 7), browsing re-engagement (Day 3), win-back (Day 30)

  HOMEPAGE / DASHBOARD:
    → "Welcome back, [Name]": Resume browsing + personalized picks
    → "Trending in your interests": Popular items in user's categories
    → "Staff picks for you": Curated + personalized mix
    → Display: 6-12 products in multiple sections
    → Personalization: Category preference, price range, brand affinity

  POST-PURCHASE / THANK YOU PAGE:
    → "Customers who bought this also bought": Cross-sell opportunity
    → "Complete the set": Complementary products
    → Display: 3-6 products
    → Timing: Immediate (while purchase momentum is high)
    → Conversion rate: 3-8% (high-intent moment)

Implementation and Optimization

RECOMMENDATION IMPLEMENTATION
===============================

TECHNICAL STACK:

  ┌────────────────────┬─────────────────────────┬──────────────────────┐
  │ Solution           │ Best For                │ Key Features          │
  ├────────────────────┼─────────────────────────┼──────────────────────┤
  │ Dynamic Yield      │ Enterprise personalization│ Real-time, AI-driven │
  │ Bassistance        │ E-commerce recs          │ Product recommendations│
  │ Nosto              │ E-commerce personalization│ Cross-channel recs   │
  │ Barilliance        │ Mid-market recs          │ Easy setup, templates │
  │ Amazon Personalize │ AWS ecosystem            │ ML-powered, scalable  │
  │ Algolia Recommend  │ Search + recs combo      │ Fast, relevant        │
  │ IBM Watson          │ Enterprise AI            │ Deep learning recs    │
  │ Custom (TensorFlow │ Large-scale custom       │ Full control, complex │
  │  / PyTorch)        │ implementations          │                       │
  └────────────────────┴─────────────────────────┴──────────────────────┘

  BUILD vs. BUY DECISION:
    Buy if: < 50K products, limited ML team, need fast time-to-market
    Build if: > 100K products, dedicated ML team, unique recommendation logic
    Cost comparison: SaaS $500-$10,000/month vs. build $100K-$500K+ development

A/B TESTING FRAMEWORK:

  TEST 1 — Algorithm Comparison:
    Variant A: Collaborative filtering (also bought)
    Variant B: Content-based (similar products)
    Variant C: Hybrid (combined)
    Metric: CTR → Conversion rate → Revenue per session
    Duration: 14 days minimum, 500+ conversions per variant

  TEST 2 — Display Format:
    Variant A: Horizontal scroll (6 products)
    Variant B: Grid layout (4×2)
    Variant C: Carousel with auto-rotate
    Metric: CTR, scroll depth, time to click

  TEST 3 — Placement:
    Variant A: Below product description
    Variant B: In-product image gallery
    Variant C: Floating sidebar
    Metric: Viewability, CTR, impact on primary CTA

  TEST 4 — Volume:
    Variant A: 4 recommendations
    Variant B: 8 recommendations
    Variant C: 12 recommendations
    Metric: CTR (total), CTR (per item), conversion rate
    Finding: More items = higher total CTR but lower per-item CTR

RECOMMENDATION DIVERSITY:
  → Don't show same category repeatedly (cannibalization)
  → Mix price points (budget, mid, premium)
  → Include new products (exploration vs. exploitation)
  → Exclude recently purchased items (within 30 days)
  → Exclude out-of-stock items
  → Boost items with margin priority (business objective)

Performance Measurement

RECOMMENDATION PERFORMANCE METRICS
=====================================

CORE METRICS:

  Click-Through Rate (CTR):
    → Formula: Recommendation clicks / Recommendation impressions × 100
    → Benchmark: 5-15% (product page), 3-8% (email)
    → Target: > 10% for on-site, > 5% for email

  Conversion Rate:
    → Formula: Recommendation purchases / Recommendation clicks × 100
    → Benchmark: 10-30% of clicks convert
    → Target: > 15% for on-site, > 8% for email

  Revenue Attribution:
    → Recommendation-attributed revenue: Revenue from recommended product purchases
    → Revenue lift: % increase in revenue from recommendation vs. control
    → Benchmark: 10-30% revenue lift from recommendations
    → Amazon: 35% of revenue from recommendation engine

  Recommendation Quality:
    → Relevance score: % of clicks on top-3 recommendations (higher = more relevant)
    → Coverage: % of catalog that gets recommended (higher = more diverse)
    → Serendipity: % of clicks on unexpected but relevant items
    → Novelty: % of clicks on new/unfamiliar items

PERFORMANCE DASHBOARD:

  ┌────────────────────────┬──────────┬──────────┬──────────┬──────────┐
  │ Context                │ Impressions │ CTR    │ Conv Rate│ Revenue  │
  ├────────────────────────┼────────────┼────────┼──────────┼──────────┤
  │ Product Page           │ 245,000    │ 12.3%  │ 18.5%    │ $89,400  │
  │ Shopping Cart          │ 85,000     │ 8.7%   │ 22.1%    │ $42,300  │
  │ Email                  │ 120,000    │ 5.2%   │ 8.4%     │ $28,600  │
  │ Homepage               │ 310,000    │ 9.1%   │ 12.3%    │ $56,200  │
  │ Post-Purchase          │ 42,000     │ 15.8%  │ 25.0%    │ $31,500  │
  │ ───────────────────────┼────────────┼────────┼──────────┼──────────┤
  │ TOTAL                  │ 802,000    │ 10.1%  │ 15.9%    │ $248,000 │
  └────────────────────────┴────────────┴────────┴──────────┴──────────┘

  Revenue per thousand impressions (RPK): $309.23
  Overall recommendation ROI: 8.2x (revenue / implementation cost)

Integration Points

Dynamic Yield / Nosto / Barilliance: Recommendation engine SaaS, on-site personalization, email recommendations, A/B testing
Amazon Personalize / Algolia Recommend: Cloud-based ML recommendations, real-time personalization, API integration
Shopify / WooCommerce: Native recommendation apps, product feed integration, cart recommendations
Google Analytics 4: Recommendation click tracking, conversion attribution, revenue measurement
Klaviyo / Mailchimp: Email recommendation blocks, behavioral triggering, personalization tokens
TensorFlow / PyTorch: Custom recommendation model development, deep learning algorithms
Apache Spark / Hadoop: Large-scale recommendation computation, batch processing
Redis / MongoDB: Real-time recommendation caching, session-based recommendations

Edge Cases

Cold start problem: New users with no history, new products with no interactions
New user: Use content-based recommendations (popular in category, trending items)
New user: Ask preference questions during onboarding (explicit feedback)
New product: Boost in category recommendations, associate with similar established products
Hybrid: 50% popular/trending + 50% personalized (gradual shift as data accumulates)
Timeline: 2-4 weeks of interaction data needed for meaningful personalization

Filter bubble and over-personalization: Users only see what algorithm predicts they'll click
Risk: Reduced discovery, catalog coverage inequality (same products always recommended)
Solution: Diversity constraint (min 3 different categories per recommendation set)
Solution: Exploration bonus (10-20% random/serendipitous recommendations)
Solution: Periodic reset (don't lock users into historical patterns)
Monitoring: Track catalog coverage metric weekly

Seasonal and inventory changes: Recommendations referencing out-of-stock or seasonal products
Real-time inventory sync: Exclude out-of-stock items from recommendations
Seasonal rules: Exclude winter items in summer (and vice versa)
Expiry handling: Exclude expired/promotional items past end date
Fallback: If personalized pool < 4 items, fill with trending/popular
Cache invalidation: Recommendation cache refresh every 15-60 minutes

Disclaimer: All rights reserved by Circulos AI. These skills are specifically designed for Claude Code, Claude Cowork, Codex, and OpenClaw. When using or referencing any skill, please provide proper attribution to Circulos AI.