Why accurate AI shopping advice may not leave consumers better off

A model of shopping, not a market experiment

AI shopping tools can recommend alternatives or select a product on a buyer’s behalf. Do those different roles lead to different prices and consumer outcomes? Xiaomeng Guo, Fasheng Xu and Xiaoli Zhang address that question in Artificial Shoppers and Attention Competition, an arXiv preprint submitted on November 6, 2026.

Their study is theoretical. It does not test a commercial assistant or measure actual shopping behavior. Instead, the authors solve a simplified economic model in which sellers anticipate how an interface will shape purchases and set prices accordingly.

The market contains three products with high, middle and low quality. A platform hosts two and collects a fixed commission on their sales. The remaining product is sold outside the platform. Without AI, consumers consider only the hosted pair. Both AI modes make all three accessible.

Consumers differ in their willingness to pay for quality. Their underlying benefit from a purchase is the value they assign to its quality minus its price. The paper calls this “true utility” and uses it to evaluate consumer welfare.

Recommending versus selecting

When consumers compare products themselves, their choices are affected by salience: they put extra relative weight on whichever attribute, price or quality, stands out in the comparison. Changing the neighboring alternative can therefore change a purchase even when the buyer’s underlying preferences remain unchanged.

Under assistance, AI knows each consumer’s preferences perfectly and recommends the two products offering the highest true utility at posted prices. Consumers then make their own salience-distorted comparison.

Under delegation, an equally informed agent selects the product with the highest true utility. The consumer retains purchase approval, modeled as an undistorted decision to buy only if that utility is nonnegative.

These are idealized rules, not descriptions of how reliably existing AI systems understand buyers. Neither rule ranks products to maximize advertising income or platform commissions.

More access can mean less aggressive pricing

The assistant’s effects depend on which products consumers could previously access.

If the platform originally hosts the high- and low-quality products, assistance introduces the middle product as an alternative. In every characterized assistant equilibrium, both hosted prices fall, platform profit falls and aggregate consumer surplus rises.

If the platform instead hosts the high- and middle-quality products, assistance can raise both hosted prices. The mechanism centers on the middle-quality seller, which appears beside the premium product for some consumers and beside the cheapest product for others.

That seller charges everyone the same price. Cutting it to attract buyers considering the cheapest alternative also reduces revenue from buyers willing to pay more. In some equilibria, the seller therefore abandons lower-valuation sales to protect its premium business. Its retreat softens competition with the high-quality seller, allowing both hosted prices to rise.

Broader recommendation exposure thus need not translate into broader sales or stronger price competition.

The alignment paradox

The paper’s “alignment paradox” separates accurate ranking from beneficial decision-making.

Some consumers who previously bought the high-quality product switch to the middle-quality product after receiving a new shortlist. Their former choice remains recommended and has become cheaper. Nevertheless, the new comparison makes its price premium loom larger relative to its quality advantage.

For harmed switchers, the lost quality is worth more than the saving relative to their original purchase. The assistant correctly ranks individual products but does not choose the shortlist to maximize welfare after accounting for the consumer’s distorted comparison.

This is not a claim that every consumer loses. With the original high–low assortment, aggregate consumer surplus increases despite possible individual losses. Here, consumer surplus means the combined underlying benefits consumers receive after paying prices; an aggregate gain does not guarantee that everyone benefits.

Delegation removes this comparison-induced problem at given prices. But sellers respond to the agent’s undistorted valuation of quality. That response can support higher prices, leaving some consumers—or consumers in aggregate—worse off than before.

Platform incentives and limits

Holding the hosted assortment fixed, both platform profit and aggregate consumer surplus can rise under assistance with the high–middle pair, or under delegation with the high–low pair. These are conditional possibilities, not universal benefits.

When the platform chooses its design, it prefers the high–low pair without AI. Among AI designs, it weakly prefers delegation with the high–middle pair wherever all compared pricing equilibria exist. Yet that preferred AI configuration can benefit both parties, hurt both, or benefit one at the other’s expense relative to the preferred no-AI configuration.

Important boundaries remain. The model assumes three products, two-item shortlists, zero marginal production costs, uniformly distributed consumer valuations and perfect AI preference information. Adoption, seller participation and hosting terms are taken as given. Some assistant parameter regions lack a pure-strategy pricing equilibrium; mixed-strategy outcomes are not characterized, limiting the platform-design conclusions there.

The practical implication is a question for evaluation, not a demonstrated market outcome: assess what people ultimately buy and how sellers adjust prices, rather than treating accurate recommendations alone as proof of consumer benefit.