Innovative AI Solutions | AI Development, Web & Mobile Apps – Delhi, India
E-COMMERCE · AI RECOMMENDATION · D2C BRAND

AI Recommendations That Drive 28% Higher Order Value

We built a personalized AI recommendation engine for a D2C skincare brand with 2,000+ SKUs — showing the right product to the right customer at the right moment, increasing AOV by 28% and generating 35% of total revenue through AI recommendations.

28%Higher Average Order Value
35%Revenue from AI Recs
18%Better Conversion Rate
40xROI in First Year
Build Recommendation Engine All Case Studies

India's D2C Beauty Market and the Personalisation Imperative

India's direct-to-consumer beauty and skincare market has grown to over ₹1,200 crore and is expanding at 25% year-on-year — one of the highest growth rates of any consumer category in the country. The market is led by digitally-native brands that have built loyal customer bases through social media, influencer partnerships, and a compelling product story. However, as these brands scale from 50 SKUs to 500 to 2,000+, they face a critical challenge that traditional beauty retail also struggled with: product discovery. A customer who comes to your store for a moisturiser needs to find the serum, SPF, and toner that complete her routine — but without active guidance, she will only ever buy what she specifically searched for. This "search and leave" behaviour is the core challenge of D2C e-commerce at scale. Amazon, which pioneered algorithmic recommendations, publicly attributes 35% of its total revenue to its recommendation engine — a figure the industry has accepted as a benchmark for what personalisation can achieve. Yet most Indian D2C brands, including this client, were running recommendation-free stores showing the same "Best Sellers" block to every visitor. Cart abandonment rates in Indian D2C e-commerce average 68–74% — among the highest in the world — partly because customers don't discover relevant products and don't feel a strong enough "complete your routine" pull to stay and purchase more. AI recommendation engines address this at every stage of the purchase journey.

💄

₹1,200 Cr D2C Beauty Market, 25% YoY Growth

India's D2C beauty and personal care market is one of the fastest-growing consumer segments in the country, driven by digital penetration, aspirational middle class expansion, and social media beauty culture. Brands that win in this market are those that can personalise the customer experience at scale — and AI recommendation is the primary lever for this personalisation.

🛒

68–74% Cart Abandonment in Indian D2C

Indian D2C shoppers abandon carts at significantly higher rates than global averages (68.5% globally). Key reasons include incomplete product discovery — customers don't find complementary products that would make the purchase feel more complete — and the lack of contextual guidance that a physical store beauty advisor would provide. AI recommendations directly address both problems.

📈

Amazon Attributes 35% Revenue to Recommendations

Amazon's recommendation engine is the most-studied example of AI-driven commerce personalisation. The 35% revenue attribution figure — confirmed across multiple Amazon earnings analyses — represents the commercial ceiling that best-in-class recommendation AI can achieve. Indian D2C brands that implement comparable personalisation systems are not just copying a tactic; they're applying the most commercially validated technology in e-commerce history.

2,000 SKUs, Zero Personalization

The D2C brand had built an impressive product portfolio over five years — 2,000+ SKUs spanning skincare (moisturisers, serums, face washes, toners, eye creams, SPF), haircare (shampoos, conditioners, hair oils, treatments), body care, and wellness supplements. The portfolio depth was a genuine competitive advantage: for almost any skin concern or haircare need, the brand had a targeted solution. The problem was that customers couldn't find the solutions that were most relevant to them — and the brand's Shopify store was not helping.

The homepage showed the same eight "Best Sellers" to every single visitor — the 45-year-old woman with dry, mature skin saw the same recommendations as the 22-year-old with oily, acne-prone skin. The moisturiser that was perfect for one was actively wrong for the other. The product detail pages had no "you might also need" or "complete your routine" logic — a customer who added the brand's vitamin C brightening serum to cart had no indication that vitamin C and their retinol product should not be used together (a real dermatological interaction), and that a niacinamide alternative would better fit their existing routine. This ingredient compatibility gap was not just a missed upsell opportunity — it was a customer experience problem causing confusion and potentially adverse reactions.

Post-purchase email performance told the same story. The CRM team sent weekly newsletters to the full customer database with the same product features and offers for everyone. The 0.8% CTR on these emails was barely distinguishable from spam — customers were conditioned to ignore them because they were irrelevant. A customer who bought an anti-ageing serum received emails featuring acne treatments; a customer in Delhi received emails promoting a monsoon haircare kit during the dry winter season. The disconnect between what the brand's products offered and what customers were being shown was the fundamental revenue gap this project needed to close.

AI That Knows What Each Customer Needs Next

A multi-signal recommendation engine trained on purchase history, browsing behaviour, skin concern quiz answers, and product ingredient relationships — serving the right product to each customer at every touchpoint.

01

Collaborative + Content Filtering

The recommendation engine uses a hybrid architecture that combines two fundamentally different types of intelligence. Collaborative filtering — "customers like you also bought" — uses purchase history patterns across the customer base to identify non-obvious product relationships. If customers who buy the vitamin C serum overwhelmingly also purchase a specific niacinamide moisturiser within 30 days, the model learns this pairing from population data even without understanding why it works. Content-based filtering adds a second intelligence layer: each product is represented as a vector of ingredient profiles, skin concern targets, usage step (AM/PM/weekly treatment), skin type suitability, and price tier. This content model enables recommendations even for new products with no purchase history and prevents ingredient conflicts — the system actively understands which product pairs are complementary versus incompatible. The LightFM library was selected for the hybrid collaborative-content model because it trains simultaneously on both interaction data and item features, achieving superior cold-start performance for new product launches compared to pure collaborative filtering approaches.

02

Skin Concern Quiz Integration

First-time visitors who haven't purchased yet present a cold-start challenge — the collaborative filter has no purchase history to work with. We designed a 5-question skin concern quiz that appears as a non-intrusive overlay on the first product page view: skin type (oily/dry/combination/normal/sensitive), primary concern (acne/pigmentation/ageing/hydration/brightening), climate/location (hot and humid/dry/cold), current skincare routine complexity (beginner/intermediate/advanced), and budget preference. Quiz completion rates reached 41% among first-time visitors — well above the industry average for similar onboarding flows — because the quiz was designed to feel like a personalised consultation rather than a data collection exercise. Quiz answers immediately seed a personalisation profile, allowing the recommendation engine to serve relevant content from the very first page view. Customers who complete the quiz have a 2.8x higher average order value in their first purchase than those who don't.

03

In-Cart and Post-Purchase Upsell

The highest-intent moment in any e-commerce journey is the cart — a customer who has added a product is actively considering a purchase decision. We designed four recommendation placements around this moment: a "Complete Your Routine" widget that appears when a product is added to cart, showing the top 3 complementary products identified by the content model as step-appropriate additions; a "Customers Who Bought This Also Got" widget using collaborative filtering signals; a "You'll Use These Together" bundle suggestion with bundle pricing when 2 compatible products are in cart; and a post-order confirmation "Your Next Step" email sent 48 hours after delivery, recommending the logical next product based on what was purchased and typical skincare routine sequencing. The in-cart routine completion widget drove the majority of the AOV uplift — when a customer adds a face wash and sees that 78% of customers who bought that face wash also buy the matching toner and moisturiser, the routine completion logic is compelling and the purchase makes genuine sense to the customer.

04

Personalized Email Sequences

Klaviyo integration replaced the brand's generic broadcast email strategy with a personalised sequence engine. Instead of weekly newsletters sent to all customers, the system generates individual recommendation packages for each customer based on: purchase history and typical replenishment timing per product category (moisturisers typically replenish every 45–60 days; serums every 60–75 days — the system times "are you running low?" emails precisely), next logical routine step based on current products owned, skin concern evolution (customers who have addressed acne might be moving to pigmentation — the system identifies this transition from browsing signals), and seasonal relevance (sunscreen recommendations increase in pre-summer, intensive moisturisers in winter, anti-humidity haircare in monsoon). This contextual, individually-timed email approach drove CTR from 0.8% to 6.2% — a 7.75x improvement — and email revenue from effectively zero to a meaningful channel contributing ₹42 lakhs annually in incremental revenue.

Six-Month Build to Full Production

M1

Shopify Data Export and Purchase History Analysis

Month 1 was devoted to understanding the data landscape. We exported 3 years of Shopify order history (148,000 orders, 94,000 unique customers), product catalogue with full metadata, customer browsing session data from Google Analytics, and email engagement history from Klaviyo. Purchase pattern analysis revealed several non-obvious insights: 34% of customers made their second purchase in the same category as their first (loyal category buyers), 28% cross-category purchased within 45 days, and 12% of customer pairs (identified by household address) shared purchase patterns suggesting gifting behaviour. These segments required different recommendation strategies — the data analysis phase shaped the model architecture decisions that followed.

M2

Collaborative Filtering Model Training (LightFM)

Month 2 built and validated the collaborative filtering foundation. LightFM was configured with user features (demographics from quiz data where available), item features (ingredient profile, skin concern target, category, price tier), and interaction data (purchase events weighted higher than browse events). The model was trained on 80% of historical orders and validated on the held-out 20%. Key metrics: precision@10 (how many of the top 10 recommended products did customers actually buy) reached 0.31 — better than comparable D2C implementations. The ingredient compatibility constraint layer was also built in Month 2: a knowledge graph of 340 known ingredient interactions (conflicts and synergies) was encoded so the model never recommends Vitamin C + Retinol as a same-routine pair, or promotes AHA after a chemical sunscreen that reduces efficacy.

M3

Skin Concern Quiz Development and Content Model

Month 3 built the user-facing quiz and the content model that powers cold-start recommendations. The quiz was designed through 3 rounds of A/B testing on completion rate and downstream purchase conversion — the final 5-question format outperformed earlier 8-question and 3-question variants. The content model encoded all 2,000+ SKUs as semantic vectors using product descriptions, ingredient lists, and skin concern mappings created with input from the brand's in-house dermatologist. This dermatologist-validated ingredient knowledge base is a key differentiator — it ensures recommendations are medically sound, not just statistically correlated.

M4

Shopify Widget Integration and Klaviyo Email

Month 4 was the integration sprint. The recommendation API (FastAPI backend serving Redis-cached recommendations) was integrated into Shopify via a custom widget that injected recommendation carousels into 5 page types: homepage, product detail page, cart page, order confirmation page, and collection pages. Widget load time was optimised to under 200ms through Redis caching of pre-computed recommendations for all customers with purchase history. Klaviyo integration used the Klaviyo API to push personalized product recommendation arrays to each customer profile, which Klaviyo's email builder then renders into the brand's email template. The total integration work was completed in 22 working days.

M5–6

A/B Test, Full Launch, and Retraining Pipeline

Month 5 ran a rigorous A/B test: 50% of Shopify sessions received AI recommendations, 50% continued to see the existing best-sellers. After 4 weeks and 32,000 sessions per variant, AI recommendations showed statistically significant improvements across all metrics (AOV +26%, conversion rate +16%, revenue per visitor +38%). Month 6 completed the full rollout and built the automated retraining pipeline: the LightFM model retrains weekly on the latest 6 months of order data, Klaviyo segments update nightly, and Redis recommendation cache refreshes daily. The system operates autonomously — no manual intervention required for ongoing optimisation.

Revenue Impact After 90 Days

💰

28% Higher AOV

Average order value grew from ₹1,240 to ₹1,587 — a ₹347 increase per order driven primarily by in-cart routine completion suggestions and bundle recommendations. The AOV uplift was consistent across all customer segments (new, returning, and lapsed-reactivated) but was highest among customers who completed the skin concern quiz (AOV ₹1,820 vs ₹1,587 average). Customers who purchased AI-recommended bundles had a 2.1x higher 90-day repeat purchase rate than customers who purchased single items — indicating that routine completion recommendations drive genuine long-term loyalty, not just short-term upsell.

🛒

35% Revenue from AI Recs

Within 90 days of full rollout, 35% of total site revenue came from products first discovered through AI recommendations — up from effectively 0% before implementation. This figure includes both immediate purchase of recommended products and deferred purchases where customers browse a recommended product, exit, and return to buy within 30 days (tracked via Shopify attribution). The 35% revenue attribution figure is particularly striking given that the brand's Amazon listing (which does not use our recommendation engine) showed no comparable growth in the same period — isolating the recommendation AI as the causal driver.

📧

Email CTR 8x Higher

Personalized post-purchase email sequences achieve an average 6.2% CTR versus 0.8% for previous broadcast newsletters — a 7.75x improvement. Open rates improved from 18% to 34%, reflecting the improved subject line relevance when email timing and product focus are personalised. Email-attributed revenue grew from approximately ₹5.4 lakhs annually to ₹47 lakhs annually — a complete transformation of email as a revenue channel. The highest-performing email flow is the 45-day replenishment reminder for moisturisers and cleansers, which achieves 9.1% CTR and an 18% conversion rate on click — meaning nearly 1 in 5 customers who click a replenishment reminder complete a purchase.

40x First-Year ROI

The recommendation engine's total Year 1 cost including build, LightFM model training, API infrastructure, Klaviyo integration, and ongoing monitoring was ₹8 lakhs. Incremental Year 1 revenue attributable to AI recommendations — AOV uplift across 8,000 monthly orders plus email revenue — reached ₹3.20 crore. This generates a 40x return on investment in the first full year of operation. The compounding nature of recommendation AI (the model improves as it collects more purchase data) means Year 2 ROI is projected even higher with no incremental system cost — only the ongoing API and infrastructure fees of approximately ₹2.4 lakhs annually.

The Revenue Math Behind This Investment

The financial case for this recommendation engine rests on two primary revenue drivers: the AOV uplift from in-cart routine recommendations, and the email channel revival from personalised post-purchase sequences. A third driver — improved conversion rate from recommendation-assisted product discovery — adds an additional layer. Together, these three streams generate exceptional returns on a modest system cost.

Revenue DriverCalculationAnnual Value
AOV uplift (₹347/order × 8,000 orders/month)₹347 × 8,000 × 12₹2.78 Cr
Email channel revenue (personalised sequences)₹42L/year incremental₹42 L
Conversion rate improvement (18% uplift on 8,000 visitors/month)1,440 additional orders × ₹1,240 AOV × 12
Total incremental annual revenue₹3.20 Cr
System build + Year 1 operating cost₹8 L
Return on Investment — Year 140x ROI

Technologies Used

Pythonscikit-learnLightFM (collaborative filtering)FastAPIShopify APIRedis (real-time recs)Klaviyo IntegrationPostgreSQLA/B Testing Framework

About This Project

How do recommendations work for new visitors with no purchase history? +
New visitors enter what recommendation practitioners call the "cold start" phase — the system has no user-specific interaction history to work with. We address this through a multi-layer fallback strategy. First, the skin concern quiz (which 41% of first-time visitors complete) immediately seeds a content-based profile — even one quiz response ("I have oily, acne-prone skin") is enough to serve highly relevant recommendations from the product content model. Second, for visitors who don't complete the quiz, we use real-time session signals: products browsed in the last 5 pages, category paths followed, search terms used, and dwell time on specific product types. Within 3 product page views, these real-time signals give the content model enough information to serve meaningfully personalised suggestions. Third, for visitors who bounce after a single page, demographic-based trending products (inferred from device type, geographic location, and time of day) provide a non-personalised but contextually relevant fallback. The system never shows generic best sellers to anyone who has provided any signal at all.
Does it work for non-Shopify stores? +
Yes — the recommendation engine is platform-agnostic by design. The core system is a FastAPI microservice that exposes a simple REST API: send a customer ID and context (current page, cart contents, session signals) and receive a ranked list of recommended product IDs with confidence scores. Integration with any e-commerce platform requires only a lightweight JavaScript snippet (under 15KB, non-blocking load) that handles the front-end display, and a server-side integration for purchase event tracking. We have implemented this on WooCommerce, Magento 2, custom React storefronts, and mobile apps (React Native and Flutter). Integration timelines vary: 2 days for Shopify via our existing plugin, 3–5 days for WooCommerce, and 1–2 weeks for fully custom platforms. The same recommendation engine also supports mobile push notification personalisation via Firebase and WhatsApp recommendation sequences via WhatsApp Business API.
How do you ensure recommendations don't hurt the customer experience? +
Customer experience protection is built into both the model architecture and our deployment methodology. At the model level, we implement three safeguards: the ingredient compatibility constraint prevents conflicting product recommendations (no Vitamin C + Retinol in the same routine suggestion, no AHA/BHA on top of a compromised skin barrier product); a price proximity filter ensures recommendations stay within ±30% of the current cart value unless the customer has demonstrated willingness to purchase at higher price points; and a diversity constraint prevents the same product category from appearing more than twice in any recommendation set. At the deployment level, we A/B test all recommendation placements and require statistically significant improvement in both revenue metrics AND customer satisfaction scores (measured via post-purchase NPS) before any placement becomes permanent. Two placements we tested — a pop-up recommendation overlay and an exit-intent recommendation modal — were rejected because they improved AOV slightly but reduced NPS. Customer trust is a longer-term asset than short-term AOV.
How does the model stay accurate as the product catalogue changes? +
Catalogue freshness is maintained through an automated retraining and enrichment pipeline. When a new SKU is added to Shopify, it is automatically ingested into the content model within 24 hours — the product's ingredient list, skin concern targets, and category are encoded into the feature vector using our standardised taxonomy. The collaborative filtering model retrains weekly on the latest 6 months of purchase data; this weekly cadence means that rapidly-selling new launches quickly acquire collaborative signal and appear in recommendations within 2–4 weeks of launch, without requiring any manual catalogue management. Discontinued products are automatically removed from the recommendation pool when their Shopify status changes to "archived." For limited-edition or low-stock products, we implemented inventory-aware recommendations: products below a configurable stock threshold are de-ranked in recommendations to prevent driving traffic to out-of-stock outcomes.
What role does the skin type quiz play compared to purchase history? +
The quiz and purchase history serve complementary roles in the recommendation system, and their relative weight shifts over a customer's lifecycle. For new customers (0–1 purchases), the quiz is the primary signal — it provides explicit, high-quality information about skin type and concerns before any purchase history exists. For returning customers (2–5 purchases), the system blends quiz data with collaborative signals — what their purchase pattern says about them may be more nuanced than what they self-reported in the quiz (e.g., a customer who said "oily skin" but consistently buys hydrating products may actually have dehydrated skin that presents as oily, a common misconception). For loyal customers (6+ purchases), purchase history and browsing patterns dominate — the model knows them better than the quiz at this point. This lifecycle-aware weighting is one of the design decisions that separates a professional recommendation system from a simple quiz-based recommender.
How do you measure the true incremental impact of recommendations versus organic discovery? +
Attribution is the hardest measurement challenge in recommendation systems, and we approach it through a rigorous A/B testing framework rather than simple click-attribution. Our A/B framework assigns customers (not sessions) to treatment and control groups at the Shopify cookie level, ensuring the same customer doesn't see recommendations in some sessions but not others. Revenue comparison is at the customer level over 30-day windows, not at the session level — this captures deferred purchases where a recommendation influences a product that is bought in a later session. For email, we use Klaviyo's holdout group functionality: 10% of customers are withheld from all personalised emails so that email revenue lift can be measured against a true control. This multi-layered attribution approach gives the brand confidence that the revenue numbers are genuinely incremental — not re-attributed revenue from customers who would have bought anyway.

Services Used in This Project

Increase Your AOV with AI Recommendations

Get a free analysis of your product catalogue and purchase data. We'll show you the recommendation opportunities you're missing.

Get Free Recommendation AI Demo