John is a Product Lead at Acast — this interview has been lightly edited for style and readability
John Burgess: Smart Recommendations is an AI powered search engine that allows advertisers to find the perfect podcast match in seconds - finding the right audience that is the most efficient for their ad spend.
Sam Sethi: So it’s AI powered, like everything these days. But, how are you using AI to get a better result?
JB: Yeah, AI gets thrown around a lot these days. When most people think about AI, hey’re thinking about Chat GPT, Grok, Claude - the typical LLMs.
What we’ve developed with smart recommendations is a way for you to use natural language, so you can come into the platform, describe the audience you’re looking for, describe the product that you’re wishing to promote, using natural language. So I could type: “My perfect audience is females, in London, interested in running”, and we take that and using LLMs, we can extract all of the nuances around that prompt. So, it might that you’re interested in certain running brands, certain apps that are related to running, all of those interests that runners may have and we extract all of that nuance using the LLM.
But then we take it further across all of the proprietary data that we have, both first-party from creators themselves, from third-party data that we have, and also all of the predictive demographic and audience data that we’ve built throughout our data sets. We understand a lot about our creators and we can map all of those different signals to find the perfect shows that you can promote against, so that you can buy and support those creators.
And then finally, after that, the most important thing is the why. Why is this show the best fit for me? And again, that’s where we’re using the LLMs with all of the data points that we’ve extracted and that reasoning, to narrate that reason as the why, that justification, and that really helps the advertiser feel confident that they are approaching the right shows for their audience.
SS: So Acast did something really smart about a year or so ago - contextual transcription, which was the idea of being able to look outside the immediacy of what the podcast was advertising at the title or description level. So you might have a podcast that talks about knitting, but it had a mention of basketball and therefore you could pull basketball ads. So that’s the basis of your data set, which is what you’re working against. Now you’re applying an AI layer to that data, is it 10% better, 20% better? What’s the return on this type of AI layer that you’ve applied?
JB: One of the problems that advertisers find, and planners for these advertisers find, is that it can take a lot of time to find the perfect podcast. There’s millions of podcasts out there. There’s 140,000 podcasts on Acast and finding that right one can be really difficult if you’re just relying on titles and descriptions and things like that. But when we’re going deeper and we’re really looking into the show and what is being talked about, that’s where we can find those perfect matches which might not be so easy to find.
We’ve been running this internally now for around six to eight weeks. It’s been used on over 200 campaign briefs during the internal test period, and anecdotally from our planners we see that we’ve gone from in the range of an hour to put together the recommendations to less than five minutes. So, 92% time-saving, which is huge. That allows those planners to really go and work on higher value campaigns. It allows them to do more creative endeavors that they may not have had the time for in the past. We’re really seeing it as a second brand, if you like, for the planners and the advertisers to really allow them to optimize their time.
SS: What AI tools are you using?
JB: We’re using OpenAI’s 4.0 mini at this point. We chose this because of the efficiency elements of it. There’s many versions and and over time we need to see how things play out. We’re really excited to be launching this week and over time we’re going to be reviewing how it’s being used, the types of prompts that people are putting in there, how long, how descriptive, how broad they are and finding that right model that most suits the need. We’re using a very efficient model for the use case at the moment and we’ll see over time if we need to make any changes on that to improve the results.
SS: Is this a free add-on to customers who are already using the Acast platform, or is this an additional function and feature?
JB: It’s included within the Acast ad platform.
SS: See you soon, John.
JB: Thank you, Sam.