AI Product Recommendations Workflows: Step-by-Step Processes

Recommendation engines predict which products a shopper is likely to want next. They learn from browsing, add-to-cart and purchase events, plus product attributes such as category, price and colour. Common models include collaborative filtering (people who bought this also bought that), content-based matching and, increasingly, models that combine both with session context.

Stores rarely build these from scratch. Shopify offers native related-product logic and apps such as Rebuy or LimeSpot; Salesforce Commerce Cloud includes Einstein recommendations; Klaviyo, Dynamic Yield, Nosto and Google Cloud’s Vertex AI Search for commerce are other common options. The workflow is similar across them.

Recommendations are only as good as the catalogue and event data behind them. Most problems that look like a bad algorithm are really missing attributes or broken tracking.

• Check that product view, add-to-cart and purchase events fire with the correct product ID on every page.
• Fill core attributes: category, brand, price, gender or age group where relevant, and stock status.
• Remove or flag discontinued products so they are not recommended.
• Make sure variants (size, colour) roll up to a parent product, or the engine will recommend the same shirt in five sizes.
• Confirm consent settings: personalised recommendations may require consent under GDPR or KVKK in Turkey.

Each page has a different job, so each needs a different recommendation type. Starting with three placements is usually enough to learn from.

• Product page: similar items for shoppers still deciding, and frequently bought together for those ready to buy.
• Cart page: low-priced complements, such as cases for phones or socks for shoes.
• Home page: recently viewed and personalised picks for returning visitors; best-sellers for new ones.
• Empty search results: popular items in the closest category instead of a dead end.
• Order confirmation: replenishment or accessories for the next visit.

Pure algorithms optimise for clicks and can push the wrong things: cheap items, out-of-stock products or brands with poor margins. Rules let merchandisers guide the model without replacing it. Start with a short list and add rules only when you see a specific problem.

Typical rules include excluding items under a margin threshold, boosting new collections for a set period, never recommending a cheaper alternative on a high-ticket product page, and filtering by the same gender or size range as the item being viewed. Write each rule down with its reason, so the next person understands why a product never appears.

Test one placement at a time against a control. The control can be no widget, a simple best-seller list or the current logic. Most recommendation platforms include A/B testing; otherwise use a testing tool and split by visitor.

Measure revenue per visitor and conversion rate for the whole session, not just clicks on the widget. A widget can collect many clicks while pulling attention away from the main product. Run tests for at least two full weeks to include weekday and weekend patterns.

The same engine that powers the site can feed browse abandonment and post-purchase emails in Klaviyo or a similar tool. Product feed segments built from recommendation logic can also shape Google Shopping and Meta catalogue campaigns, for example by grouping frequently bought together items into a bundle feed label.

Keep timing in mind. A shopper who bought a sofa does not want three more sofas by email the next day. Post-purchase flows should switch to complements and care products.

Recommendations drift without an owner. New collections arrive, suppliers change, a sale ends and a rule written for it stays active. Assign one person, usually in e-commerce or merchandising, to review placements every two weeks.

• Spot-check five product pages in different categories on mobile and desktop.
• Review the top 20 recommended products and confirm stock and margin.
• Remove expired campaign rules and boosts.
• Read the test results dashboard and note any placement that lost revenue per visitor.
• Log every change with a date and a reason.

This takes about an hour. It prevents the slow decline that makes teams think the engine stopped working, when in fact the store changed around it. It also gives the merchandiser a regular view of what shoppers actually pair together, which often informs bundles and promotions.