Why a Recommendation Engine for E-Commerce Boosts Sales and Where Business Rules Fit In

Written by

Łukasz Niedośpiał

Published on

23 Aug 2023

Last update

28 Sep 2026

Key takeaways

A recommendation engine suggests products to each shopper based on their behavior and the behavior of similar shoppers.
Personalization most often drives a 10–15% revenue lift, according to McKinsey's cross-industry research.
Algorithms pick candidates. Business rules decide what can actually be shown: in stock, sellable in the shopper's region, within margin and promotion policy, and appropriate for that shopper.
Recommendations need ongoing testing and rule reviews. They are not a set-and-forget feature.

What is a recommendation engine for e-commerce?

A recommendation engine is software that suggests products to each shopper based on data: what they browse and buy, what similar shoppers bought, and what is popular right now. It powers widgets such as "Recommended for you," "Frequently bought together" and "Customers also viewed," as well as product suggestions in email and in apps.

A production setup has two parts. An algorithm scores which products a shopper is likely to want. Business rules filter and rank those candidates against the retailer's policies before anything reaches the page. Both matter for the result.

Why do recommendation engines increase sales?

Recommendations help shoppers find relevant products faster, which lifts conversion and order value. McKinsey found that personalization most often drives a 10–15% revenue lift, with company-specific results ranging from 5% to 25%; 71% of consumers expect personalized interactions and 76% are frustrated when they do not get them (McKinsey, November 2021).

AI-driven recommendations now shape a large share of online spending. Salesforce reported record online holiday sales of $1.29 trillion globally and $294 billion in the US for the 2025 season, and estimated that AI and agents drove 20% of all retail sales, or $262 billion, through personalized recommendations and engagement (Salesforce, January 2026). Salesforce does not publish a detailed methodology for "AI-influenced" sales, so treat the figure as a directional signal.

The metrics to watch are specific:

Click-through rate on recommendation widgets.
Conversion rate of sessions that engage with recommendations.
Average order value and items per order.
Revenue per visitor, compared with a control group.
Return rate of recommended products - a recommendation that gets returned is not a win.

How do recommendation algorithms work?

Three families of methods cover most e-commerce use cases:

Method	How it works	Good for	Limit
Collaborative filtering	Recommends what similar shoppers bought or viewed	"Customers also bought," large catalogs with lots of traffic	Cold start: weak for new products and new shoppers
Content-based filtering	Recommends products with similar attributes to what the shopper liked	Similar products, new items with good product data	Tends to recommend "more of the same"
Hybrid models	Combine both, often with machine learning ranking models	Home page and personalized feeds	More data and tuning required

Most retailers today use a hybrid approach, often supplied by their commerce platform or a specialized personalization vendor.

Why do recommendations need business rules around the algorithm?

Because the product a model scores highest is not always the one a retailer should show. Business rules handle the decisions a model does not know about:

Availability. Do not recommend out-of-stock items, discontinued products or items that cannot ship to the shopper's region.
Margin and pricing policy. Avoid pushing items sold below margin, or respect minimum advertised price agreements.
Promotions and contracts. Give agreed placement to a brand campaign, or exclude competing brands on a partner's landing page.
Appropriateness and compliance. Exclude age-restricted products for unverified shoppers, or products that cannot be sold in certain states.
Journey context. On the cart and checkout pages, show only low-price accessories that do not pull the shopper away from completing the order.
Diversity. Limit how many items from one brand or category appear in a single widget.

These rules change often, and they are usually owned by merchandising, legal or category managers, not by the data science team. Keeping them in a separate rules layer lets those teams change them without retraining a model. The same pattern is common in regulated industries; for how insurers apply it, see Product Recommendation Engines in Insurance.

Which recommendation tactics work best, and what rules sit behind them?

Tactic	Where it appears	Typical business rule behind it
Recommended for you	Home page, account page, email	Exclude items already purchased; respect category preferences
Frequently bought together	Product page, cart	Only in-stock items; bundle price must stay above margin floor
Similar products	Product page	Same category and price band; exclude out-of-stock
Complementary accessories	Product page, cart	Compatible with the main product; low price on cart page
Newer version available	Product page, post-purchase email	Only when the successor is in stock and in the same price range
Best sellers and top rated	Category pages, empty search results	Minimum number of reviews; regional availability
Bundles with a discount	Product page, checkout	Discount within promotion budget and dates
Post-purchase suggestions	Order confirmation, email	Wait period by category; exclude consumables bought recently
Event and seasonal picks	Home page, email	Active campaign dates; regional relevance

How do you keep recommendations working after launch?

Recommendation engines are not a set-and-forget feature. Shopper behavior, catalog and promotions change constantly. Three habits keep results up:

Test against a control group. Use your storefront's experimentation tools to compare recommendation strategies, and judge them on revenue per visitor and returns, not only on clicks.
Review rules on a schedule. Promotions end, contracts change and new product lines arrive. Stale rules are a common reason recommendations underperform.
Watch the edges. Check what new shoppers, returning shoppers and mobile users actually see. Cold-start and small-screen experiences are where most weak recommendations show up.

How does Higson fit into a recommendation setup?

Higson is a business rules engine built by Decerto, used mainly by insurers and financial services companies. It is not an e-commerce personalization platform. In a recommendation setup, it can serve as the rules layer around your algorithm:

Decision tables for merchandising rules. Availability, margin, promotion and exclusion rules live in tables that business users edit and test, instead of in application code.
Your models inside the rules. Higson's ONNX runtime runs machine learning models inside rule execution, so a model score and the business rules around it are evaluated together.
Real-time performance. Rule execution takes 0.23 ms per decision, with throughput of 9,000 requests/second.
Versioning and audit trail. Every rule change is versioned with rollback, so a team can see what changed when results moved.

If you work in insurance and want to see how carriers apply the same pattern to bundling, coverage and renewal offers, read Product Recommendation Engines in Insurance. For rules engine fundamentals, see What Is a Rules Engine?.

‍

Sources

McKinsey, „The value of getting personalization right - wrong - is multiplying" - https://www.mckinsey.com/capabilities/growth-marketing-and-sales/our-insights/the-value-of-getting-personalization-right-or-wrong-is-multiplying
Salesforce, „Salesforce Reveals 2025 Holiday Shopping Data" - https://www.salesforce.com/news/stories/2025-holiday-shopping-data/