Marketing, Meet AI

Artificial Intelligence is often discussed as a future technology. The reality is that AI is here, and is deeply embedded in digital marketing. Existing ad networks like Google and Facebook are betting heavily on AI as growth drivers (think Google SmartBidding or Facebook lookalike audiences), and new networks have cropped up offering custom machine learning algorithms to help zero in on the perfect audience for any company.

Although it has been a growing presence in digital marketing for years, AI is relatively new to digital marketers from a direct management perspective. Networks, old and new, are offering a cure-all for finding your target audience, and getting them to convert at your target cost-of-acquisition, but we’ve found that it’s not the magic wand the network sales reps make it out to be. In this article we identify four problem scenarios created by the use of AI and potential solutions.

Problem Scenarios:

The Black Box

When networks talk about custom, proprietary algorithms, they may not just be concerned about their intellectual property. By nature, machine learning algorithms are more organic than synthetic, and there is often no way to describe the decisions the algorithms are making. We can speculate about the specific types of machine learning being used by advertising networks, but features such as lookalike audiences (now offered by most big programmatic, social, and display networks), can be created using a neural network, an algorithm that actually learns as it gets more data to make better predictions. These decisions are often complex, and likely cannot be distilled down to the tidy audience descriptions that marketers are used to, such as “ males, age 35-44, married, no children, associate’s degree”. Not knowing all of the variables that are in use, and how they are weighted and applied in the algorithm means that marketers have less control or understanding of the causes of failure or success.

Incremental Value

Is this adding value? How much? These two questions are vitally important to any digital marketing program, but can be trickier to measure with AI campaigns. Programmatic campaigns are designed to get ads in front of a relevant audience, and particularly individuals who are showing indications of being interested in converting. We have found evidence that the algorithms may be too good in some cases. In these instances, we can see the network in question reporting huge numbers of channel conversions (seriously, ridiculously huge numbers of conversions), but no corresponding increase in total conversions. This poses a problem that is not prevalent in any other type of digital marketing (with perhaps the exception of remarketing): are we paying for conversions that we would have garnered anyway? Are the algorithms so good at identifying people who are likely to convert, that they’re serving ads to people who would already be converting?

Metrics in Isolation/Top of Funnel Bias

We have had the same conversation with at least a dozen different clients over the past few years where we say “If that’s your priority, we can impact {some metric} in isolation, but we don’t recommend it.” Once a specific KPI has become the topic of conversation, it can be hard to see the forest through the trees – or the profit through the interim metrics (see our blog post on the dangers of Local Optima here). Networks are now offering products that can optimize to a target for you. You can maximize your conversions, set a target cost-of-acquisition, or just let ‘er rip and set everything to enhanced cpc. Just like with any other campaign, optimizing to a single metric in isolation is probably not going to get you where you want to go. For example, a 50% decrease in cost-of-acquisition sounds great, but can actually be terrible if we decreased overall acquisition numbers by 50% to get there (think: you saved 75% of your ad dollars, but your net profit fell by 50%). The algorithms in use by these networks are focusing on specific top-of-funnel metrics, in isolation. This means that your bottom line isn’t safe if you take a set-it-and-forget-it approach to your AI testing – at least not until these networks figure out how to incorporate down-funnel value into their optimizations.

Limited Data

In instances where you have a very limited testing budget, AI may be problematic. The issue we’ve run into is that the machine learning algorithms need plenty of data in order to actually learn. If you’re trying to run a smart shopping test with a $500 all-in budget, you will almost definitely see worse results than in your standard shopping campaign, since the campaigns are not designed to perform well initially, they are designed to gather as much data as possible. Many networks have actual thresholds for the amount of data you need in order to even launch a test, but those thresholds don’t necessarily indicate that you will be achieving optimal volume for the campaigns to properly optimize.

The Solutions:

Methodical Testing

In instances where data is scarce, you will need to employ some patience in watching spend slowly trickle through a campaign that doesn’t seem to be doing much of anything for days and weeks on end. Be prepared to battle with the temptation to pause your tests until they have left the learning phase and start showing you the kind of performance you can expect in the future. Some networks are nice enough to provide a “learning phase” status update, so you have a sense of how long you have to wait.

If you’re not so sure that there is value being added from your AI campaigns, set up a split test or pulse test, double-check that your data is squeaky clean (and includes all the information you will need to make a good decision), and let it run. The clean and complete data will also help you assess the down-funnel impact of the AI campaigns, so you can adjust based on financial outcomes, and not just the top-of-funnel results the ad networks will report to you.

Innovate

It can feel like its Google’s world, and we’re just living in it. The reality is that each of the challenges presented by the transition to AI bidding, targeting, and optimization, are also opportunities. We continue to identify unanticipated side effects of automated optimization in networks like Facebook and Google, and turn them into opportunities to use algorithm biases to our advantage and help our clients reach their goals. It’s easy to imagine digital marketing as an increasingly automated field. In actuality, the shift to more AI in digital marketing requires a savvier and steadier hand on the wheel than ever before.

Learn & Iterate

One of the exciting benefits of the ad networks processing so much data is that they can share some insights with us. Look for opportunities to learn more about the audience that’s engaging with your ads and site by mining the data available from the ad networks and the data you collect yourself. Then, use this information to prioritize targeted landing pages, update copy, and adjust your targeting.

AI is here to stay and will only become deeper ingrained in digital execution. While the use of AI requires continued thoughtful testing and potential revision of best practices from account structuring to “targeting” to budgeting, AI does present massive opportunity for profit-driven campaigns.