Before Clarence Saunders opened Piggly Wiggly in Memphis in 1916, shoppers usually handed a list across a counter and a clerk fetched the goods. Self-service removed the clerk, and the clerk's persuading had to happen somewhere else. As Hagley Museum's packaging history puts it, "Without clerks, stores relied on package visibility to sell their products." Boxes were designed to win on crowded shelves. Selling moved onto whatever surface the person doing the choosing could see.
In 1974 the package gained a machine reader, when a cashier in Troy, Ohio, scanned a pack of Juicy Fruit. The scanner needed to know only what the item was. Everything else printed on the pack was still addressed to a human.
Shopping agents put a reader between the shelf and the buyer, and the reader makes the choice. Sellers are already redesigning the box. Google's Merchant Center invites retailers to submit conversational attributes, such as product Q&As, related items and popularity rankings, to help AI agents understand products. Shopify lets merchants feed their policies, FAQs and brand voice to AI shopping tools. The new package is text written for a reader that reasons.
History might seem reassuring here. Shoppers learned to discount the front of the box, and that skepticism became an ordinary skill. One might expect smarter agents to learn it faster.
DECEPTICON, a peer-reviewed study presented this year, tested agents against dark patterns: website tricks such as false urgency, invented popularity and deliberately confusing steps. Across sizes of one model family, the tricks worked 43.8 percent of the time on the smallest model and 73.7 percent on the largest. In a separate test on a different model, letting it think longer before acting raised its rate from 38.5 to 71.2 percent. Experienced human users given the same tasks were fooled about a third of the time.
In one case from the study, agents thinking briefly and agents thinking at length both closed a popup pushing a limited-time air fryer deal. The longer-thinking agent then treated the popup's sales copy as evidence of quality and chose the advertised fryer. The authors call this overthinking. They did not establish its cause, and their pages were built for the test rather than taken from live stores, but the pattern is worth taking seriously.
Packaging worked on impulse, and people can learn to distrust their impulses. A reasoning agent is built to take what it reads seriously, which means more reasoning can draw more of the seller's material into the decision.
Sellers are rewarded for writing text that agents find persuasive, and little in that market rewards restraint. The historian Ai Hisano observed that cellophane let shoppers see food while blocking touch and smell, so the seller controlled what looking could reveal. Conversational attributes extend that control to the agent. The seller decides much of what the agent can know, and how it is phrased.
In an earlier column, I asked how a user could tell whether an agent had secretly sided with someone else. This case is harder: an agent entirely loyal to you, talked into a choice by material it mistook for evidence.
A more plausible defense than smarter reasoning may be labeling what agents read, so that "the seller says" carries less weight than "an independent source shows." Shoppers did not get ingredient lists by becoming savvier. Rules required them, as rules in many places later required unit prices. Agents may need their own version, and someone will have to require that as well.