How AI agents actually buy things
The mechanics of agent-driven purchasing: how agents find products, evaluate price, and what makes a catalogue legible to them.
An agent purchase has four steps: understand the intent, find candidate products, evaluate them, and complete the transaction. Only the last has a protocol; the first three are where most agents are weakest.
Finding is a search problem across sellers, not within one. Agents compare exhaustively rather than sampling, which means a product invisible to structured retrieval is invisible full stop, regardless of how well it ranks for humans.
Evaluating is where agents fail most often. Presented with a price, an agent has no basis to judge it — no history, no market context, no memory of what this shopper considers expensive. It will either accept the first plausible option or spend enormous context trying to reconstruct context it cannot reach.
The practical implication for merchants: publish identity and a structured catalogue. The practical implication for anyone building agents: give the agent a verdict, not a number, and give it somewhere durable to remember what it learned.
Keep reading
What is commerce intelligence?
Commerce intelligence is the layer above price monitoring: an index of what exists, who sells it, at what, and what that means.
Why percentile beats percent-off
A percentile against a product's own history is a defensible measure of discount depth; percent-off a reference price is not.
Product matching and why it fails quietly
How cross-seller product matching works, the three ways it fails, and why published match confidence matters.
GTIN, UPC, EAN and ASIN
How product identifiers relate, why check digits matter, and what to do when there is no identifier at all.
Data provenance in commerce data
Why every price should carry its origin and a confidence weight, and what goes wrong when it does not.