Recommendations trained on your customers, not Shopify's.
Custom engine using real behavior data. Collaborative + content-based hybrid. A/B tested. Monthly retainer delivery.
Who this is for
DTC or e-commerce operator at $5M–$50M scale where Shopify's default recommendations underperform and full recommendation vendors (Nosto, Klevu) cost too much at current margin.
The pain today
- Shopify Related Products picking wrong items 60% of the time
- Klaviyo recommendations generic, not tuned to your catalog
- Nosto/Klevu vendor fees $500–2,000/mo with unclear ROI
- No visibility into which algorithms are driving revenue
- Can't test new recommendation hypotheses without vendor change orders
The outcome you get
- Custom recommendation engine trained on your real purchase and browse data
- Hybrid algorithm: collaborative filtering + content similarity + session-aware
- A/B testing built in — measure lift vs current recommendations
- Integration with Shopify or BigCommerce via official APIs
- Continuous learning — model retrains weekly on fresh data
Recommendation types (collaborative, content, hybrid)
Three approaches to recommendations. Collaborative filtering: 'customers who bought X also bought Y' — needs purchase data, works best on high-volume catalogs. Content-based: 'this product is similar to that product' — needs rich product metadata (category, materials, price, style), works for long-tail or cold-start (new product with no purchase history). Hybrid: combines both with session context (what's the user looking at right now) — best performer in production for most catalogs. Modern engines add LLM-based semantic similarity for matching intent to product when metadata is thin. I pick the combination based on your catalog and data volume.
Shopify and BigCommerce integration
Shopify: App Store app using Storefront API and Admin API. Theme integration via Shopify-native 'apps blocks' or direct Liquid embedding. Behavior tracking via Shopify's Web Pixel API (privacy-compliant). BigCommerce: Stencil theme integration with BigCommerce APIs. Both: real-time recommendations rendered server-side for SEO, with client-side personalization layered on. Rendering approach: edge-cached common recommendations (category pages) + personalized last-mile layer (user-specific recommendations on PDP and cart). This pattern keeps LCP fast while still showing personalized content.
A/B testing and revenue attribution
Recommendations must prove ROI. Every algorithm change A/B tested against control. Metrics tracked: CTR on recommendation tiles, add-to-cart rate from recommendations, revenue per session attributed to recommendations, revenue lift vs control. Attribution model: last-click within session counts the recommendation as source of purchase. Dashboard: real-time A/B results, win/loss detection at 95% confidence, automatic rollout of winners. I build the measurement infrastructure alongside the recommendation engine because 'we added recs' without measurement is indistinguishable from placebo.
Case study: Imohub search at scale
Imohub is a real estate portal with 120k+ property records. The matching engine — finding the right property for a buyer's criteria — shares technical DNA with product recommendation engines. Sub-0.5-second query response across 120k items. 70% infrastructure cost reduction vs legacy stack. Meilisearch for keyword-and-semantic search, Next.js frontend, Laravel backend. The same stack — fast search over rich catalog, semantic matching, scalable infrastructure — applies to product recommendations. The engineering principles don't change because the catalog is cars, homes, or products.
Pricing
E-commerce recommendation engines fit the AI Automation retainer at $3,000/mo. First-version timeline: 5–7 weeks from kickoff to A/B testing against current recommendations. Retainer continues through algorithm tuning, seasonal adjustments, and expansion (recommendations on email, recommendations on post-purchase). 14-day money-back, cancel anytime, Work Made for Hire. LLM or ML infrastructure costs typically $100–1,000/mo depending on catalog size and recommendation volume.
When Nosto or Klevu is enough
Nosto, Klevu, Bloomreach, and Algolia Recommend cover most e-commerce recommendations well at $500–2,000+/month. Custom is worth building when your catalog or business model doesn't fit their standard approach (configurable products, rental marketplaces, service bookings), when margin pressure makes vendor fees prohibitive, or when you need specific algorithms they don't offer (cross-sell logic tuned to your economics). I'll say honestly in the first call — Nosto is strong out of the box; custom earns its cost when generic doesn't fit.
Recent proof
A comparable engagement, delivered and documented.
Rebuilt a real estate portal at a fraction of the cost
Rebuilt Imóveis SC's real estate portal as ImoHub — a faster, more scalable successor — handling 120k+ properties with sub-second search and drastically reduced AWS costs.
Frequently asked questions
The questions prospects ask before they book.
- How much data do I need?
- For collaborative filtering, 1,000+ monthly purchases across 100+ unique SKUs. Below that, content-based recommendations using product metadata perform better. Cold-start for new products handled with content similarity. Most DTC stores at $1M+/year have enough data for meaningful personalization.
- What lift should I expect?
- Typical 5–15% lift in revenue per session on pages where recommendations are surfaced. AOV lift of 3–8%. Specific numbers depend on catalog, current baseline, and traffic quality. I share targeted lift estimates after reviewing your data in week 1, not as generic promises.
- Can it handle seasonal catalogs?
- Yes — models retrain weekly on fresh data, so seasonal shifts (summer inventory, holiday categories) reflect in recommendations within a week. Explicit seasonal tuning (Black Friday, Christmas, back-to-school) configurable as campaign-specific rules overriding standard recommendations.
- Does it work for subscription or rental products?
- Yes — subscription recommendations use 'likely next subscription' logic based on cohort behavior. Rental recommendations factor availability and category. Both are common use cases that generic vendors handle poorly, making custom worthwhile.
- What about email recommendations?
- Email recommendations are natural extensions — same engine, rendered at email-send time via API call, images and links prepared for email clients. Klaviyo, Customer.io, and Iterable integrations standard. Email recs typically drive higher revenue per impression than on-site recs due to intent signal of email open.
Ready to start?
Tell me what you need in 60 seconds. Tailored proposal in your inbox within 6 hours.