Shoppers given picture-and-text suggestions for refining broad searches bought more over time in an experiment on a large Chinese ecommerce platform, according to research distributed by the National Bureau of Economic Research.
The feature had no immediate effect on orders or spending. Over subsequent months, however, the researchers found higher purchasing among people assigned access to the tool, with a particularly pronounced response for products outside the platform’s top sellers.
The working paper by Lu Fang, Yanyou Chen, Chiara Farronato, Zhe Yuan and Yitong Wang examines one platform and a particular design. It was circulated in 2024 and is not an NBER peer-reviewed publication. One coauthor was employed by the company that supplied the data.
Turning a broad search into a specific one
Someone searching for headphones may know they want a wireless pair without knowing which product terms will produce useful results. A list of suggestions can help them express that requirement and show them choices they might not have considered.
In the experiment, the platform presented pictures alongside suggested refinements when a customer entered an eligible general search. Selecting a suggestion changed the search to a narrower query. A search for a dress, for example, could be refined by style.
The researchers call the feature picture-text guidance. It differs from a conventional filter: instead of restricting an existing list by price or brand, the feature helps a shopper choose the words used for a subsequent search.
Not every search triggered the tool. The platform selected eligible general queries according to their popularity and whether they could be divided into narrower categories. Suggestions were created through a combination of algorithmic methods and human curation.
Why the first day was not the whole story
The experiment ran from March to December 2021 on the platform’s mobile app. Customers were randomly assigned to a group that could see the feature or a control group that could not. The analysis focused on logged-in customers in China who searched for an eligible general term.
On the first day after entering the experiment, people with access to the feature were more likely to use a suggested specific query. There was no immediate increase in the number of orders or amount spent.
Over the following months, purchasing rose in the group with access to the feature. The researchers found no corresponding increase in overall searches or product views, which they interpret as evidence that shoppers were finding suitable items more efficiently, not simply spending more time browsing.
The paper also reports changes beyond searches directly eligible for the tool. The authors argue that some customers learned to use more specific search terms on their own. Their data cannot identify every reason why an individual selected a particular product.
What smaller sellers can take from the results
Search design affects whether a customer finds an item in the first place. In the experiment, the benefits were especially visible for products outside the best-selling group, suggesting that guidance can help people discover a wider range of options.
A retailer assessing its own search could start by examining broad queries that return many unrelated items. Search logs, customer-service questions and unsuccessful searches may show where shoppers need clearer product attributes or plain-language alternatives.
Any change should be tested against measures beyond clicks on the new feature. Completed orders, returns, customer satisfaction and the ability to find relevant products can show whether the tool helps customers as well as attracting attention. Our earlier coverage examined how a product-page cue affected the next step in a shopping journey.
The study does not establish that pictures alone caused the improvement: images and text were introduced together. Nor does it show that the same results would follow on a smaller retailer’s site, in another country or with a different set of suggestions.
For a business running its own experiment, the delayed result is a reason to agree on an appropriate observation period before declaring a new search feature ineffective.