Livestream shopping’s AI assistant: orders rose, but authority stayed with the buyer

External-source research synthesis, not an OfferSignal client result. Based on the retained October 2026 research and source-specific dates.

A livestream seller can present products to a large audience but cannot answer every viewer’s question about features, coupons, returns or an existing order. A randomized field experiment in China tested whether an information assistant could address that service-capacity problem without giving the system autonomous buying authority.

What the experiment tested

The peer-reviewed study by Lingli Wang and six co-authors examined an unnamed leading Asian e-commerce platform’s livestream-selling unit. The platform and participating merchants remain confidential; an author’s Alibaba Group affiliation does not identify the service. The intervention ran 19–24 February 2021. Assignment used the final two account-ID digits: 00–49 received the assistant and 50–99 did not. From more than 10 million broader participants, the platform supplied a random 1% analytic subset of 132,199 consumers who watched at least one livestream: 65,902 treatment and 66,297 control. Reported balance checks found no significant observed differences across consumer, streamer or livestream characteristics. 1 2 3

The assistant monitored comments using speech recognition and natural-language processing. It combined product-page detail, the seller’s return/exchange policy and livestream-room discounts. When it detected needs such as product information, coupons or order help, an icon invited the viewer to ask. A private thread could provide text or voice support, product recommendations, coupon information and order tracking. The buyer still selected products and placed orders. The study does not show supplier comparison, negotiation, payment authorisation, fulfilment commitment or a binding purchase by the assistant. 2

The result was an order effect, not autonomous commerce

Across the 132,199 analytical consumers, treatment access produced a reported 3.00% increase in orders placed (zero-inflated Poisson coefficient 0.030, p < 0.05). Product clicks and cart additions were not statistically significant: coefficients 0.0081 and 0.0052, respectively, both p > 0.1. The study’s #Pay measure counts order placements, not payment settlement, merchant revenue or profit. 1 2

The paper also reports 10.95% higher normalized consumer spending (OLS coefficient 0.00023, p < 0.01). This is not a currency amount, gross merchandise value, margin or return on investment: raw prices were withheld and normalized against the experiment’s highest product price. Assistant costs are not published. 2

Among 35,164 consumers who placed an order, the return rate was 0.03511 with assistant access versus 0.04015 without it: an absolute decline of 0.504 percentage points and a relative decline of 12.55% (generalised linear model coefficient -0.139, p < 0.05; coarsened-exact-matching check -0.140, p < 0.05). This denominator is purchasers, not all viewers. The return follow-up window, reasons and raw values are not public. 2

Information can help and interrupt

A separate online video experiment analysed 150 Chinese public-university participants after removing 18 of 168 completions for an attention-check failure. Those seeing an assistant scenario reported more intelligent information provision (5.755 versus 4.446), more interruption (4.026 versus 2.914), less uncertainty (3.763 versus 4.399), higher purchase intention (5.270 versus 4.547) and higher decision confidence (5.632 versus 4.797). This suggests a trade-off, but participants imagined shopping; it was not a live-customer purchase or return test. 2 3

Field effects were more evident for new products—without prior sales, reviews or ratings—and for streamers above the pre-experiment median follower count. Of recorded AI interactions, 74.66% were proactive and 25.34% reactive. The authors explicitly treat these mode comparisons as correlational because consumers’ choices were endogenous, not as proof that proactive prompting caused the outcome. 2

A bounded application for commerce teams

For an illustrative manufacturer, distributor or retailer, the transferable pattern is an evidence-grounded shopping copilot for compatibility or fit questions, current price and policy explanation, and a human-reviewed request. Its record should retain product or configuration identity, source, seller, timestamp, price basis, availability or lead time, returns or warranty terms, and the correcting owner. The assistant may answer, explain, recommend or prepare a request; a named buyer or commercial owner should approve any quote, commitment or payment.

Test the whole path: source-grounded answer, prompt, engagement, order, return and return reason. Include non-engagement and interruption as possible harms. This short historical experiment does not establish a modern large-language-model architecture, long-term adoption, European or business-to-business performance, autonomous procurement or economic ROI.

The final paper supersedes its 25 February 2023 SSRN working paper, which reported +2.61% sales and -62.86% product returns for the same stated sample. The final version reports +3.00% orders and -12.55% return rate; no public revision memo explains the change, so the figures must not be combined. 4

Sources and next reading

Explore OfferSignal · Focused Business & AI Workshop · All research

References