The question that decides whether it makes sense
Recommendation systems are one of the AI use cases with the best-documented ROI in e-commerce and retail. They’re also one of the hardest to evaluate properly, because the correlation between “we showed a recommendation” and “the customer bought it” doesn’t prove causation — the customer might have bought the product anyway.
The starting question has two parts: do you have a large enough catalog and enough active customers for personalization to make a real difference? And can you measure the incremental impact of recommendations, not just the click-through rate?
A ten-product catalog with a hundred customers doesn’t need an AI recommendation system. A catalog of thousands of items with varied buying patterns across thousands of customers is where the ROI is clearest.
Assessing the opportunity
Recommendation systems create value in four scenarios: product discovery (the customer doesn’t know something they’d like exists), cross-sell (adding related products to a purchase), upsell (offering the higher-value version when the customer is inclined to buy it), and retention (reminding customers of products they had a good experience with).
To size the opportunity, start with three questions: what share of purchases are first-time purchases in that category vs. repeat purchases? How many categories does an average customer browse per session? What’s the current conversion rate on product pages? These numbers show where the biggest lever is.
Data requirements
Transaction history per customer. Collaborative filtering (recommending based on what similar customers bought) needs enough interaction density: customers with few purchases have too thin a preference profile for personalized recommendations. New customers create the classic “cold start problem.”
Behavioral data (not just purchases). Product page views, items added to the cart and abandoned, searches, time on page — all these signals enrich the customer profile beyond what they end up buying.
Product attributes. For content-based filtering (recommending products similar to those the customer has viewed or bought), product attributes (category, price, features, brand) make recommendations possible even for new products with no purchase history.
Real-time inventory data. Recommending an out-of-stock product is a bad customer experience and a lost sale. You need real-time inventory integration to filter out unavailable items.
Cost structure and complexity
Well-established recommendation libraries and platforms keep model development costs down. The most variable part is integration: how recommendations appear in the online store, in email, in the mobile app, and in customer service communications.
Real-time personalization (recommendations update based on the current session) requires different computing infrastructure from batch personalization (recommendations are calculated once a day and served statically).
Risk assessment
Filter bubble risk. A system that only recommends what the customer already knows and buys limits product discovery and narrows the basket. Well-designed systems balance exploration (recommending something new but potentially interesting) with exploitation (recommending what the customer is most likely to like).
Cannibalization risk. If the customer would have bought product B anyway, showing B from product A’s page adds no revenue — it just changes the navigation path. Measuring real incremental impact requires A/B tests with control groups that don’t see recommendations.
Privacy risk. Recommendation systems built on intensive behavioral tracking have privacy implications you need to assess under the GDPR, especially with cross-session or cross-device tracking.
Success metrics
- CTR (click-through rate) on recommendations
- Conversion rate from recommendations vs. standard browsing
- Revenue per session with and without active recommendations
- Average order value with and without active cross-sell
- Catalog coverage: % of items that appear in recommendations (a large catalog whose recommendations always show the same 50 products has a diversity problem)
- Incremental revenue measured through A/B testing (the only metric that measures causation, not correlation)
How it delivers ROI
The main levers are higher average order value (more items per purchase through cross-sell) and higher purchase frequency (customers come back sooner because the system reminds them of products they care about). The secondary lever is lower acquisition cost: a customer who finds more value in your store has a higher CLV and is less likely to leave and have to be won back.
In B2B, a recommendation system can also cut order management costs: if the system suggests the products a customer usually buys as part of the ordering process, the customer spends less time building the order by hand.
Frequently asked questions
Are email recommendations as effective as on-site ones? They work differently. On-site recommendations act on active purchase intent (the customer is browsing). Email recommendations act on latent demand (they create the intent). Email has lower absolute conversion rates but can drive visits that wouldn’t otherwise happen. The two channels complement each other.
Can a small catalog benefit from a recommendation system? It depends on how complex customer preferences are. With a small catalog but a lot of variation in individual preferences (food tastes or reading habits, for example), recommendation systems are still useful. With a small catalog and similar preferences across customers, personalization adds little, and a curated editorial selection can work just as well.