Scraping podcast data for advertisers
The skills I demoed here can be learned through taking Data Science with Machine Learning bootcamp with NYC Data Science Academy.
GitHub repo
Introduction
Podcasts are one of the largest growing entertainment mediums, both in size and outreach. Advertisers are also reaping the rewards of this rapidly growing industry. The total data market estimate of podcast advertising revenue is projected to exceed $1 billion in 2021.1 Consumers have better brand recall and are more likely to purchased goods or services heard on a podcast than other forms of digital advertisements.2
However, unlike other digital ads, podcast ad costs are calculated per 1000 downloads.3 The ideal target audience is loyal enough to purchase whatever the content creators advertise, yet small in relative size to minimize advertising costs.
Therefore, the purpose of this project was to accomplish the following objectives:
- Collect data on the top podcasts in the United States.
- Determine which networks produce the most successful podcasts.
- Discover the most popular podcast genres.
- Create a custom metric that best explains describes how highly ranked a podcast is.
Data Collection
To collect podcast information and complete the project objectives, I scraped a third-party website that tracks chart rankings for Apple, Spotify, and Google. Scraping more than the first page requires users to log in with a free account, so I used Selenium to enter my information and change pages.
The following podcast information was scraped for each podcast:
rank, name, network, genres average stars, number of ratings
There were data quality issues due to scraping from a third-party site. For example, the podcast Office Ladies uploaded its most recent episode on February 17th, 2021. However, according to the website, they haven't posted an episode since July of 2020. The website updates certain podcast information more frequently than others, and in some cases does not include certain items such as genre, network, or ratings. Therefore, parts of the dataset may be unreliable. Collecting data directly from the streaming platform would yield more accurate and up-to-date information.
Data Cleaning
If genres were unavailable, I classified a podcast's genre as 'Unknown'. In addition, podcasts that did not have network information were marked as 'Unaffiliated'. Finally, I had to unpack podcasts that had more than one genre in order to analyze the performance of different genres in the top 100.
Comparing platforms
While Selenium provides additional flexibility that other methods of web scraping do not, the tradeoff is the additional time required for data collection. Collecting podcast data from multiple platforms (Spotify, Apple, etc.,) is computationally demanding and time-consuming. Instead, if we could assume there are no differences between listeners across platforms, we could simply collect data on a single platform.
First, I selected all the available podcasts on each platform chart. This third-party site compounds the top 200 for Spotify and top 250 for Apple. Of those podcasts, 98 had both an Apple and a Spotify rank. To test whether a podcast's Apple rank was statistically different from their Spotify rank, I created a bootstrapping function to randomly sample podcasts and compare rank differences.
The results revealed no differences between platforms. Therefore, using rank as our measure of a successful podcast, we can make generalizations about Apple users from Spotify data. While this may hold true now, platforms are starting to sign exclusive deals with podcast creators4. As this content becomes more popular yet unavailable to other platforms, differences in rank may eventually diverge.
Network Data results
For the remaining analyses, I examined the top 100 podcasts on Spotify. Seven networks have more than one podcast in the top 100, including Parcast Network (19) and NPR (10). Parcast Network has exclusive Spotify podcasts for each horoscope sign, explaining why they have so many podcasts within the top 100.
NPR has the most archived episodes, followed by Parcast Network and BBC World Service. NPR and BBC each have a podcast that are news updates uploaded every 2-4 hours. That sheer volume of daily output is the primary driver as to why these networks produce the most content.
Genre Data results
Society & Culture is the most popular podcast genre within the top 100, followed by Comedy, Health & Fitness, News, and True Crime. Similar to networks, the podcasts by NPR and BBC with an overwhelming number of episodes also belong to the News genre. Therefore, News and Daily News have the most episodes by genre, followed by Society & Culture and Comedy.
Custom metric results
As mentioned in the objectives, the cost is determined per 1000 downloads. However, downloads nor views are not available in this dataset. Hence, I had to find an alternative metric to identify successful podcasts for advertisers to target. The most highly correlated features to rank are average stars
and number of episodes produced
. 'NPR News Now' skews the distribution of the number of episodes, and when removed from the dataset, the relationship between the number of episodes and rank becomes negligible.
To create my metric of success, I made the following calculations:
success_metric = stars^4 / log(number of ratings)
The correlation between my success metric and rank was no different than the most highly correlated features that are not skewed by the number of episodes (including episode density). In future work, I plan on creating a regression model to determine how each feature influences Spotify rank. However, collecting data directly from each streaming platform will likely yield the most accurate and reliable information to determine which podcasts will maximize return on investment.
Conclusion
In this project, I identified which networks and genres most comprise the top 100 podcasts on Spotify in the United States. Also, I determined stars and the number of ratings are the most valuable metrics in determining rank. However, the limitations of the third-party site information made it difficult to make confident business decisions based on these findings.
If you have any additional questions or want to find more of my content, connect with me on LinkedIn.
Sources
- https://iab.com/wp-content/uploads/2019/06/Full-Year-2018-IAB-Podcast-Ad-Rev-Study_6.03.19_vFinal.pdf
- https://speakerdeck.com/brendanbrits/podcast-market-deep-dive-bdmi-jan-2020?slide=40
- https://www.chartable.com