E-commerce now regularly features livestream shopping, a format where sellers showcase products on live video to engage audiences and generate sales. While industry revenues are projected to reach hundreds of billions of dollars globally, individual seller success is highly variable. Many sellers abandon their livestreaming efforts within a year.
A study published in the International Journal of Research in Marketing sought to identify the specific factors that predict a livestream seller’s sales performance. A research team featuring Zekun Liu of Indiana University, along with colleagues from Rice University, New York University, and Alibaba Group, developed a model to evaluate how various livestreaming strategies relate to business success.
Looking at Seller-Hosted Streams
Past research on livestream shopping has primarily focused on sessions hosted by paid influencers. However, the industry has experienced a major shift toward broadcasts hosted directly by business owners. In these sessions, sellers promote their own brands and maintain full control over what products they feature and how long they stream.
The researchers wanted to build a predictive tool that went beyond basic store characteristics, such as past transaction volume or subscriber counts. They designed an investigation to test whether incorporating specific livestream metrics could more accurately forecast which sellers would generate the highest sales.
Tracking Store Data
To test their model, the team analyzed data from Taobao Live, a major livestream shopping platform owned by Alibaba. They selected a sample of 4,061 small business sellers operating in the fashion, food, and jewelry categories. All the sellers in the sample launched their livestreaming channels between August 2019 and February 2020.
The researchers tracked each seller’s sales performance and livestreaming activities for five months following their first broadcast. The collected data covered session habits, product assortment choices, social engagement metrics, and platform-generated scores related to content quality and seller responsiveness.
The team fed this data into a machine learning algorithm designed to predict monthly sales. They also applied an analytical technique to break down the algorithm’s predictions, allowing them to see exactly which individual metrics carried the most weight.
Better Predictions
The analysis revealed that adding livestream-specific metrics improved the accuracy of sales predictions by 25.6 percent compared to relying solely on traditional store history. This improvement was consistent across different types of sellers but was especially pronounced in the jewelry category.
The length of time a seller had been operating also changed which factors mattered most. For newly established sellers, product assortment choices and content quality scores were the strongest predictors of sales. For more experienced sellers, content quality and responsiveness scores held the most predictive weight.
Session and Assortment Strategies
The data showed clear patterns regarding how sellers structured their time. High livestreaming frequency and a high total number of broadcast hours within a single month were linked to higher sales. However, an extremely high accumulation of lifetime broadcasting hours over many months was associated with lower sales. The researchers suggest this negative association might reflect audience fatigue or content oversaturation.
Product selection decisions were also strongly tied to performance. Streams that featured higher-priced items tended to predict better sales outcomes. The pace of the broadcast played a role as well, with higher sales linked to a rapid presentation style that introduced a large number of products per hour.
Variety was another key indicator. Sessions that presented a mix of recently added inventory alongside older items performed better than those focusing exclusively on long-standing products. Sellers also saw better outcomes when they featured products that were moderately popular on the platform, rather than items that were either completely unknown or overly saturated.
Meaningful Engagement Over Passive Views
When evaluating how audiences interacted with the broadcasts, the researchers found that raw view counts and simple sharing metrics did not reliably predict sales. Instead, the model prioritized metrics that reflected sustained attention and individual participation.
Average viewer watch time emerged as the most important engagement metric for predicting sales. The number of unique individuals who commented on the stream or clicked the “like” button also served as strong predictors. In contrast, the total number of times a stream was shared had little to no predictive value.
The researchers note that these predictive indicators can serve multiple business purposes. Lending institutions and digital platforms can use these models to assess credit risk when offering financial services to small businesses. Sellers can use the identified metrics to benchmark their performance against competitors.
The authors specify that their model is designed for predictive analytics rather than establishing direct cause and effect. A strong association between a specific metric and high sales indicates a reliable pattern, but it does not guarantee that artificially inflating that metric will directly cause a corresponding spike in revenue.




