When ChatGPT goes shopping: what brands overlook in their customer journey

“Which pram fits into a small car boot and folds with one hand?” Anyone who asks an AI assistant this may already receive a list of suitable models before visiting a single shop. The example is hypothetical. The question for brands is concrete: How is the shortlist formed with which customers arrive on their website in the first place?

A web analytics dashboard can show which models visitors look at. It usually does not capture which alternatives they ruled out in the chat beforehand. Anyone who studies only the visit to the shop starts too late for some purchase decisions.

A recommendation is not yet a purchase

Writing on marktforschung.de, Julia Eymann describes how regular AI users build such assistants into their product search. The qualitative GIM trend study offers insight into the decision paths of this group. It does not tell us how common these paths are among all buyers.^[1]

The KPMG and EHI Consumer Snapshot from October 2026 shows why it is worth distinguishing between individual steps. When buying a smartphone, 72 per cent of respondents would use an AI agent to search for information, but only 45 per cent for payment. These figures reflect willingness, not observed purchases.^[2]

In Quirk’s, Jon Arthurs of Toluna suggests testing the visibility of brands in AI answers with queries based on real needs:

“These should use prompt entry points derived from real consumer needs, not generic prompts”^[3]

This is a recommendation from practice. Such a test shows what an assistant outputs under certain conditions. Whether people follow the advice has to be examined separately.

How to examine the influence of AI

Start with the purchase occasion. In a short shopping diary, participants facing an upcoming purchase record what they do first, when they use an assistant and which suggestions they adopt or reject. The study should also include people who search differently. Otherwise, recruitment alone makes the AI channel more important than it needs to be for the research question.

Accompany a specific search. An observed product search makes it possible to trace which criteria people enter into the chat, what they double-check on the retailer’s website and when they switch to another source. With the pram, it might turn out that the model is convincing but its folded dimensions are missing. That points to a different measure than more visibility in AI answers.

Follow up after the decision. In the interview, we clarify what tipped the balance: the recommendation, a familiar brand, the price or perhaps the advice of another person. A later conversation also helps to understand cases in which a recommendation was convincing at first but did not lead to a purchase after all.

What brands can take from this

For clients, this results in different tasks. If the product hardly appears for relevant queries, it is worth looking at its discoverability. If it is recommended but described incorrectly, reliable and consistent product information is a starting point. If the recommendation is correct and the search still breaks off in the shop, a user test there can help.

At WMM, we would align the research design with the decision that comes next. A small number of well-documented purchase journeys can show where a closer look is worthwhile. How often a pattern occurs, or whether a change increases sales, requires an appropriate follow-up measurement. This turns the general question about AI into a concrete research task.