Amazon’s AI hosts are now selling you products: Here’s how

Amazon has quietly launched one of its stranger AI features to date: a short, podcast-style audio segment where two synthetic hosts discuss a product's reviews, specs, and merits and take questions from you as if it were a call-in show.

When the feature is available on a product page, shoppers can generate an audio segment voiced by two AI hosts who walk through the product's highlights, common customer complaints, and relevant context pulled from multiple sources.

According to Amazon, the content isn't limited to the product listing and user reviews it draws from other online sources as well, meaning third-party reviews and editorial coverage can feed into the discussion.

The technology runs on Amazon Bedrock, Amazon's cloud-based AI platform, described by the company as "several AI technologies working together." Shoppers can also submit their own questions, which the hosts will answer mid-episode in a call-in format.

Google's NotebookLM introduced a similar text-to-podcast tool at least two years ago, letting users upload PDFs or web pages and receive a conversational audio summary. That tool had a clear use case: turning dense, difficult reading into something easier to absorb.

Amazon's version applies the same format to commerce, which is a different proposition. The product pages it targets are often genuinely overwhelmingly stacked with specs, sizing charts, brand copy, and hundreds of user reviews. Whether audio helps cut through that or simply adds another layer is the real question.

Testing revealed some predictable limitations. When asked to compare the Bissell Little Green mini vacuum to the full-size model, one AI host acknowledged it lacked the specs to make the comparison and directed the listener to check the other product page a non-answer a basic search would have solved faster.

Amazon has not explained which products qualify, and attempts to generate podcasts for toilet paper, wet wipes, and certain adult product categories returned no results. The feature appears to be expanding gradually, with no clear criteria made public.