Case Study: Determining Optimal Ad Targeting To Drastically Reduce Cost Per Lead Using Facebook Audience Insights Data 

By Kian Xie, Digital Marketing & Analytics Consultant

Process and documentation are copyright Heal Your Marketing, 2018, all rights reserved. Specific keywords and user information are withheld for the purpose of client confidentiality.

The Process

We started with aggregate data sourced from FB Audience Insights, which included audience proportions related to gender, education level, audience behavior (clicks, likes, comments, shares on posts), income level, and purchases of subscription services (numbers on the last two metrics were sourced from third-party Cambridge Analytica data, and thus are no longer available through Facebook). Each row represents one audience segment, categorized by interest keyword. Each column represents a particular metric on which we would like our targeting audience to score high.

These numbers don't tell us too much on their own. We can try to order the list by one metric, such as desired income proportion, but that does not create any clarity regarding other important metrics, such as audience behavior. Instead, we want to view this data in a way that makes it easy to optimize for ALL desired qualifications.

Just knowing that one audience segment scored higher than another segment on a certain metric isn't enough to give us significant insight. We need to measure the differences in scores in a way that reflects the context of the whole combined audience. Thus, for each metric, we calculate the mean and "pseudo-deviation" - not quite a standard deviation, but a similar measure of variation that adequately serves the purposes of this test. The columns shaded in gray show part of the process of calculating these measures.

Once the mean and pseudo-deviation are calculated for each metric, we establish a color-coding system that will allow us to categorize and visually observe the distribution of scores in each metric:

Finally, we create an overall "Favorability Score" based on the value of each metric and its expected effect on lead quality, and order the rows by this overall score. Some patterns become apparent, such as more blue/green on top, and more yellow/orange/red in the bottom rows. The levels of visual consistency in the upper rows guide us intuitively as far as where to draw the "cutoff" line between favorable and unfavorable segments. We decided, based on visual observation of rows, to set the cutoff at a favorability score of 73.

Some Insights We Observed After Applying The Process:

  • Some metrics show more regularity and consistency in their distribution than others. For example, the metric "clicked" (average number of times each person clicked on an ad in the past 30 days) shows a large amount of variability that conflicts with other desirable metrics, and the distribution of "percent_gradschool" (proportion of users whose highest level of education is a graduate degree) is less variable and more consistent with the overall Favorability Score.
  • Some audience segments are consistently close to the mean, or slightly above the mean, in their scores across all metrics, while other audience segments are unusually high on some metrics and unusually low on others.
  • In one particular segment, post engagement (likes, clicks, shares, and comments) was extraordinarily high, while desired income level proportion was extraordinarily low. We later obtained insider information that the marketing department of the company most associated with this segment keyword was investing significantly in paid and organic social media strategy that focused solely on increasing brand recognition and social sharing, without paying any attention to conversions, sales, or any demographic audience quality. Due to inconsistency with other audience segments, and low overall favorability, we knew we would have to exclude this segment from targeting if we wanted to obtain consistent results.

The Result: 

Drastic, Statistically Significant Reduction In Cost Per Lead!

In conclusion, we have a ridiculously high level of confidence that this process is repeatable and scalable in its effectiveness to reduce advertising costs, provided the advertiser has access to data regarding audience income level proportions and buying habits.