What Data Do AI Product Recommendations Need to Work?
Introduction
Personalisation in Asian retail has moved from a competitive edge to a basic expectation. Whether shoppers are browsing an online store, receiving a promotion through a messaging app or scanning a loyalty QR code at the counter, they expect product suggestions that reflect their preferences.
For retail decision-makers, including CEOs, COOs and IT heads, using artificial intelligence for personalisation sounds promising. However, many recommendation engines fail to deliver a strong return on investment (ROI), not because the algorithms are weak, but because the data feeding them is incomplete, inconsistent or out of sync.
AI models do not work in isolation. They rely on continuous feeds of customer, product and operational data. This guide explains the core data that AI product recommendations need in order to drive cross-selling, larger baskets and faster stock turnover across physical and digital stores.
The Core Data Behind AI Product Recommendations
To generate accurate "Frequently Bought Together", "You Might Also Like" or personalised reorder prompts, recommendation engines draw on three main streams of retail data: customer behaviour, product attributes and live inventory.
When these streams are not synchronised, AI product recommendations quickly become irrelevant. Typical examples include suggesting items that are out of stock, promoting a product the customer bought minutes ago, or showing seasonal goods at the wrong time of year.
1. Customer Behaviour and Transaction Data
Behavioural data tells the AI who the customer is and what they are likely to buy next. Recommendation engines typically look at two types:
Transaction history (POS and online): Past receipts, purchase frequency, average order value (AOV), customer lifetime value (CLV) and brand preferences. This data powers market basket analysis and collaborative filtering.
Real-time browsing and in-store signals: Items viewed, cart additions, search terms, abandoned carts, time spent on category pages, and store visits recorded through loyalty card scans or CRM.
Real-time intent modelling stops recommendations from going stale. If a customer adds a coffee machine to their online cart, the engine should immediately switch from suggesting other coffee machines to complementary items such as coffee beans and descaling solution.
2. Structured Product Attribute Data
If customer data shows who is buying, product attribute data explains what is being bought. AI models use content-based filtering to group similar products and identify meaningful relationships between them.
Product Data Field | Purpose | Example |
Hierarchy and category | Primary, secondary and tertiary product classification | Beverages > Plant-Based Milk > Oat Milk |
Functional attributes | Brand, volume, size, colour, ingredients, dietary tags and price band | Organic, gluten-free, 1 litre, premium |
Margin and velocity tags | Internal data that guides promotion priority | High margin, fast-moving, markdown candidate |
Substitution mapping | Linked alternative SKUs for when an item is out of stock | Brand A Oat Milk to Brand B Oat Milk |
Without rich product data, AI product recommendations cannot identify sensible pairings or offer suitable alternatives when a first-choice item is unavailable.
3. Real-Time Inventory and Operational Data
Recommending a product that is out of stock at the customer's nearest store or fulfilment centre creates frustration and damages trust. Effective recommendation engines therefore need a direct link to back-office inventory systems:
Omnichannel stock levels: Live visibility of warehouse and store inventory, so recommendations are limited to items customers can actually buy.
Local pricing and promotions: Recommended items should reflect active store promotions, loyalty tier discounts or bundle prices.
Supplier lead times: Slow-to-replenish items can be deprioritised to protect stock for core demand.
Data Governance: PDPA and Clean Transaction Records
In Malaysia, using customer data for personalisation must comply with the Personal Data Protection Act 2010 (PDPA). Retailers should build consent management for behavioural tracking into their data pipeline across both online and in-store touchpoints. The Personal Data Protection Department publishes the latest guidance and requirements.
Clean transaction records matter too. Since the same POS and sales data feeds both recommendation models and the Inland Revenue Board of Malaysia's (LHDN / HASiL) e-invoicing process, master data such as customer IDs and buyer Tax Identification Numbers (TINs) should flow accurately into accounting ledgers without corrupting customer profiles. A single, well-governed data source supports both personalisation and compliance.
Building a Data Foundation for AI Product Recommendations
Before investing in a recommendation engine, it helps to check your data readiness:
Are POS, e-commerce and loyalty transactions linked to a single customer record?
Is your product catalogue tagged consistently by category, attribute and price band?
Can your systems share live stock levels with customer-facing channels?
Do you have clear customer consent for using purchase and browsing data?
If the answer to any of these is no, fixing the data first will do more for your results than switching algorithms.
How QR Retail Automation (QRRA) Brings Your Retail Data Together
QR Retail Automation (Asia) Sdn Bhd (QRRA) helps Malaysian retailers turn fragmented operational data into usable business intelligence and personalised campaigns.
AI product recommendations as a service: QRRA's Product Recommendation as a Service (PRaaS) combines repeat purchase prediction, RFM-based member segmentation and market basket analysis to suggest relevant items for each customer. It sits as a layer on top of your existing retail systems, without replacing them.
A single source of truth: QRRA's Data & Analytics solution and AgoraCloud Merchandising & Material Management (ACMM) bring store POS, online orders and warehouse stock into one central view.
Built-in e-invoicing: AgoraCloud E-Invoice (ACE) validates transactions with LHDN's MyInvois system while keeping sales records clean and audit-ready.
MSME tax incentive note: Micro, small and medium enterprises can claim a tax deduction of up to RM50,000 per year of assessment for consultancy fees incurred to implement e-invoicing, from year of assessment 2024 to 2027. Check with your tax adviser which incentives apply to your wider system upgrade.
Frequently Asked Questions
What data do AI product recommendations need?
They need customer behaviour and transaction data, structured product attributes, and live inventory and pricing data.
Why do some recommendation engines show irrelevant products?
Usually because the data is incomplete or out of sync, so the engine suggests out-of-stock items or products the customer has just bought.
Can AI product recommendations work in physical stores?
Yes, when loyalty scans and POS receipts are linked to customer records, recommendations can power in-store offers and follow-up promotions.
Do I need customer consent to use their data for recommendations?
Yes. In Malaysia, using personal data for personalisation must comply with the PDPA, including proper consent management.
Do I need to replace my systems to use AI product recommendations?
Not necessarily. Services such as QRRA's PRaaS sit on top of existing retail systems.
Power Your Personalisation Strategy with Clean Data
AI product recommendations are only as good as the data behind them. By bringing customer behaviour, product attributes and real-time inventory together in one platform, your business can grow average basket sizes and improve customer retention.
To see how this would work for your stores, request a demo with QR Retail Automation and book a retail data assessment with the team.
Comments