In the previous post, I listed four outdated truisms of media buying that no longer apply in the era of AI. Did AI make Media Buyers and Campaign Managers hopelessly obsolete, or can they add value in some way? Yes. I strongly believe that humans are still useful, after all–if they play to their strengths.
These are the useful things that Media Buyers and Campaign Managers can and must do:
1. Carefully define the right success metrics. Enable tracking of the desired events. Refine the goals over time.
AI is great at doing exactly what you task it to do, but it can’t assign a goal to itself. It has no “free will” of its own, nor does it understand what goals are important for your advertisers. The task of converting business objectives into clearly defined goals for the DSP ML algorithms thus falls on the human Campaign Managers. This task could be trickier than it seems.
For most verticals, including gaming, fintech, and e-commerce, determining and communicating the success metrics is particularly challenging. Typically, mobile app advertisers start by shooting for a certain maximum Cost per Install (CPI). However, installs are just the means to the end. What you’re really after are in-app purchases (good CPA and ROAS).
In this regard, the success metrics present significant complications:
- Not all installers turn into spenders.
- Even among spenders, LTV greatly varies. Not only the amount of the first purchase, but also the number of future purchasers, varies drastically.
Modern ML solutions, like the one my team and I have launched, can optimize multiple metrics at the same time. In the app environment, they start by driving installs (lowering the Cost Per Install [CPI]). However, as installers become purchasers, sophisticated multi-layer ML models prioritize the highest-spending installers. This is why it is critical to pass not just installs but all downstream revenue metrics into the ML system.
The new customers that your ad campaign attracts spend money within your advertisers’ accounts over time. Thus, you start with an estimate of the new customer average LTV, and then you refine that estimate. Over time, you should adjust the CPA goals you input into your ML system.
2. Produce a Diverse Set of Creatives
Predictive AI solutions are great at matching users with creatives. Gen AI can help you produce creatives quickly and at a low cost. However, even when assisted by Gen AI, deciding on the creative strategies, requirements, and ideas is where humans still excel.
If your product has several benefits, as most products do, emphasize some of the benefits in some creatives and other benefits in the others. Do the same thing with the other characteristics of creatives such as:
- Calls to action
- Psychological triggers
- Design & layout
- Images/videos
Don’t just test minor variations of the creative, like a slightly different shade of purple. Be bold.
Define a variety of creative sets for all formats and sizes. Do not assume that some formats are inherently better or worse than the others. I’ve seen media buyers in one company assume that “banners work better than video, because this is the results that they have seen.” The other company reached the opposite conclusion, so they would only run video and never the other formats. Both of these conclusions are examples of Hasty Generalization Bias. Programmatic markets are fairly efficient, so they price the different formats in proportion to their performance across all campaigns. Yes, the CTR and Conversion Rate (from impressions) for banner ads are much lower than for video, but they are also much cheaper.
Pro Tip: Map similar creatives into themes/groups/strategies, so your ML enables cross-learning/data sharing among similar Creatives within the same theme. Different creatives might appeal to different audiences and perform differently depending on the context. However, when you have a lot of creatives, the ML model might not learn enough about the performance of each due to the high granularity/cardinality.
Grouping similar creatives into themes and making the themes available as features to your predictive AI engine resolves this issue.
3. Instead of “Targeting”, Help your Data Science Team Come up with New Sources of Data for ML to Rely upon.
In the ancient era, “Targeting” was the approach used by Media Buyers. They would typically have ideas about which audiences would convert better and which would convert worse for a particular campaign. So, they would target the “better” audiences and avoid the audiences that are “worse.” As I have shown in the previous post, this is an outdated tactic. Except for obvious things like not advertising alcohol to people below legal drinking age or not showing ads made for one country in another country, you should not target.
Every user and every piece of inventory is valuable–at a price. Let your ML system pick that price. But you should help it by giving it the signals to learn from. A modern approach is to feed the audiences that you’d target or avoid to your ML system, so it determines how to bid appropriately given the combination of all available signals. You are the expert about what these signals might be. Help your Data Engineering, Data Science & ML Ops Teams to acquire and enable those valuable signals.
4. Don’t Overlook “Retargeting” Signals
Retargeting usually means showing the ads to:
- Users who clicked but haven’t installed the app.
- Users who installed but haven’t purchased.
- Users who purchased only once, but have not yet subscribed to automatic re-orders.
- Users who purchased one product but did not purchase related products.
- Users who installed your app but then became inactive.
Each action is a strong signal of users’ propensity for future purchases. The old practice would be to create separate retargeting campaigns. Indeed, those often perform very well because users have already confirmed their interest in the product. A more powerful approach, however, is to feed the above signals as features into your ML. This way, when the system sees these users, it will bid on them appropriately (most of the time, much higher) than on all other users.
Summary
Predictive AI is extraordinarily powerful. It has been outperforming humans on many tasks years before the first ChatGPT model went live. This includes, first and foremost, the task of campaign optimization. Predictive AI performs billions of calculations before determining which bid price, if any, would be appropriate for each user in a given context.
However, ML systems perform the tasks you give them using the predictive signals that you and your team enable. As sophisticated as the predictive AI systems are, they have no way to determine if they are serving a useful purpose. They also have no “awareness” of what source data might be useful in making predictions. As the media buying expert, you should be in charge of your Campaign Optimization AI by assigning it appropriate goals and feeding it with relevant data.
Being able to set the correct goals and frame the problem in a mathematical way that a Predictive AI system can understand becomes an ultra-valuable skill that will make you extraordinarily valuable in the age of AI.
