Impression Frequency Analysis

Problem Statement

Working with large-scale lead generation clients and spending hundreds of thousands of daily budgets on various in-house marketing channels and affiliate networks, we have seen how easy it is to get lost in remarketing and retargeting initiatives. Typically, all channel managers would have their own retargeting (pixel-based) and remarketing (emails from CRM systems) initiatives/campaigns within their channel efforts. It’s also not rare to have upper funnel retargeting-oriented channels of their own (for example, Criteo). 

One of the marketing analytics projects handled by our team was the development of an MVP dashboard to track the impression frequency analytics for a display programmatic channel.

The project’s goal was to identify the maximum limit of showing an ad to the same audience during a given period, not exhausting the audience while ensuring optimal brand visibility for conversion.

Planning

One of the main benefits of MediaMath, the programmatic channel chosen for our initiatives, was the impression-level log data, enabling tracking of each of the conversions previously exposed to MediaMath. For example, what exactly were those exposures; were they just impressions, did they yield clicks/conversions, what was the campaign, audience, device, publisher, and many more metrics. 

The available data sources for this particular analytics project were:

• Impression-level log data from programmatic channel
• User ID graph
• CRM database of leads and converted clients

Having this data available together with the final conversion data (CRM database), we have planned to map the impressions by users and any available user-level data to final conversions to understand the ideal frequency and duration of retargeting. 

Execution

Step 1 

User-level impression frequency analysis was conducted based on the impression-level log data. Merging all the impressions of a single user into one data point, we were able to identify what was the frequency and duration of a single user, getting impressions from the same programmatic channel. 

Step 2

The user IDs were then mapped to the actual conversions and leads’ database to identify which percentage of the overall population exposed to MediaMath ads had become a lead at the end of the day. The data allowed us to understand the impression of conversion rates for the channel overall and calculate the conversion metrics by frequency and duration of the retargeting initiatives.

Step 3 

A final touch to the analysis was putting the whole data together and building a comprehensive dashboard that would display the main metrics - the quantitative and relative volumes of impressions by frequency and duration of retargeting and conversion efficiency metrics for the same dimensions. 

Surprise Conclusion

Though we were hoping to see meaningful exposure exhaustion rates from the data, we were able to identify something more significant for the business: the majority of the users were actually served only one impression from this programmatic channel, meaning that the retargeting initiatives hardly served the initial purpose of exposing the user to the upper funnel communication messaging, but getting a random impression from the channel, which would later be claimed as part of the multi-touch attribution for the company. 

Suppose we ignore the finding above and try to get the answer to the original question raised by the company marketing management. In that case, we could conclude that there were big spikes in conversion efficiency around the days 19, 24, and 30 after the first seen impression for the user as total days of retargeting the same person until they convert and 5-10 was the most optimal bucket of a number of impressions by the time that the user converts.

After this analysis, the company decided to switch its attribution logic to account for only the conversions exposed to the display programmatic ads at least 2 times before the conversion. Shortly after, the channel was closed altogether because of a lack of evidence of incremental conversion efficiency.