Can AI chatbots make your holiday shopping easier?

However, Rajiv Mehta, vice president of search and conversational shopping at Amazon, said the company is now able to offer more helpful recommendations by programming Rufus to ask clarifying or follow-up questions. Customers also use Rufus to search for offers, some of which are personalized, Mehta said.

Of course, chatbots are prone to hallucinations, so Rufus and most similar tools can get it wrong.

Juozas Kaziukenas, founder of e-commerce intelligence company Marketplace Pulse, wrote in a November blog post that his company tested Rufus by soliciting gaming TV recommendations. The chatbot’s response included products that were not televisions. When asked about the cheapest options, Rufus responded with suggestions that were not the cheapest, Kaziukenas said.

An Associated Press reporter recently asked Rufus to provide some gift recommendations for a brother. The chatbot quickly spit out a few ideas for “thoughtful gifts,” ranging from a T-shirt and a charm keychain to a bolder suggestion: a multi-function knife engraved with “BEST BROTHER EVER.”

After a five-minute written conversation, Rufus made more bespoke suggestions – a few Barcelona football shirts sold by third-party sellers. However, it could not be said which seller offered the lowest price. In another search, when Rufus was asked to compare prices for a popular skin serum, he showed the product’s pre-discounted price instead of the current price.

“Rufus is constantly learning,” Amazon’s Mehta said in an interview.

Shop AI, a chatbot that Canadian e-commerce company Shopify launched last year, can also help shoppers discover new products by asking its own questions, such as asking for details about the intended gift recipient or features that the buyer wants to avoid. However, Shop AI has difficulty recommending specific products or identifying the cheapest item in a product category.

The limitations show that the technology is still in its infancy and has a long way to go before it becomes as useful as retailers – and many shoppers – want it to be.

To truly transform the shopping experience, shopping assistants must be “deeply personalized” and able to autonomously remember a customer’s order history, product preferences and purchasing habits, consulting giant McKinsey & Company said in an August report. says the McKinsey report.

Amazon has determined that Rufus’ responses are based on information contained in product listings, community questions and answers, and customer reviews, including fake reviews used to increase or decrease the sales of products on its marketplace .

The large language model that powers the chatbot was also trained on the company’s entire catalog and some public information on the Internet, Trishul Chilimbi, an Amazon vice president who oversees AI research, wrote in October in the electrical engineering magazine IEEE Spectrum.

According to Nicole Greene, an analyst at Gartner, it is unclear how Amazon and other companies weight different training components – such as reviews – in their recommendations or how exactly the shopping assistants come up with them.

Perplexity AI’s new shopping feature allows users to enter searches like “best phone case” and get answers from multiple sources, including Amazon and other retailers like Best Buy. Perplexity also invited retailers to share data about their products, saying those who do so would have a greater chance of having their items recommended to shoppers.

But Perplexity CEO Aravind Srinivas suggested in a recent interview with Fortune magazine that he didn’t know how the new shopping feature recommended products to customers. But in an interview with the AP, Chief Business Officer Dmitry Shevelenko rejected that characterization, saying Srinivas’ comment was “probably taken out of context.”

The context, he said, is that with generative AI technology, “you can’t know in advance exactly what the output will be just because you know what the inputs are” from the training materials.

Shevelenko said retailers and brands need to know that their products cannot be recommended in Perplexity’s search engine because they incorporate “keywords” into their websites or use other techniques to better show up in search results

“The way you show up in an answer is to have a better product and better features,” he said.