How Marketers Can Prepare for Unintended Consequences of AI

Goals need to be specific and practices should be honed to achieve them

AI and machine learning can positively impact marketing, but there are also aspects that marketers need to be prepared for. Getty Images
Headshot of Brian Stempeck

Today, a great deal of attention goes into applying artificial intelligence and machine learning to advertising. And the value that AI can bring to marketing is unquestionable, from creating thousands of variations of creative formats to optimizing campaigns across channels and screens.

But it’s worth pausing for a moment to consider what might happen when extremely powerful technology is pointed at certain marketing goals. What unintended consequences might be in store, and how can marketers set themselves up for success? What questions should marketers consider to ensure that human and machine each have input on an ad campaign?

It’s time to ditch last-touch attribution

Last touch attribution, where the last ad seen receives credit for an online sale, is still the de facto standard in our industry. As machine learning drives more campaign optimization, that has real ramifications for how and when ads are shown.

As machine learning becomes a bigger driving force in every marketing plan, marketers need to get more precise about the goals they’re putting in place.

One subset of AI called reinforcement learning “rewards” the machine for hitting certain goals. Last-touch attribution, which is one such goal used to train machines, can be problematic, as complex algorithms trained to master last-touch attribution don’t necessarily lead to good marketing. Perfectly executed, that algorithm would serve an ad to a consumer a fraction of a second before that consumer buys the product in the ad. Doing that would deliver outstanding performance for the advertiser. But did the machine learning drive results? Or did it instead perfect the art of taking credit for good marketing without actually doing it?

Consider the consumer who has seen three months of branding ads for the new Alfa Romeo on connected TV. When that consumer visits the dealership online, the last seen Alfa Romeo banner ad gets all the credit, “rewarding” the machine for the wrong outcome.

Media quality is one goal, but not the only goal

Advertisers also commonly build goals around the quality of the advertising inventory itself. Was the ad viewable? Did the viewer watch the entire video ad? While these goals solve real challenges and clean up the quality of digital inventory, they shouldn’t be the only goals. Rather, agencies and marketers should view media quality as table stakes, not the end-all-be-all.

Using AI to maximize media quality alone can have unintended consequences as well. For example, in-app interstitial ads are highly viewable. But without the right guidelines, an algorithm acting alone could push the majority of campaign ad spend to in-app. That would check the box on viewability while potentially missing the mark strategically on what the campaign was really after.

Combining goals is a better answer. Clients who use both viewability and on-target CPM to judge campaign performance, for example, ultimately get confirmation that they hit the right age/gender target with viewable inventory.

Use sales goals whenever possible 

One of the biggest opportunities for marketers is training AI to optimize business results instead of proxies for business results.

Some clients use multi-touch attribution to credit different touch points along the conversion path and pass those scores back to train the machine. Multi-touch, granular attribution may not always be the answer, though, as it can be challenging, particularly with some walled gardens making it even harder to stitch the pieces together.

We’re seeing other advertisers tie their campaigns to sales data: Whether automotive sales measured by Oracle, grocery store purchases from Nielsen Catalina, a retail store visit counted by Placed or a lead conversion inside Salesforce. And in many cases, sales lift with a test versus control group is the source of truth.

Some advertisers also bring their own sales data to the table without needing user-level grain. Rather than seeking a perfectly described attribution path on every individual consumer, they look directionally. Take a retail chain with 300 locations. With sales data split out by store, zip code and product category, it’s not a big leap to train an ad buying platform to optimize those goals instead of clicks or online conversions.

By passing its sales data to a DSP and trying to drive sales by store, the retail chain provides the perfect opportunity for the media buyer and AI to combine powers. The role of the media buyer then shifts to developing and testing hypotheses. What creative messages drive sales in the northeast versus southeast? How far are customers willing to drive to the store, and how does that impact the campaign’s geo-targeting? What product categories are resonating in different seasons or regions?

As machine learning becomes a bigger driving force in every marketing plan, marketers need to get more precise about the goals they’re putting in place. Training AI on actual business outcomes, or at least better proxies, is a better long-term approach.


Brian Stempeck is chief client officer at The Trade Desk.
{"taxonomy":"","sortby":"","label":"","shouldShow":""}