A simulated florist shows how retail AI can escape a no-advertising trap

A retailer’s decisions determine not just what it earns, but what it can learn. Stop bidding for sponsored-search traffic, and there may be no new evidence about whether paid advertising would attract customers. Change discounts, and shoppers who postpone buying can look like shoppers who never wanted the product.

In a preprint, Hadi Mohammadi of Utrecht University develops an agent for these intertwined problems. Tested in a simulated online florist, the method repeatedly samples a plausible market and follows the selling decisions that would be optimal if that market were real. The results favor frequent updates, while revealing how customer waiting can undermine an otherwise effective learner.

Learning from decisions that hide information

The simulated florist starts each season with 15 bouquets and has 20 selling periods. Bouquets cost $50, an always-stocked vase costs $20, and the seller chooses both a sponsored-search bid and a discount for buying them together. Unsold bouquets have no value at season’s end.

The seller knows the mathematical structure of demand but not three parameters describing advertising responsiveness, market strength and variation in customers’ willingness to pay. It also operates under an assumed scarcity effect: displaying lower remaining stock raises willingness to pay.

The information problem has two layers. Without an advertising bid, organic visitors still arrive, but their arrival rate reveals nothing about paid-search responsiveness. Purchase choices become visible only after someone clicks. A seller can therefore continue making sales while learning nothing about whether advertising would help.

Mohammadi’s agent maintains probabilities over 125 candidate markets. Before selling begins, it calculates the optimal season plan for each candidate. During selling, it updates those probabilities from clicks and purchases, samples a candidate and looks up that candidate’s recommended bid and discount for the remaining stock and time.

This is posterior sampling, also called Thompson sampling. Its randomness selects among plausible market explanations rather than arbitrary actions. The preferred version draws again every period, allowing it to respond immediately when observations contradict its previous choice.

Frequent updates reduced the learning bill

The study measures performance using regret: revenue forgone relative to a seller that knows the true market and chooses optimally within the same model.

Averaged over markets drawn from the starting distribution, the frequently updated agent accumulated $33.60 in regret over 500 seasons. Sampling once per season produced $109.40, while a planner using posterior-average estimates produced $74.00. These are cumulative losses over all 500 seasons, not losses repeated each season.

In the two weak-market tests, frequent sampling lost $91.30 and $160.70 over 500 seasons, compared with $680.50 and $712.50 for once-per-season sampling. Early unsuccessful offers could quickly discredit an overly optimistic market sample, allowing the agent to change course before the season ended.

The method did not win everywhere. In the strong-market baseline, the average-estimate planner lost $11.40, versus $25.40 for frequent sampling. Its decisions already generated useful information, so additional exploration carried a cost.

A separate stress test exposed the no-advertising trap. With 98% of its initial advertising belief assigned to an unresponsive click curve, the average-estimate planner never bid during the reported 1,000-season experiment. Frequent sampling first bid after about three seasons on average; once-per-season sampling took about 51.

When changing discounts backfires

Scarcity simplified one part of the problem. In markets where the assumed scarcity response kept every shopper’s willingness to pay above the bouquet’s list price, the optimal bundle discount depended only on the known vase preferences. Distinguishing among those strong-market explanations was unnecessary for choosing a policy.

Waiting customers created the opposite problem. In an extension, shoppers could delay when the current discount was smaller than its recent level. Their initial visits appeared as clicks without sales, pushing a seller that ignored waiting toward weaker-demand explanations.

Holding one discount throughout each season eliminated waiting under that particular behavioral rule. The best constant-discount plan sacrificed 5.9% of the fully informed, no-waiting benchmark’s revenue in a weak market, but nothing in the strong market.

Evidence from a model, not a shop

No retailer records informed the calibration. The model also charges every click at the bid, including organic clicks, overstating advertising costs. Its waiting shoppers respond to recent discounts rather than anticipating future markdowns, and its scarcity response is assumed.

The paper provides a regret guarantee for once-per-season sampling, not the better-performing frequent-update version. Its practical promise therefore rests on these simulations: keeping plausible alternatives alive can prevent decisions from cutting off the evidence needed to improve them.