While it would appear at cursory glance that human marketers are being replaced or transformed into designers as a result of algorithmic efficiency, rest assured that there is still plenty for the human marketer to do. Just as Microsoft Excel did not replace, but rather reinforced the work of data analysts, we human marketers can now benefit by leveraging a new technology that brings huge improvements to ad targeting, so that we can focus on strategic thinking and activities that drive higher value.
The age of algorithmic marketing enables marketers to shift their focus from surface-level traits (such as people who like Airbnb or search for “flights to New York”), to behavior-level targets (such as people who are likely to be interested in signing up for a travel app based on a multitude of correlated signals).
Here are a couple core examples of how marketers can take advantage of this new era of mobile ad targeting.
While already an option in the mobile marketer’s toolkit, algorithmic marketing increases the importance and opportunity to focus on targeting audiences for lookalikes (AKA actalikes, AKA audience hints) and re-marketing.
Lookalike audience targeting does not and should not be as simple as stopping after creating a 1% lookalike from a seed audience of purchasers. While this is perfectly fine as a starting point, it’s now more important than ever for marketers to deploy their full faculties towards building and exploring more advanced audiences -- and continuously testing the ROI that each group returns.
Some intermediate methodologies include:
- Building lookalikes from the highest-value users uploaded into a basic seed audience;
- Building combinations of different lookalikes targeting different levels in your app’s funnel (i.e. bidding to LTV for users likely to install/sign up/trial/purchase/repeat purchase);
- Using Facebook’s LTV seed audiences with a full range of users, from lowest to highest-value users.
An example of an advanced methodology includes leveraging an analytics tool like Amplitude or running a multiple linear regression model yourself to identify the correlation strength between independent variables (e.g. number of days retained) and the desired, dependent variable (i.e. LTV), and thus the behaviors to target in order to influence your bottom line the most.
For example, running such an analysis could uncover the fact that there is an inflection point that causes LTV to increase significantly when users complete 10 levels of a game, vs only 9 levels. This research could inform the targeting of a lookalike based on a seed audience of users who have completed 10 levels, and ultimately close the loop on the human-machine optimization cycle by uncovering a more effective way to target higher LTV users.
Facebook famously used this type of analysis to identify that users who reached a certain number of friends within a certain time period were drastically more likely to continue using the service than the average user, and altered their onboarding based on this fact.
Creating categorical or custom variables is in this Facebook vein of analytical audience targeting. This involves grouping users by a combination of behaviors, such as users who have consumed their full trial within 2 days or those who have completed their 10th level and played the game for over 60 minutes. This analysis is more difficult to manage as it involves proliferating and prioritizing metadata, but it can be a powerful way to uncover the types of key insights that power the top 1% of marketing campaigns.
Lastly on the subject of lookalike audiences, while determining the seed/member list is the most visible and sexy part of audience targeting, it’s important not to forget about the importance of determining the closeness of match, the price to pay per engagement, and the creative-audience interaction rate (creative optimization is certainly another strong opportunity in the age of algorithmic marketing).
Remarketing is somewhat more straightforward than lookalike exploration, with audiences managed by moving user IDs from more shallow-funnel remarketing lists to deeper-funnel remarketing list. It’s also important to adjust bidding through the funnel by subtracting the customer acquisition cost to get the user to the current level from the LTV of the next lowest level.
Because audience targeting is easier to adjust than event-based targeting (although it does require an SDK or raw user ID capture), it is a more ideal place to start; creating a seed audience takes little effort and has a minimal lasting impact on the ad account.
Comparatively, event targeting first requires events to be tracked using code in the app via an SDK, and second requires those events to be integrated into the ad platform to be used. If done using the ad platform SDK this second step is not necessary, but both methods require events to be mapped to an event category. Some events (e.g. purchases/sign ups) are easy to map, while for others (e.g. retention events or custom events) it can be more difficult to find an appropriate category mapping.
Event targeting is also integrated into the ad bidding and targeting systems, giving it a big impact on who in the audience targeted sees ads and powering more proactive reactions to user auctions (i.e. the dynamic between your own app’s performance, user engagement, and the competition).
The question sometimes arises as to why a downstream event-optimized campaign can produce significantly better ROI than an installs-optimized campaign, if both are targeting a high value user-seeded lookalike.
While you can bid higher for installs-optimized targeting and hope to extract more installs from valuable users within such a lookalike, the goal of an installs-optimized campaign is only to acquire installs, regardless of each user’s propensity to complete downstream events. A downstream event-optimized campaign leverages different machine learning algorithms to focus on the conversion probability of the user for the event specified, rather than the conversion probability of just an install, which is the fundamental difference in the way both campaign types operate.
For example, a Facebook value-optimized bidding setting selects and bids for users based on their predicted ROAS, while an installs-optimized bidding setting selects and bids for users based on their predicted installation conversion rate. If a user with a track record of making a lot of purchases does not happen to have a high click-through-rate or conversion rate from the ad, then the two campaign types would target this same user in the same audience differently.
Important to note is that using event targeting is not effective unless you have enough event volume to power the targeting algorithm’s learning phase. Ensure that you have 30-50 targeted events occurring per day for several weeks for best results. Marketers can address low target event volume by using a set of tiered campaigns optimizing for higher event completion rates at higher levels of the funnel and working to improve conversion rates at each progressive level through the funnel.
Lastly if you want to tap into event-optimized campaigns, be prepared to pay a premium for each install. CPIs often start at 2x as high as compared with an installs-optimized bidding and can run as high as 5-7x depending on the campaign type.
Exploring for valuable correlations between independent and dependent variables is what machine learning seeks to do and is a task to which human marketers can lend highly valuable contextual reasoning.
The age of algorithmic marketing does not mean the end of targeting or the human marketer. It marks the evolution of the human marketer.
The glint from the chrome Cadillac convertible era of the Mad Man marketer has officially disappeared below the horizon. In the old icon’s place gleams the eerie light of the Mad Scientist marketer; still human, but now part machine.