AI Shopping Assistants That May Have Their Own Agendas

A growing share of online shopping now begins with a prompt rather than a search box. Adobe Analytics clocked generative-AI traffic to US retail sites surging almost 8-fold during the 2025 holiday season, and industry projections put agentic commerce at roughly US$1 trillion of retail revenue by 2030. The question of consumer economics is no longer whether these advisers work, but rather whom they are working for.

Choosing what to buy through key word searches, user reviews and livestream promotions each carry a visible bias. Search answers often cough up paid ads as top choices. Reviews can be fabricated. Livestreaming leverages the personality of the sales host. In China's, iResearch estimated that product returns in livestreaming e-commerce were about 36 percent in 2024, more than double that of conventional platforms, with clothing returns hitting as high as 50 percent.

A conversational shopping agent is a different kettle of fish. It can unpack a vague request, condense thousands of specifications and weigh trade-offs in real time. The danger, however, is behavioral rather than technical. Users instinctively grant some trust to machine-dialogue partner that sounds neutral and impartial. A sponsored result can be scrolled past or filtered out, but a model that both answers question and carefully frames recommendations blurs that boundary until it disappears.

In a Princeton University study, 2,012 participants were asked to each select a book from a real e-book catalogue, with one in five titles secretly earmarked as sponsored. The results found that persuasion generated by a large language model nearly tripled the share of people who chose a sponsored product to 61 percent. Inclusion of an explicit "sponsored" label barely moved the needle, trimming the rate to 55.5 percent.

More troubling, participants rated persuasive chatbots more satisfying to use than a plain search, even as their choices were quietly being manipulated.

Adobe figures capture the ambiguity. Shoppers who used AI assistants lingered 45 percent on a sales site, and 72 percent of said they were less likely to return what they bought. They also spent more on purchases.

The situation also exists in China's e-commerce. JD.com's AI shopping guide Jingyan has served more than 50 million consumers. During the peak hours of its 2025 sales during the Singles' Day shopping festival, the platform's algorithmic guide fielded over 4.8 million sessions while lifting pre-sale conversion to sales by 37 percent. On the merchant side, JD's AI customer-service tool Jingxiaozhi, now supporting more than 900,000 vendors, claimed a customer-retention lift of up to 36 percent,

Such numbers are genuinely useful metrics to sellers but do they serve buyers? Using algorithms that serve both sides flatters a platform's profile but obscures the impact on shoppers. That is why headline "improvement" figures deserve a second look.

A shopper can run simple tests. Ask the assistant to spell out its ranking rules – that is, whether commissions, ad weighting or inventory drive recommendations. Rephrase the same request and compare the answers. If results shift markedly once sponsored items are excluded, commercial incentives are at work. And since detection of promotions can be near-blind even in controlled trials, healthy skepticism is the sensible default.

Time saved is a weak yardstick for seeking out value in purchases. A hasty purchase can simply be based on a faster impulse. In assessing the benefits to shoppers, one needs to consider shorter decision times, lower return rates, narrower gaps between expectations and the delivered product, and satisfaction that endures. Notably, in the Princeton trial, many choices made under persuasive conditions did prevail even after full disclosure.

Governance completes the picture. Disclosing commercial weightings, separating advisory functions from sales targets and letting users switch off sponsored rankings would align these tools more with consumer welfare. An AI shopping assistant is not inherently a good or bad decision-maker. It serves whoever writes its software, but final choices still remain with humans.

(The author is an assistant research fellow at the Institute of Journalism at the Shanghai Academy of Social Sciences.)