A recommendation can start a shopping journey. Search can refine it, rescue it or take it somewhere new. Looking at either tool alone misses what the customer is trying to do.
The decision is how to make recommendations and search work as one product-finding journey. More searching is not automatically better. Neither is less. The useful outcome is a suitable purchase reached without unnecessary work.
🔎 What research says
Read search growth in context. Follow the same shopping task across recommendations, category browsing and search. Test query suggestions alongside refinement tools, then judge the result by useful choices and purchases.
🗺️ In this guide
1. Investigate search growth before calling it success
📈 Recommendation
When search use rises after a recommendation change, check the rest of the journey. People may be exploring a new idea—or looking for what the recommendations failed to show.
For a luggage store, distinguish a broad query such as “weekend bag” from a precise model search. Check whether each follows a recommendation and ends in a suitable purchase. This classification is our proposed store analysis.
Compare recommendation clicks, search clicks and completed purchases for the same randomized visitors. A channel can gain activity while the overall journey gets worse.
🎓 Findings
Yuan and colleagues randomized 555,800 customers on a mobile marketplace. One group’s homepage recommendations stopped using personal data; search itself stayed unchanged.
In their 2023 author manuscript, Table 1, search clicks rose 6.3%, while recommendation clicks fell 83.6%, relative to the personalized group. These are changes in average click counts, not conversion rates. Search growth was partly a response to worse recommendations.
The published paper also distinguishes broad queries associated with developing an interest from more specific queries used to pursue an existing need. One total search count cannot distinguish those tasks.
🧠 Why it works
Recommendations offer a starting point. If that starting point misses the task, shoppers may take over through search. If it sparks an idea, search may help them refine it. The same extra search can therefore have different meanings.
✋ Limitations
The numerical estimates come from the author manuscript’s one-day experiment and have not been matched to the final published table. Removing personal data was a large intervention; it does not predict the effect of a small carousel change in your store.
The study supports keeping the channels under joint review. It does not establish that every increase in search is a failure or that broad queries always indicate useful discovery.
2. Measure progress through the whole product-finding task
📈 Recommendation
Judge search and recommendation use against the product the shopper needs. A customer who finds a suitable item with fewer search pages has not necessarily disengaged.
For a clothing store, follow one product task across recommendations, category browsing and keyword search. Keep a clear route to each tool; do not remove search just because its use falls.
Measure suitable purchases, returns and total journey effort. Use one shared task denominator so a shopper moving between tools is not counted as several different customers.
🎓 Findings
Wan, Kumar and Li randomized recommendation availability at an apparel retailer. Their author manuscript analyzed 573,665 purchase funnels from 430,349 visitors. A funnel followed a visitor’s search within a subcategory until purchase or the data period ended.
Table 12 estimates 0.02 fewer search-result pages and 0.10 fewer category pages per funnel when recommendations were available. Expressed across 100 comparable funnels, those averages are about two fewer search-result page views and ten fewer category-page views. They are not two or ten fewer shoppers.
The paper links benefits to finding lower-priced or better-fitting products, rather than easier navigation alone. That is why lower search use needs an outcome check.
🧠 Why it works
A useful suggestion can replace a trip back to the category page. That saves an extra step only if the product fits. Counting fewer searches as either success or failure skips the actual shopping task.
✋ Limitations
The page-count estimates come from the author manuscript and one apparel retailer. One visitor can contribute several funnels; the results are not a per-visitor conversion lift. Product fit was partly inferred from algorithmic measures.
Our suitability-and-returns scorecard is a local test proposal, not the exact intervention the researchers tested.
3. Pair query discovery with a way to refine it
📈 Recommendation
Test suggestions of what to search alongside tools that help shoppers express their own intent. Suggesting a new direction and completing a query do different jobs.
For a grocery store, distinguish an adjacent suggestion such as a different meal idea from autocomplete that finishes the food or brand someone already started typing. Keep the typed query under the shopper’s control.
Compare the existing search interface with a version adding query suggestions while retaining autocomplete. Track orders, new useful categories and unsuccessful searches; inspect frequent and infrequent autocomplete users separately.
🎓 Findings
Zheng and colleagues (2025) studied query suggestions in an Asian mobile food-delivery app. The electronic companion reports 58,519 users; the field comparison lasted 30 days.
The published paper reports a 1%–2% relative increase in purchase volume with query recommendations. Illustration, not a measured study total: if a comparable group placed 1,000 orders, that relative change would mean 1,010–1,020 orders. It does not mean one or two extra buyers per 100 visitors.
The electronic companion shows more generic queries without a clear fall in specific queries. Benefits also varied with prior autocomplete use. This supports testing discovery and refinement together, rather than treating every suggestion as the same feature.
🧠 Why it works
An adjacent query can introduce a direction the shopper had not considered. Autocomplete can then help turn an idea into an actionable search. These are plausible ways to support exploration without taking control away from the customer.
✋ Limitations
Food delivery involves frequent purchases and a mobile interface. The reported range is not a conversion-rate gain or a forecast for durable goods. The main article’s detailed uncertainty remains unresolved.
Prior autocomplete reliance was observed, not separately randomized. The subgroup pattern does not prove that adding autocomplete causes a special extra boost. Test the combination in your own setting.
💡 The takeaway
Give shoppers a useful starting point and a clear way to take control.
Start with one product-finding task. Compare recommendation and search behavior for the same visitors, distinguish exploration from precise retrieval, and check whether people find suitable products. Expand a change when the whole journey improves—not merely because one tool gets more clicks.
📚 Sources and study types
Yuan, Z., Chen, A. J. Y., Wang, Y., & Sun, T. (2025; online 2024). How Recommendation Affects Customer Search: A Field Experiment. Information Systems Research, 36(1), 84–106. Randomized marketplace experiment. The cited click estimates are from the linked 2023 author manuscript, Table 1; final-table parity remains unchecked.
Wan, X. (S.), Kumar, A., & Li, X. (2024; online 2023). How Do Product Recommendations Help Consumers Search? Evidence from a Field Experiment. Management Science, 70(9), 5776–5794. Randomized apparel-retailer experiment. The cited search-tool estimates are from the author manuscript’s Table 12.
Zheng, S., Tong, S. (J.), Kwon, H. E., Burtch, G., & Li, X. (2025). Frontiers: Recommending What to Search: Sales Volume and Consumption Diversity Effects of a Query Recommender System. Marketing Science, 44(3), 516–524. Randomized mobile food-delivery experiment; published effect summary and electronic-companion query/subgroup analyses.
