Data is often called the new oil, but we’d like to argue that algorithms are the golden jewel. The impact of personalization and recommendation algorithms on business performance is staggering.

For example, Netflix has reported that the combined effect of personalization and recommendations saves them more than $1 billion per year. Fewer people cancel their subscriptions, so they make more money from existing customers and don’t have to spend as much on attracting new ones. Another industry leader, Amazon, reports that about 35% of their sales originate from cross-sales, which is a direct result of their recommendation engine. This highlights the significant revenue potential of personalization and recommendation algorithms. 

The value of these deep learning algorithms goes beyond just revenue and cost. They can also profoundly impact the overall value of a company. Take TikTok’s algorithm as an example. It’s estimated to be worth around $60-70 billion, and many analysts believe that without it, the company would be worth significantly less, around $30-40 billion.

There’s also a McKinsey article stating that companies that grow faster drive 40% more of their revenue from personalization than their slower-growing counterparts. In fact, personalization generates roughly $1 trillion in value by top-quartile performance players. 

So, what sets leaders apart from others?

Their ability to customize their offerings and outreach to the right individual, at the right moment, with the right experiences. 

But how do you know what those experiences are? 

To answer this question, we have to look at the expectations of modern shoppers. 

A study conducted by Google and BCG revealed a striking finding: if an algorithm that provided personalized product recommendations were to be eliminated, it would put around 10% of companies’ total revenue at risk. This highlights personalization’s major impact on a retailer’s bottom line. But it’s not just about having personalization—it’s about having the right personalization. 

Shoppers want experiences that are convenient and relevant. They expect to quickly and easily find the products or services they need without digging through numerous options. In fact, a remarkable 80% of global shoppers prioritize speed, citing the importance of getting things done quickly. To meet this demand, retailers must provide frictionless experiences that make engagement and purchasing effortless.

Customers also want to get value from their purchases, with 87% of customers considering it necessary to feel they’ve secured a good deal. This is an excellent opportunity for retailers to stand out. Personalized offers and discounts give customers that sense of value. When recommendations match what they want, shopping becomes easier and more enjoyable, motivating them to buy more.

The most telling fact is that 62% of global shoppers rely on recommendations when deciding which brand or retailer to support. This highlights the importance of offering timely, relevant products or services. It lets shoppers find things they love easily, building a sense of trust and loyalty with their preferred brands. 

The best way to deliver those targeted recommendations and offer content that resonates with shoppers is to use machine learning (ML) and artificial intelligence (AI) technologies.

This white paper examines recommendation engine algorithms, their role in retail, and why they are vital for creating great shopper experiences.

What is a recommendation engine?

A recommendation engine’s primary goal is to promote sales through personalized recommendations. It also influences buying behavior, e.g., prompting shoppers to opt for more profitable items or increasing average order value through cross-sell and upsell opportunities.

At their core, recommendation engines are complex customer engagement software that use machine learning (ML) and artificial intelligence (AI) to track and analyze customer data. This includes browsing history, search queries and the user's past behavior, as well as product information, such as attributes and pricing. Engines then use this data to generate personalized product suggestions and deliver them to users through channels like web storefronts, mobile apps and offline points of sale.

The data is typically aggregated and managed by a Customer Data Platform (CDP)—a centralized hub that enables recommendation engines to train their algorithms and make predictions or suggestions. 

The best recommendation engines integrate with the retailer’s product catalog and CRM and marketing automation platform to combine particular user data with product data and provide accurate suggestions. 

These personalized suggestions or recommendations are delivered within the original retailer’s design, providing a consistent and intuitive shopping experience that mirrors the brand aesthetic.

The diagram above illustrates a VTEX commerce platform solution example, showcasing how user interactions, catalog data and personalization (recommendations system engine) data flows are integrated.

What algorithms are used in recommendation engines?

With over a decade of experience in recommendation engines, Loadstone has become a trusted retailer partner. We focus on increasing customer lifetime value (LTV) and driving long-term growth. The goal is simple—help businesses achieve the best e-commerce performance.

As part of this white paper, we conducted a detailed analysis using a massive dataset that includes trillions of data points from hundreds of retailers. These retailers operate across multiple geographies and verticals, including apparel and accessories, home and garden, electronics, food and beverages, health and beauty, sports and leisure, pets, kids and baby, gifts and more. 

The analysis highlights key performance metrics, including attributed revenue (sales directly linked to the recommendation engine), average order value (AOV) and units items per transaction (UPT). While these metrics provide valuable insights into the effectiveness of recommendation engines, other important e-commerce metrics, such as conversion rates (CR), click-through rates (CTR), customer retention rates, bounce rates, and overall customer lifetime value (LTV), weren’t explicitly examined in this study. A more in-depth analysis of each retailer’s business would be required to assess these metrics accurately.

The findings of this analysis provide a perspective on preferences in retail recommender engine algorithms, summarized in the table below. 

Note. - Attributed Revenue is the revenue generated by shoppers who interact with recommendations under a given attribution policy; Engaged AOV Factor is the difference in average order value for shoppers who interact with recommendations compared to those who don’t, under a given attribution policy, after being influenced by the algorithm; Engaged UPT Factor is the difference in units per transaction for shoppers who interact with recommendations compared to those who don’t, under a given attribution policy, after being influenced by the algorithms.

The most compelling insight from our analysis is that top-performing retailers are using algorithms to outpace average retailers, achieving impressive gains of 30-40% in key e-commerce metrics. These leaders are realizing 35-40% higher attributed revenue, 25-30% higher AOV and 15-20% higher average UPT. They also got an impressive 500% increase in total revenue and a 300% surge in orders where shoppers interact with personalized recommendations. 

Recommender engine algorithms that top-performing retailers use are designed to accomplish specific objectives and can be categorized into four groups:

• Category-based
• Product-based
• Shopper-based
• Search-based

Here’s a deep look at each group’s strengths and capabilities, which will help retailers adjust their approach to meet unique business requirements. 

Category-based algorithms

This group includes algorithms that display products across the entire store or within selected categories. The category-based algorithms increase the discovery of new and popular items, driving moderate attributed revenue and engaging with shoppers with higher AOV. 

Examples of category-based algorithms:

• Popular for category: This algorithm helps shoppers find new and popular items. It leads to a moderate attributed revenue and boasts a high engaged AOV factor. It identifies popular products and displays them, making it easier for shoppers to discover new and trending products without browsing through dozens of products in a category.

• Popular discounted for category: This algorithm increases sales of trending items at a discounted price. It brings moderate attributed revenue and moderate engaged AOV and engaged UPT factors. The Popular discounted for category algorithm connects with shoppers looking for top-selling products on sale. It highlights popular items with active discounts, making deals easy to spot. Shoppers can quickly find great offers and buy confidently, knowing they’re getting popular products at a better price.
• Novelties from categories: This algorithm keeps shoppers engaged with the latest products, has a low attributed revenue and moderate engaged AOV and engaged UPT factors. The Novelties from the categories algorithm attract shoppers with a higher AOV and purchase more items per transaction. This algorithm identifies new and trending products and suggests them prominently, making it easier for shoppers to discover new, exciting items.

Loadstone clients love using our category-based algorithm—Popular products. It creates an e-shelf that shows the most viewed and liked items that align with what a customer might buy next. It does so by looking at their actions and what other people are interested in. For instance, if someone often browses electronics, the algorithm will show them similar popular items, like gadgets that are getting a lot of attention or sales in that category.

Product-based algorithms

This group includes algorithms that showcase products closely related to a specific product.

They engage shoppers interested in purchasing additional items that go well with their initial purchase.

Product-based algorithms include: 

• Accessories: This algorithm has the highest engaged AOV and engaged UPT factor. It suggests relevant accessories that increase the potential for cross-selling. The Accessories algorithm identifies products frequently purchased together, e.g., a phone case for a shopper who just bought a phone. This way, shoppers can find valuable accessories quickly.

• Alternative: This algorithm drives the highest attributed revenue and orders. When a shopper is uncertain about their current selection, the Alternative algorithm—a type of collaborative filtering algorithm—helps them discover products popular among people with similar interests. The algorithm serves as social proof, showing the shopper that others with similar tastes have chosen these products. Suppose someone is looking at a biker jacket but spends a long time on a page or keeps switching between options without purchasing. In that case, the collaborative filtering recommender system might suggest other popular jackets that similar users have purchased. Social validation helps build trust and confidence in the shopper’s purchasing decision, making it more likely that they will convert and discover new products they may not have considered otherwise.
• Related: This algorithm has a high-engaged AOV and engaged UPT factors. It helps shoppers find complementary products and ensures they complete their purchase with items that match their selection. Loadstone Total Look AI Stylist algorithm is a perfect example of a Related algorithm. It’s designed specifically for the fashion industry. If a shopper selects a dress, the algorithm might suggest shoes or a handbag that match the initial product’s color and style and the user's preferences. 

Shopper-based algorithms

As the name suggests, shopper-based algorithms show customers specific products that interest them. They do so based on their web storefront or mobile app actions. 

Examples:

• Personal: This algorithm drives high attributed revenue and boasts high engaged AOV and engaged UPT factors. It analyzes a shopper’s browsing and purchasing history and identifies and suggests products likely to attract them. If a user has shown interest in specific products, the Personal algorithm suggests alternative options, helping to find the best fit and moving closer to making a purchase. If a user has previously made a purchase, the Personal algorithm recommends complementary products, further increasing the potential for upselling and cross-selling.

• Personal for category: This algorithm drives high attributed revenue and has a moderate engaged AOV factor. The personal for category algorithm shows products only from the category that the user is currently in.
• Viewed: This algorithm allows shoppers to quickly revisit and purchase items they’ve already shown interest in. This drives moderate attributed revenue and boasts a moderate engaged AOV factor and engaged UPT factor. The Viewed algorithm connects with shoppers who have a higher AOV and purchase more items per transaction. By displaying products that a shopper has previously viewed, the algorithm makes it easier for them to find and buy products they’re interested in, without distracting them with other products.

Search-based algorithms

These algorithms analyze shopper behavior after search queries to provide relevant recommendations. As such, they increase the effectiveness of the search function, making it easier for shoppers to find appropriate products. For example, if a shopper searches for “running shoes with extra cushioning”, the algorithm uses content-based filtering to suggest running shoes with different cushioning levels.

Search-based algorithms are a perfect fit for behavioral marketing tactics as they analyze a shopper’s search queries, recommend products that match the user’s search and offer alternatives. This way, they help shoppers find the most suitable product, even with non-standard or complex search phrases. A shopper may type “waterproof red leather boots for hiking”, and the search- and content-based recommender system can show a selection of boots that meet those specific requirements, like red color, durability and water resistance.

Personalization strategies for recommender engine algorithms

Now that we’ve covered the key recommendation engine algorithms for great shopping experiences, let’s see how personalization works within and alongside them.

To set the record straight: Best practices in personalization combine multiple algorithms across the shopper buying journey stages to create a customized shopping experience.

Here are the exact stages businesses should consider when implementing recommender engine algorithms:

Everything begins with Awareness, when the shopper is introduced to the brand or product. As the shopper progresses through the journey, they move through Exploration, Consideration, Evaluation and Conversion stages. The final stage, Post-Purchase, involves the shopper engaging with the brand after they buy.

The buying journey for non-vertical specific retailers is illustrated below, including stages, tasks, pages and algorithms.

To get the most out of a recommendation engine, retailers must align it across all buying journey stages and e-commerce touchpoints. This includes smooth integration with the home page, product detail pages, category pages, search results page and shopping cart page, among others.

It’s something we pay close attention to at Loadstone. Our AI Recommender System automatically connects with retailers’ systems and SKU databases to offer relevant and available products in every buying journey stage. 

Below are personalization strategies that enable retailers to outmaneuver the competition with their recommender engines and maintain a best-in-class position in the market.

Combining Alternative, Popular and Search algorithms

Putting Alternative, Popular and Search algorithms to work together lets businesses create a unique shopping experience that resonates with shoppers and fuels UPT and AOV growth:

• Popular algorithm showcases trending and best-selling items on key pages, such as Home, Category and Search results. It creates a sense of urgency and increases the likelihood of conversion.
• Alternative algorithm serves as a discovery catalyst on Product pages, encouraging shoppers to explore more products.
• Search algorithm provides precision-guided recommendations on the Search results page, ensuring that shoppers quickly find relevant products that match their search queries.

Serving high-intent shoppers with relevant product recommendations

This strategy creates a personalized shopping experience that increases revenue. It targets high-intent shoppers with recommendations that fit their preferences, browsing history and purchase behavior.

By combining Accessories, Personal for category and Novelties from categories recommender engine algorithms, businesses can deliver an experience that resonates with shoppers and drives orders and attributed revenue: 

• Novelties from categories algorithm keeps shoppers engaged with the latest products. It makes it easy for them to stay on top of trends and discover fresh favorites on the Home, Category and Subcategory pages.
• Personal for category algorithm delivers personalized product recommendations on Category and Subcategory pages, allowing shoppers to discover new products that align with their interests.
• Accessories algorithm acts as a complementary product suggester on Product and Cart pages. Companies can set it up to propose relevant add-ons to improve the shopping experience and increase AOV.

Synchronising Personal, Viewed and Related algorithms

Personal, Viewed and Related algorithms combined make a true revenue machine:

• Personal algorithm provides shoppers with custom product recommendations on the Home page, Order Tracking and Search results pages, aligning with their unique preferences and interests.
• Viewed algorithm uses customers’ browsing history to suggest products they’ve previously shown interest in on the Product and Product Details pages. This rekindles their engagement and encourages them to complete a purchase. 
• Related algorithm identifies complementary products likely to resonate with the shopper on Product, Product Details and Cart pages, making it easy for them to discover new products and increasing the chances of a sale.

Cost of implementing and operating recommendation systems

Although recommendation engine algorithms offer undeniable advantages, such as increased revenue and AOV, some businesses may hesitate to adopt them. The main concerns include complexity, cost, implementation and ongoing operations. 

But the truth is that with the right e-commerce tools and expert guidance, deploying a recommendation engine is a straightforward process. 

The Loadstone Recommender System is one such tool. 

It includes over 35 machine learning algorithms that assist businesses in delivering personalized recommendations on all online and offline touchpoints—websites, apps and offline stores.

Plus, whether you’re looking to align with new market trends or strengthen the impact of your recommendation engine, you can count on the Loadstone team. Our experts have decades of experience helping retailers every step of the way, from initial deployment to ongoing optimization and improvement.

Let’s look at the objections and the cost of building a recommendation engine in-house versus getting a pre-made solution.

The estimated cost of building and testing an in-house basic recommendation engine is around $400,000 annually, excluding IT infrastructure costs. This includes the cost of recruiting and onboarding a team of developers, data scientists and DevOps engineers, as well as the cost of A/B testing and data analysis. Conducting each A/B test necessitates a substantial allocation of time and resources. It requires at least 100 hours of specialized expertise and costs approximately $5,000 to $10,000 per test. Fine-tuning each algorithm involves a series of tests, adding to the overall cost and complexity.

As a result, the total cost of developing and optimizing a recommendation engine in-house could be substantial, making it a massive undertaking for many businesses. 

Besides financial costs, building an in-house recommendation engine also requires considerable time. More precisely, developing a recommendation engine can take 5-7 years.

A recommendation engine as a ready-made product is a more cost- and time-effective choice. And it comes with even more advantages. When bought as part of a broader martech ecosystem, such as Loadstone, the recommendation engine includes technical support and dedicated customer success management. This means retailers receive expert guidance and proven strategies to get the most out of their solutions.

With the Loadstone composable commerce platform, companies can also expand their tech stack with other Loadstone products like AI Commerce Search and CRM and Communications.

Our cost-benefit analysis reveals that opting for a recommendation engine as a ready-made product can yield a cost savings of $1,672,000 over 5 years for a customer base of 1 million CDP. 

This considerable cost advantage stems from eliminating upfront development costs, reducing personnel expenses and minimizing A/B testing. By selecting this option, retailers can reallocate the savings to other strategic initiatives, getting greater value and return on investment (ROI) for their organizations. This approach ensures retailers can capitalize on the latest advancements in recommendation engines, ML and AI technologies.

Customer privacy 

Another common concern is the potential impact on customer privacy. And it exists rightfully so—it’s vital to ensure customers’ and your business’s data is always safe. 

Fortunately, the best recommendation engines comply with all relevant legislation, ensuring that data is protected and used responsibly. 

At Loadstone, with over a decade of expertise in recommendation engines, we have a deep understanding of the importance of balancing personalization and privacy. Our AI Recommender System provides highly personalized recommendations that improve the shopping experience while respecting shopper boundaries and adhering to the highest data protection standards. The commitment to transparency, security and compliance with regulatory requirements, such as GDPR and US privacy law compliance, ensures that customers can trust the handling of their data. 

How do you measure the recommender engine’s success?

Measuring the impact of a recommendation engine is another potential challenge that may prevent businesses from implementing this technology. Tracking all the effects of the recommendations on sales, customer behavior and overall business performance can feel overwhelming. It can also be challenging to determine the direct impact of a recommendation engine, as there may be other factors at play.

However, just like it’s the case with every other business process, tracking performance is necessary as it lets retailers understand how the engine impacts their results. And there are proven tactics and key metrics that businesses can use to evaluate their algorithms’ effectiveness. 

One of the most insightful metrics for measuring a recommendation engine’s business value is revenue lift, which quantifies an engine’s influence on sales. Here, it’s important to distinguish between incremental sales generated by the recommendation engine and those that would have occurred organically, as shoppers may have purchased recommended products regardless of the recommendation engine’s suggestions. This attribution challenge can obscure the true impact of the recommendation engine. That’s why retailers should use analytics and testing tools to measure algorithms’ effectiveness accurately.

The Loadstone Recommender System makes it simple to track performance. Its built-in analytics tools organize data in a user-item matrix, showing key indicators like impressions, clicks and purchases for each recommended product. This helps businesses see how well their recommendations are working.

The key metrics retailers should track to evaluate the effectiveness of their recommendation engine include:

• Click-Through Rates (CTR) and Conversion Rates (CR): The percentage of users who click on a recommended item, and the percentage of users who complete a purchase after clicking on a recommended item.
• Attributed Revenue: The revenue generated by shoppers who interact with recommendations, under a given attribution policy. 
• Average Order Value (AOV): The average amount a shopper spends in a single transaction.
• Engaged AOV Factor: The difference in average order value for shoppers who interact with recommendations compared to those who don’t, under a given attribution policy, after being influenced by the algorithms.
• Units Per Transaction (UPT): The average number of items a shopper purchases in a single transaction.
• Engaged UPT Factor: The difference in units per transaction for shoppers who interact with recommendations compared to those who don’t, under a given attribution policy, after being influenced by the algorithms.
• Revenue Lift: The increase in revenue generated by the recommendation engine compared to a control group. 

Drive success with the right recommender engine algorithms

Recommendation engines are powerful technologies for businesses looking to drive engagement, conversion rates and sales. By implementing the strategies and algorithm mixes we’ve covered, companies can achieve their goals. Whether increasing orders and attributed revenue or scaling UPT and AOV, recommendation engine algorithms can help retailers succeed.

The Loadstone team has years of firsthand experience with the impact that well-designed and well-implemented recommendation engine algorithms have on a business’s bottom line. Our clients like Koröshi and Garmen have seen a 110% increase in average order value and a 378% jump in revenue with the Loadstone Recommender System. These results aren’t luck—they’re proof of what the right technology can do. We’re here to make sure retailers get the most out of it.

Schedule a call with our team today to see what Loadstone can do for you.