Recommendation quality depends on context. A product detail page, cart, checkout, home page, search results page, and post-purchase email each represent a different customer intent. Reusing one recommendation algorithm across all of them usually produces predictable but weak results: the same products appear everywhere, high-margin items are over-promoted, and recommendations ignore what the shopper is trying to do at that moment.
For WooCommerce agencies, recommendation surfaces should be treated as separate optimization problems. Each surface needs its own candidate products, ranking signals, business rules, and success metric.
Product detail pages: optimize for discovery and comparison
A product detail page usually answers one of two questions: “What else might I need?” or “What are comparable alternatives?” Those are different recommendation goals and should not always be combined in one carousel.
Useful product detail page strategies
- Frequently bought together: Recommend complementary products that are commonly purchased with the viewed item.
- Similar products: Recommend alternatives using category, attributes, price range, brand, and product type.
- Recently viewed alternatives: Help returning shoppers continue product comparison.
- Accessory recommendations: Show compatible or required items, such as filters for a coffee machine or a case for a camera.
For example, a shopper viewing a DSLR camera may benefit from lenses, batteries, and memory cards in an accessory module. A “similar cameras” module should instead rank products with comparable specifications and price points. Combining both intents into a single list makes the ranking difficult to interpret and harder to test.
Useful signals include product co-purchases, co-views, taxonomy relationships, attributes, price distance, inventory status, and historical conversion. The primary metrics are often add-to-cart rate, attach rate, and revenue per product detail page view rather than raw click-through rate.
Cart pages: optimize for completion and basket expansion
A cart recommendation is made after the shopper has demonstrated stronger intent. The algorithm should focus on products that complement the current basket without distracting the customer from checkout.
Cart recommendations can use:
- Products frequently purchased with the items already in the cart.
- Missing accessories or consumables required for the selected products.
- Threshold products that help the customer qualify for free shipping or a promotion.
- Low-friction add-ons with suitable price and fulfillment characteristics.
Cart logic should be basket-aware. If the cart contains a printer, recommending another printer is usually less useful than recommending ink or paper. If the cart already contains a product, its duplicates should normally be excluded unless the store supports quantity-based purchasing.
Agencies should also apply operational filters before ranking: exclude out-of-stock products, products unavailable in the customer’s shipping region, and products that conflict with cart-level rules. Measure incremental items per order, average order value, attach rate, and checkout completion rate. A recommendation that increases average order value but reduces completed orders may be harmful.
Checkout: use restraint and protect conversion
Checkout is a high-intent surface with limited tolerance for distraction. In many stores, recommendations should be restricted to essential add-ons, digital products, warranties, or services that are directly relevant to the order.
At this stage, a simple rules-based algorithm can outperform a complex model. For example, an extended warranty can be shown only when the cart contains an eligible product. A setup service can be offered only for products that require installation. The eligibility rule is more important than broad behavioral personalization.
Track checkout completion, offer acceptance, refunds, and customer support contacts. Avoid placing recommendation components where they interfere with payment fields, express checkout buttons, or required legal disclosures.
Home pages: optimize for broad discovery and returning intent
The home page has less explicit intent than a product or cart page. It should combine merchandising priorities with behavioral personalization rather than relying only on immediate-session events.
Suitable modules include:
- Recently viewed products for returning visitors.
- Popular products within the shopper’s relevant category or region.
- New arrivals for customers who frequently browse new inventory.
- Personalized category shortcuts based on prior browsing and purchases.
- Seasonal or campaign collections controlled by the merchandising team.
A home page algorithm should maintain diversity. Showing six nearly identical products may maximize predicted clicks for one category while reducing the chance that the shopper discovers a different product family. Use category, brand, price-band, and attribute diversification constraints after the initial ranking.
For anonymous visitors, session behavior and contextual signals such as device, geography, referral campaign, and current category interest may be useful. For authenticated customers, purchase history and long-term preferences can be added with appropriate privacy controls.
Search results: prioritize relevance before personalization
Search is an explicit request. If a shopper searches for “black waterproof hiking boots,” lexical and semantic relevance should dominate the ranking. Personalization can refine results, but it should not replace query intent.
A practical search ranking sequence is:
- Match the query against product names, SKUs, attributes, categories, and searchable descriptions.
- Apply hard constraints such as stock status, catalog visibility, region, and customer eligibility.
- Rank by textual relevance, product quality, and commercial signals such as conversion rate.
- Apply limited personalization using brand preferences, price sensitivity, or previous category engagement.
Recommendation modules can still appear alongside search results, but they should be labeled and separated from the requested results. For example, a “You may also like” module can use collaborative filtering, while the main result list remains query-driven.
Post-purchase and email: optimize for the next useful action
After purchase, the best recommendation may depend on product lifecycle rather than immediate co-purchase behavior. Consumables can be recommended near the expected replenishment date. Accessories can be recommended after delivery. Educational content or setup services may be more appropriate immediately after purchase.
For example, a subscription store might estimate the next reorder window from order intervals. A home improvement retailer might recommend compatible maintenance products several weeks after delivery. These algorithms require order history, product compatibility, delivery status, and timing signals.
Measure repeat purchase rate, replenishment conversion, subscription enrollment, and unsubscribe rate. A high click rate is not sufficient if messages arrive too early or promote products the customer already owns.
Use a shared platform with surface-specific policies
Different algorithms do not require completely separate infrastructure. A recommendation service can share event collection, product metadata, feature computation, candidate generation, experimentation, and reporting while allowing each surface to define its own policy.
A useful recommendation request might include the following context:
{
"surface": "cart",
"customer_id": "customer-123",
"product_id": "product-456",
"cart_product_ids": ["product-456", "product-789"],
"currency": "USD",
"shipping_country": "US",
"limit": 4
}
The response layer can then apply surface-specific rules. A product page may request complementary and similar candidates. A cart may request compatible add-ons and threshold products. Checkout may request only eligible services. The same product catalog and event pipeline support all three, but the candidate sets and ranking policies remain distinct.
Define business rules before model tuning
Machine learning cannot correct incomplete catalog relationships or invalid commercial rules. Before tuning a ranking model, document exclusions and constraints for each surface.
- Inventory and backorder policy.
- Product visibility and customer-group restrictions.
- Compatibility and variation rules.
- Brand, category, and price diversity requirements.
- Promotion, margin, and supplier constraints.
- Recently purchased or recently viewed exclusions.
- Regional shipping and tax availability.
In WooCommerce, these rules should be evaluated against the same product and variation data used by the storefront. Agencies should account for parent products, individual variations, product visibility settings, stock management, and custom taxonomies. A recommendation that points to a visible parent product but cannot select an available variation creates unnecessary friction.
Measure each surface against its real objective
Use surface-level experiments instead of one global recommendation metric. Product detail pages may optimize add-to-cart rate or attach rate. Cart modules may optimize incremental items per order while monitoring checkout completion. Search personalization may optimize query conversion and zero-result reduction. Post-purchase messaging may optimize repeat purchase within a defined time window.
Where possible, use holdout groups to measure incremental revenue rather than attributing every order containing a clicked recommendation to the algorithm. Also report coverage, recommendation freshness, out-of-stock exposure, and the percentage of sessions receiving a valid recommendation. These operational measures reveal when a model appears accurate in aggregate but performs poorly for important customer segments.