Clicks Were Always a Terrible Proxy for Intent

Digital retail knows an extraordinary amount about what people do and surprisingly little about why they do it.

A typical analytics stack can reconstruct acquisition source, search terms, page views, filters, dwell time, cart additions and abandonment. From these traces, the retailer infers intent. A customer who opens paint, filler and roller pages may be redecorating a room. They may also be pricing materials for someone else, repairing a leak or merely comparing brands. The behaviour is observed; the objective is guessed.

This is not a defect in analytics. It is a consequence of the interface. Websites produce clicks because clicks are how users communicate with them.

Conversational interfaces produce a different kind of data. “I need to repaint an 18-square-metre bedroom; there are two cracks and I own no tools” contains the project, scale, resources and one defect before a product has been returned. The retailer does not have to infer the objective from navigation because the customer has already stated it.

From zero-result searches to zero-result intentions

This could materially improve merchandising intelligence.

Traditional site-search analysis is good at identifying zero-result queries. A conversational system can identify zero-result intentions: projects that repeatedly fail because the assortment lacks a component, use cases that customers care about but the taxonomy does not represent, or price-and-feature combinations the catalogue cannot satisfy.

The distinction is subtle and useful. A failed search tells the retailer that a term did not map to a product. A failed intention explains what the customer was trying to achieve when the catalogue failed.

At sufficient scale, this begins to resemble continuous qualitative research embedded in the transaction flow.

The old engagement metrics do not transfer cleanly

The danger is to measure conversation with the same metrics used for websites. More messages are not necessarily better. A ten-turn interaction may indicate expert guidance or an assistant that asked nine unnecessary questions. Longer session duration can mean engagement or confusion.

Agentic systems need task-level evaluation. IBM emphasizes whether the system completed the task, remained faithful to the available information, acted safely and used tools correctly. Retailers can add commercial measures: time to a valid basket, basket acceptance, replacement frequency, missing-item rates, conversion, assisted revenue and the share of intents the catalogue could not satisfy.

This is harder than counting clicks because the unit of analysis is no longer the page. It is the job the customer came to complete.

A richer signal, with a higher governance burden

Conversation data is not automatically superior. People omit information, change their minds and sometimes describe needs poorly. Language can contain sensitive personal information that a clickstream never captured. Model-generated summaries can introduce interpretation errors of their own.

Retailers therefore need strict governance: retention limits, separation between stated facts and inferred attributes, aggregation before merchandising use, and explicit rules for sensitive domains.

Handled well, however, conversational data closes an old information gap. Retailers have spent years inferring intent from clicks because clicks were the only visible trace. Once customers can state their objective directly, some familiar analytics begin to look like elaborate proxies for a variable that has finally become partially observable.

The commercial value of conversation may ultimately be as important on the merchant side as on the shopper side: it reveals not only what people bought, but what they were trying to solve.

Clicks record how customers adapted to the retailer’s interface. Conversation can capture what they wanted before that adaptation.

“The most significant challenge in scaling AI is not technological but organizational.” — Przemek Czarnecki, CTO of ASOS, via McKinsey