AI in Practice
AI recommendations that actually move revenue
Neha5 min readMost ‘AI-powered’ recommendation engines are rules-based with a label. Here’s what real collaborative filtering does to AOV on a commerce site.
The Label Versus the Reality
Every recommendation app in the Shopify ecosystem claims to be AI-powered. Most of them run static rules. Bestsellers get shown to everyone. Products frequently bought together appear in a carousel. Items from the same category fill a “You might also like” section. This is a database query with a marketing label.
Recommendation systems can combine learned product relationships, catalog attributes, merchandising rules and customer signals. Individualized output is one design choice, not the test of whether a system uses machine learning. Compare the inputs and ranking behavior available in the configured implementation, then evaluate the result for the customer task.
We built a recommendation system for a DTC supplement brand that replaced their rules-based app. Average order value increased by 23% in the first 90 days. The difference came from one thing: recommending products the customer hadn’t already considered but was statistically likely to buy.
How Rules-Based Recommendations Work
A rules-based system follows predetermined logic. If a customer views Product A, show Products B, C, and D. If a customer adds Product E to their cart, suggest Product F. The rules are set by a merchandiser or auto-generated from purchase correlation data.
This works when the recommendation is obvious. Someone buying dog food probably needs treats. Someone buying a moisturizer might want the matching cleanser. For those scenarios, simple rules perform fine.
Rules fail when the connection between products isn’t obvious. A customer who buys a glucosamine supplement for their dog might benefit from a probiotic, but a rules engine won’t surface that recommendation unless enough previous customers bought both products in the same order. If the probiotic is new, or if most customers buy them months apart, the rule never triggers.
Rules can use customer context when the implementation supplies it. A product-only recommendation request and a system using an individual customer’s history have different inputs; the label alone does not tell you which is in use. Inspect the configured inputs and behavior before deciding that a separate recommendation service is needed.
Recommendation Approaches
Collaborative filtering identifies patterns across your entire customer base. It finds groups of customers with similar purchase and browsing histories and recommends products that others in the group bought but this customer hasn’t. If customers who buy glucosamine also tend to buy fish oil supplements eight weeks later, the system learns that pattern and surfaces fish oil to glucosamine buyers around that timeline.
Content-based filtering looks at the attributes of products a customer has engaged with and finds other products with similar attributes. If a customer consistently browses grain-free, chicken-flavor products for senior dogs, the system identifies that preference profile and surfaces matching products across categories.
Contextual signals add a time dimension. Recommendation quality changes based on day of week, season, device type, referral source, and what the customer did in the previous session. A customer returning to the site 20 days after purchasing a 30-day supply is likely ready to reorder. A customer browsing at 11 PM on a phone has different intent than the same customer on a laptop at 2 PM.
These three approaches work together. Collaborative filtering provides the candidate set. Content-based filtering refines it for the individual. Contextual signals determine timing and ranking.
The Shopify-Specific Challenge
Shopify’s related-product documentation describes recommendations based on purchase history and product descriptions, with related collections as a fallback. Search & Discovery customizations can affect the results. Shopify also documents complementary-product recommendations that can be displayed in a theme.
The documented related-product flow starts from the product a visitor is viewing. That is different from a promise of real-time individual browsing-history personalization. Check the endpoint, available inputs, configured merchandising and returned results against your requirements; do not infer unsupported limitations or revenue impact from the word native.
Compare native recommendations, an installed app and a separate service before choosing an implementation. A custom service needs an authorized data source, a defined ranking method and a storefront integration; owning a model is not a prerequisite for using learned recommendations. Theme sections can fetch and display recommendations with JavaScript, while a headless frontend implements its own presentation.
Placement and presentation experiments can be implemented in a theme or a separate frontend. Choose based on the changes required, the available integration points and the maintenance burden. Measure any commercial effect rather than assuming the architecture produces it.
The Revenue Impact
The supplement brand we built this for was running a popular Shopify recommendation app that showed “frequently bought together” and “customers also viewed” carousels. Those recommendations generated about $4,200 in attributed monthly revenue.
After switching to the custom recommendation system, attributed revenue from recommendations reached $11,400 per month within 90 days. The increase came from two sources: higher conversion on recommendations because they were more relevant, and higher AOV because the system surfaced products customers hadn’t considered rather than products they’d already seen.
The most effective placement was the post-purchase page. After checkout, the system recommended products based on what the customer just bought, their full purchase history, and what similar customers bought next. Post-purchase recommendation conversion was 4x higher than product page recommendations, because the customer had already committed to buying and the payment information was already on file.
The system cost more to build than a Shopify app subscription. On a catalog large enough that the right next product is not obvious, the revenue difference covered the build inside the first quarter.
What would this look like for your brand?
Talk directly with a founder about one customer journey, the systems behind it and a focused prototype using your approved products and content.
