The Product Is Not Always the Unit of Demand

Retail systems are built around the product because products are what merchants stock, price and fulfil. Customers are less consistent. Sometimes they want a product; often they want a result for which several products happen to be inputs.

The distinction is economically important because it changes the unit that a shopping system should optimize.

A customer replacing a known coffee filter is making a product-level decision. A customer hosting dinner for twelve is managing a bundle of complements. A person buying a drill may want a drill; a person putting shelves on a wall wants a completed project. The correct purchase can depend on whether the customer already owns bits, screws, anchors or a level.

In these cases, the value of the retailer is not only the quality of each recommendation. It is the completeness and coherence of the basket.

Joint demand is normal, but ecommerce treats it as exceptional

Economists have long distinguished complements from substitutes. Ecommerce recommendation systems do too, usually through “frequently bought together,” bundles and cross-sell modules. Yet the conventional interface still requires the shopper to construct the bundle incrementally.

AI changes the feasibility of doing this in reverse. A user can describe an outcome and let software derive the component purchases. Google’s 2026 updates to the Universal Commerce Protocol are telling: agents can retrieve live price and inventory and add several products to a cart together. Google’s Universal Cart similarly treats the basket as a coordinated object rather than a sequence of isolated clicks.

Alibaba is pursuing the same shift at much larger catalogue scale. Reuters reported in May that Qwen will gain access to more than four billion Taobao and Tmall products, allowing users to browse, compare and purchase conversationally while the system also handles logistics and after-sales functions. The important feature is not chat. It is access to enough commercial context to turn an objective into a feasible set of products.

Why grocery and DIY are revealing test cases

Grocery and home improvement expose weaknesses that fashion demos can hide.

A grocery assistant must reason about quantities, dietary constraints, pack sizes, budget and what the household already has. The result can contain twenty products and still be wrong because one essential ingredient is missing. In DIY, compatibility is even harsher: an individually sensible product may make the whole project infeasible if it does not fit the other components.

These are joint-decision problems. They are also where AI can save meaningful time because the customer’s intent is naturally expressed at a higher level than the catalogue.

The same logic applies to skincare routines, travel kits, home furnishing and parts of B2B procurement. The common characteristic is not the vertical. It is that demand is compositional.

The basket becomes the measure

This has implications for recommendation metrics. Product-level click-through is a weak measure of whether an assistant solved the task. The system should be judged on basket completeness, constraint satisfaction, replacement rates, budget adherence and the amount of editing required before checkout.

That is a more demanding standard than “the recommended item was relevant.” It also aligns the system with the consumer’s actual objective.

There is no reason every purchase should become conversational. Known-item retrieval is already highly efficient. The opportunity begins where the catalogue forces customers to decompose an outcome into a collection of items and reconstruct the logic themselves.

In those journeys, the product is not always the natural unit of demand. Sometimes the unit is dinner.

The product is the merchant’s unit of inventory. It is not always the customer’s unit of demand.

“Agents can now save multiple items to a cart at once.” — Google, on UCP’s 2026 multi-item capability