An Experiment in Conversational Shopping

Internet search changed how we discover and buy products. As more of our attention shifts from keyword-based search to multi-turn ‘conversations’ with LLM’s – its possible there’s another step change coming. Shopping conversationally can be a personalised journey where a description of a person’s needs can progress through a range of potential answers and refinements leading finally – and only as necessary – to products.

It’s not clear yet how the e-commerce landscape will shift as companies compete to have their products “seen” by people using generative AI systems. There are open questions like whether LLM’s will have paid-for bias represented in their responses like search does – and how legislation might emerge to protect the public from AI-powered manipulation. Something that can be studied right now however is how conversational interfaces might change people progress along what marketers call the “purchase funnel”

Method
A group of 20 research participants was given a shopping challenge of finding a roughly $100 birthday gift online for a friend interested in computers. Participants were split into two groups. A control group was given a standard Google search page and asked to their preferred online shopping method to complete the task. The second group was given the same instruction with a single difference that they launch their search from Google Gemini. The participants’ screens were recorded throughout the task, capturing the steps they took as they searched for a suitable gift.


Results
Each discrete action a participant took during the product search process was captured and categorised. The graphs below captures the results. The main difference was that the group using LLM’s took far fewer actions related to seeking information about product reviews. They depended on the LLM as a source of validation of the value of the product.

The sites people used to source their information where also captured and categorised. (“Aggregator” is a site like Amazon or Bestbuy). The main difference observed here is that the group using Gemini didn’t seek very much information outside of Gemini despite having freedom to do so.