Most recommendation engines retrieve. They do not recommend.

Covington, Adams and Sargin at Google published "Deep Neural Networks for YouTube Recommendations" at ACM RecSys 2016. The system has two stages: candidate generation, which narrows billions of videos to hundreds by similarity, and ranking, which orders those hundreds with richer features. The paper is explicit about the split.

The method, step by step: embed products by what appears together in sessions, carts and orders; embed each shopper by their history; measure the distance between shopper and product; return the closest items. It learns that people who did X tend to do Y. It never learns that this shopper should see Z because Z leads to the next order.

Most ecommerce vendors implement the first stage and call it recommendation. The ranking stage, where a value model could live, is missing or optimizes for engagement. Clerk.io's own figures show how much is at stake: recommendations draw 7% of ecommerce traffic but produce 24% of orders and 26% of revenue.

What is similar?

Retrieval. Method: similarity in an embedding space. Optimizes the probability of a click. Result: more of the same. Discount buyers get more discounts.

Which order?

Ranking. Method: learning to rank with more features. Optimizes a mix of clicks, time on page and add-to-cart. Result: a better order of the same similar items.

Which action?

The agent. Method: predict each shopper's next basket and value, then pick the product, message, channel and moment that move them toward the goal. Result: the most valuable next step, which may be a product outside the shopper's pattern.

The research community has measured the gap. At IJCAI 2015, Theocharous, Thomas and Ghavamzadeh showed that of two ad recommendation policies, the one with the lower click rate produced more revenue and a higher lifetime value. Ferraro and colleagues' 2023 systematic review in Expert Systems with Applications found that most recommender research still optimizes engagement rather than economic value. The ACM RecSys 2023 tutorial on customer lifetime value calls the shift from clicks to value a change of paradigm.

A shopper buys a discounted item. The engine places them next to other discount buyers. It shows them more discounts, because those have the highest similarity. They click, which confirms the prediction, so the engine moves them deeper into the discount cluster. The mid-range product that would have opened a new category is never shown.

The engine's accuracy goes up. The customer's value goes down. That is the failure mode of optimizing the click: the system gets better at predicting what the shopper will click and worse for the store.

The opposite is visible in a store's own orders. In one online grocery, 30% of orders include a recommended item and 68% of recommendation revenue comes from a single cart-page block that reminds each shopper of their usual items (case study). In a DIY store, orders placed after a recommendation were nearly 4 times the store's average (case study).

The most valuable question in ecommerce is not "what will this shopper click next?" It is "what will their next basket contain, and when will they buy it?" Next-basket prediction models the purchase sequence over time: buying cycles, category expansion, seasonality, replenishment.

That one prediction changes everything downstream. The browse-abandonment email goes out three days before the predicted purchase window, not 24 hours after the visit. The products in it fit the next basket, not the last click. The channel is the one this shopper answers at this point in their cycle. A shopper with a large predicted basket sees a different offer from one with a small one.

Retrieval has no concept of time, no concept of value and no concept of what comes next. It has a similarity score and a catalog. The agent has one memory of each shopper and a goal you set; the how it decides page shows what it does with them.

Does it know what a customer is worth? Can it recommend something the shopper has never browsed? Do the email picks differ from the on-site picks? Does a one-time buyer get the same logic as your best customer? Can you measure whether a recommendation changed what a customer bought later, not just whether they clicked?

Three or more answers of no means the engine retrieves. In a 50/50 test at a kitchen appliance store, the agent produced 2.3 times as many add-to-carts per visitor, significant at 95%. In a motorbike marketplace, viewers who saw a recommendation booked test rides at 6.5 times the rate (case study).