Modeled vs. Unmodeled Attribution

Podsights uses a modeled approach to attribution. Due to the data available in Podcast media and the nature of mobile and corporate networks, Podsights cannot reliably run attribution with a standard attribution window on all impressions.

As an example, if Podsights observes an impression from a cellular network, and then sees a conversion on the same cellular network 7 days later, chances are, that is not the same person since cellular networks use the same IP address for many customers. Attributing that conversion to the impression would likely be a false positive, and therefore inflate performance metrics. Therefore, we won’t attribute that conversion to the exposure.

Instead, Podsights only runs a standard attribution window on Impressions originating from households, then assumes those impressions and performance are indicative of the full population of Impressions. With that assumption in mind, Podsights models results to provide a more holistic view of the impact of your podcast advertising campaign. 

Podsights captures two main pieces of data (IP address and user agent) from podcast downloads. Note that this is widely restrictive when compared to other digital marketing channels. There is not an identifier that ‘follows’ a device or browser, i.e. a cookie ID, device ID, hashed email, etc… Given the nature of IP addresses, a device is assigned an IP address based on where the device is physically and what connection it’s connected to, such as the router at home or the cellular carrier’s towers.   The main challenges for Podcast attribution is to reliably define the connection type of the listener and to extrapolate results from listeners that we can measure reliably to showcase the full impact of the podcast campaign. To solve these two challenges, Podsights implements the following solutions: 

1. Identifying the connection type of the listener
   1. Finding confirmed households: When we receive an IP address from an  impression, the Podsights process will attempt to match the IP address to  Tapad’s cross device graph to help define the IP connection type as a household. Once we figure out the distinct household from download, we are able to reliably see if that same household showed up on the advertiser’s website within the standard attribution window.
   2. Using an IP Address contextualization database: Podsights uses MaxMind to help define the connection type of an IP address. The database will tell us if the IP address is likely to be a household (i.e. cable/dsl/dial-up connection type) or a noisy connection (i.e. cell tower or commercial IP block). For the connection types that are likely a household, those IP addresses will be considered a household unless they exhibit behaviors of noisy IP addresses (e.g. more than 10 different devices showing up from the same IP address on the advertiser’s website).
2. Modeling: When Tapad is unable to match an IP address to a household, this is where modeling comes into play.
   1. Based on the total IP space which we capture through the downloads, we project a multiple of additional households that would have also shown up on the advertiser’s website had we been able to accurately define them as a distinct household.
   2. The multiplier will vary for each line-item as it is based on the amount of non-noisy IP data we can use for attribution and the amount of noisy IP data we have to throw away. 
   3. The multiplier is calculated by taking the number of total IP reach and dividing it by the number of attributable IP reach.  
      
      
      • As an example: Sally fills a cup with water, but spills a bunch of it on the floor. Sally uses the water in the cup to estimate how much she spilled.

Soo... When should I use modeled data vs. unmodeled data?

[Recommended]  Use Modeled Data when:

• You want to pull the dashboard data as is, following the Podsights modeling methodology

Please Note: Conversion exports will pull unmodeled data ONLY. Unmodeled data includes location and user agent details. The modeled data will not provide such information since the modeled data is derived from the multiplier.