A 6M-Customer Field Experiment Found More Purchases With Less Exploratory Browsing

Generative search did not simply add another interface in a large commerce experiment. It changed where the customer journey began — and made some familiar digital signals harder to read in isolation.

2 October 2026 · 4 min read

What happened

Marketing Science Institute highlighted Generative Search: Evidence from a Large-Scale Field Experiment on 18 September 2026. The paper, accepted by Management Science, studies more than six million customers on Meituan, a major Chinese commerce platform.

The experiment compares a traditional search experience with a generative-search experience that can respond to a customer's natural-language expression of intent before conventional results.

The authors report three material findings: generative search significantly increased purchases; it enabled more effective and more diverse keyword queries; and it reduced exploratory browsing and clicking while concentrating evaluation within more relevant categories and merchants.

A separate NielsenIQ survey released on 24 September provides market context rather than causal confirmation: 51% of U.S. consumers in its tracker reported using at least one AI-powered tool to support shopping in the previous month. That does not prove the Meituan effect travels to the U.S., but it makes the underlying behavior worth watching.

What the evidence actually says

Observed fact. The Meituan study is a large-scale field experiment, not a survey of intentions. Its public abstract reports a statistically meaningful increase in purchases alongside less exploratory browsing and clicking.

Reported mechanism. The authors say the evidence is most consistent with generated answers helping people interpret their underlying needs, expanding awareness of relevant attributes and directing attention toward more relevant categories.

Inference. Search is moving one step upstream. Traditional keyword search starts after a person has translated a need into words a search engine can use. Generative search can participate in that translation. That makes problem formulation — what the person is actually trying to achieve — part of the observable journey.

Unknown / limitation. The public abstract does not provide the purchase-lift effect size. The experiment is on one Chinese platform and should not be generalized to every category, market, search engine or conversational surface. It also does not establish that every reduction in clicking is positive.

Why this matters for marketing

Digital marketing has spent years learning to read journeys through observable actions: searches, impressions, clicks, sessions, product views and conversions.

That logic becomes less complete when an interface can absorb part of the exploration itself.

A customer may ask a broad question, receive a synthesized explanation, refine the need, and arrive at a smaller set of relevant options with fewer intermediate clicks. In that journey, lower browsing activity can coexist with better commercial outcomes.

The consequence is not "stop measuring clicks." It is more specific: do not confuse the amount of visible exploration with the quality of the decision process.

That matters for search, commerce, content, customer research and measurement. If the upstream problem is changing, a downstream dashboard can remain technically accurate while describing less of what caused the choice.

The operating implication

Marketing teams should add one object to the way they review search and discovery: the expressed need.

For material journeys, capture the problem the customer is trying to solve, the evidence or attributes surfaced in response, and the eventual choice. Then review clicks and conversions inside that context.

The second move is to separate efficiency from disappearance. If exploratory clicks fall, ask whether people are reaching better-fit categories and outcomes — or simply losing access to alternatives. The Meituan experiment supports the first possibility in its setting; it does not remove the need to test the second.

What not to do: rewrite the search strategy around one study. Treat this as a strong signal to improve the measurement question, not as a universal rule.

The Olymuse lens

This connects to Marketing Process Intelligence.

In Marketing Is Not a Linear Workflow, the core idea is that marketing decisions loop through evidence, interpretation, action and revision. Generative search makes that loop visible earlier in the customer journey.

Our point of view is simple: when a journey starts from an intention rather than a keyword, the useful memory of that journey should not begin at the first click. It should preserve the need being interpreted, the evidence that shaped the choice and the assumptions used when the team later judges the result.

That is an operating principle, not a product-performance claim. A team can begin applying it in research and journey reviews today.

What we would watch next

First, replication: do comparable field experiments outside Meituan and China find the same combination of fewer exploratory actions and stronger outcomes?

Second, effect sizes by category and task: does generative search help most when needs are ambiguous, advice-heavy or attribute-rich, and less when the customer already knows exactly what to buy?

Third, measurement practice: do search, commerce and analytics platforms begin exposing intent-level evidence rather than forcing teams to reconstruct the journey only from traffic events?

Those three observations would tell us whether this is a local interface effect or the beginning of a broader change in how marketing should read discovery.

Read next: Marketing Is Not a Linear Workflow · Olymuse is in founder-assisted private alpha. Apply for the Private Alpha →