A search box can wait for a customer’s words or offer a starting point. That small choice can change what the customer considers buying. It can also divert an existing intention.
The operator’s decision is which terms to suggest and how to judge them. Suggestion clicks are useful diagnostics. The test is whether customers find suitable products and whether the store gains useful purchases after displaced activity is counted.
🔎 What research says
Test relevant prompts against the current search bar. Count displaced queries, follow effort beyond typing, and distinguish newly discovered purchases from purchases redirected away from familiar products or sellers.
🗺️ In this guide
1. Test relevant prefilled terms against an empty bar
📈 Recommendation
For shoppers who have not chosen a product, test a relevant category prompt before filling the search bar for everyone. Include the existing empty bar as a comparison.
For a homeware store, a seasonal term such as “picnic blanket” could be one test candidate. Compare it with a term based on recent browsing and the current search bar. These are proposed store examples, not the study’s exact keywords.
Judge each version by purchasing customers and spending per assigned customer. Decide in advance how you will detect people losing a route to the product they wanted.
🎓 Findings
Sun (2025) randomly assigned 72,587 customers on an ecommerce platform to an empty search bar or prefilled trending, personalized or niche category terms. These were real shopping outcomes, not imagined choices.
The published abstract reports 10.4% higher purchase incidence and 8.8% higher spending with trending terms versus an empty bar. These are relative changes, not percentage-point gains. Niche terms produced little detectable benefit.
Illustration, not a study result: if average spending started at €10 per assigned customer, an 8.8% relative increase would take it to €10.88: €10 multiplied by 1.088. That is 88 cents more per customer in the illustration, not €8.80 and not a forecast for your store.
🧠 Why it works
A useful term gives an undecided shopper somewhere relevant to start. A narrow, unrelated term may simply occupy valuable space. Relevance is the hypothesis to test; filling the box is only the implementation.
✋ Limitations
The headline estimates are verified in the primary abstract; the complete main article was not accessible. The accessible appendix supports the direction but does not resolve every outcome definition, time window or confidence interval.
This study does not establish that trending always beats personalization. A niche result without a clear benefit also does not prove that every specialist category is ineffective.
2. Count the queries your suggestions displace
📈 Recommendation
When personalizing the prefilled term, measure whether it replaces the shopper’s own search. More use of the suggested term can coexist with less searching overall.
For a shoe retailer, retain separate counts for accepted suggestions and self-entered terms. Keep purchases and spending on the same randomized-customer denominator.
Make the prompt easy to replace. Check repeat purchasers and customers returning to a previously viewed category. This editable design is our proposed safeguard; its exact interface was not separately tested.
🎓 Findings
Sun’s official appendix, Table B, separates searches using personalized terms from other searches. In its robustness analysis, personalized prefill added 2.234 suggested-term queries per customer, while other queries fell 4.252. Total queries therefore fell 2.018 per customer: 2.234 minus 4.252.
The reported standard errors are 0.361, 0.393 and 0.172 respectively; each difference is statistically clear in this table. These are average query counts, not percentages or lost buyers. The decomposition uses constructed control-group comparisons, so it is not a literal record of which individual query replaced another.
🧠 Why it works
A customer who recently viewed running shoes may accept that suggestion instead of typing “school shoes.” Some accepted suggestions redirect activity that would already have happened. The business question is whether the redirection improves the purchase, rather than whether the prompt attracts a tap.
✋ Limitations
Fewer queries did not automatically mean worse commercial performance in this study. Personalization also increased product exploration and purchasing measures. Do not treat a displaced query as a lost order.
The main article’s detailed measurement definitions remain unresolved. The count decomposition is stronger evidence of a trade-off than a universal target for how many typed queries a store should preserve.
3. Measure effort after the query is entered
📈 Recommendation
Separate the work of starting a search from the work of choosing among results. A prompt can save typing while creating a longer comparison task.
For grocery search, record time to submit a query separately from time spent comparing products and reaching checkout. Inspect abandonment alongside completed purchases.
For a broad suggestion such as “quick lunch,” test whether the results let shoppers narrow the range by a meaningful need. That result-page treatment is a local test proposal, not a proven finding from this experiment.
🎓 Findings
Zheng and colleagues (2026) randomized exposure to a list of popular search genres in Meituan’s food-delivery interface. The primary abstract reports shorter, broader queries and less time generating them, but more subsequent product exploration, alongside more purchases.
No trustworthy estimate of minutes saved, extra browsing time or purchase uplift was recovered from the accessible record. The finding establishes a directional trade-off; it cannot supply a numerical efficiency target.
🧠 Why it works
“What should I search for?” and “Which of these should I buy?” are different jobs. A broad prompt can solve the first while expanding the second. A seller should investigate where time moves, not celebrate a faster first tap in isolation.
✋ Limitations
The main paper and its detailed numerical tables were not available for review. The supplement link was identified, but its content could not be fetched reliably. The sample size and precise uncertainty remain unverified.
A ranked genre list in food delivery is not the same treatment as a single prefilled ecommerce keyword. Treat the proposed grocery scorecard as a way to test transfer.
4. Separate new purchases from redirected purchases
📈 Recommendation
If suggestions introduce new products or sellers, count the purchases they replace as well as the purchases they create. Assess the result against the business objective agreed before the test.
For a marketplace, compare orders at previously unvisited sellers, familiar sellers and all sellers. For a single retailer, a similar category-based analysis is a proposed adaptation.
For a rollout decision, review total paid orders, customer suitability and seller distribution together. A wider spread of demand and a larger total are different goals.
🎓 Findings
Zheng and colleagues (2025) studied a query recommender in an Asian mobile food-delivery app. Its primary abstract reports higher purchase volume and broader consumption across customers and the market.
The precise split between newly discovered and familiar sellers was not independently verified for this guide. No numerical estimate of that split is given here. The proposed scorecard checks whether discovery creates purchases or moves them between destinations; it is an operator’s test, not a claimed result from an unverified table.
🧠 Why it works
A useful discovery feature can change both the size of demand and where it goes. A seller gaining orders and a marketplace gaining orders are different outcomes. Counting only the destination that benefits can hide the full commercial result.
✋ Limitations
The complete main paper and detailed merchant-level tables were not available for independent review. The abstract supports the direction of broader purchasing, but cannot establish the size of a new-versus-familiar merchant trade-off.
This paper and the 2026 Meituan study share researchers and related settings. Their dataset independence remains unresolved; they should not be presented as two confirmed independent replications.
💡 The takeaway
Choose suggestions by their contribution to the whole shopping task.
Start with one relevant category and an empty-bar comparison. Keep the customer’s own route available. Expand when purchasing outcomes improve on the same customer denominator and the result remains worthwhile after changes in search effort and purchase destinations are understood.
📚 Sources and study types
Sun, C. (2025). How Does Prepopulating Search Bars with Keywords Affect Online Consumer Behavior? A Field Experiment. Marketing Science, 44(6), 1217–1231. Randomized ecommerce experiment; primary abstract and official appendix inspected. Full main article not reviewed.
Zheng, S., Tong, S. (J.), Fang, S., Gopal, A., Li, X., & Jiang, Q. (2026). Priming in Search: A Large-Scale Field Experiment on the Impact of Popular Search Genres in Mobile Shopping. Information Systems Research, articles in advance. Randomized Meituan food-delivery experiment; primary abstract inspected. Numerical table extraction remains unresolved.
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 food-delivery experiment; primary abstract inspected. Main-article and merchant-level numerical review remain incomplete.
