The rise of AI tools in the advertising industry has brought with it a new set of expectations. Marketers demand to see automation, speed and efficiency right off the bat, but as they adopt new tools, they also expect ROI.
As in many cases with marketing, if the ROI isn’t immediately apparent, marketers are weary to invest more. In this age of rapid adoption, an unsuccessful campaign that leverages generative AI tools can sour a brand on all AI tools. The brand may even blame the AI itself for the poor outcome.
This is obviously a hasty decision, and it’s far too soon to completely right off agentic AI tools. Advertisers that feel they got short shrift from a campaign that leveraged AI need to look under the hood a bit to understand what may have impacted performance. Small mistakes – call them modelling malpractice – can undermine the success of an entire campaign. Here are three areas advertisers can examine to make sure they maximize their return on AI.
The models lack sufficient records
For all that AI offers, it is not a replacement for the other critical elements that go into successful model production. The point of audience models is to find correlations and points of connection between sets of consumers, then use those insights to group individuals into a targetable segment of likely customers. Models deliver far greater fidelity and performance when they are built off of large data sets.
It’s entirely possible to build an audience model on top of only 500 records. But the number of connections and the granularity you get with 500 records is far less than what’s possible with 20,000 or 200,000 customer records.
If a model is built on too few audience profiles, then it may not deliver the kind of performance that the advertiser is expecting. In fact, it may miss swaths of the consumer population all together because they weren’t represented in the initial small data set.
This is much the same challenge that panel-based measurement firms like Nielsen are running into lately. The consumer landscape is vast and varied, which makes it full of pockets of opportunity for advertisers. More data opens the door to greater insights, which in turn leads to better targeting.
The data source is suspect
Let’s stay with the theme that AI outputs are only as good as what you put into them. The number of records used to build a model matters, and so does the origin of those records and the recency. So yes, models generally perform better when they are built on top of larger seed data sets, but those need to be fresh, accurate profiles.
Consider a model built on 50,000 consumer profiles that contain purchase data. Sounds appealing, right? What if I told you the purchase data was from 2020? That changes things dramatically. No AI engine in the world is going to build a successful model if it’s fed outdated data . In this case, outdated is synonymous with inaccurate.
Confirmation bias.
Seasoned marketers know that the size and accuracy of an initial dataset are important considerations when building audience models. But confirmation bias is something of the silent enemy in audience construction and deployment.
For many, the goal of a custom-built audience model is to surface insights from a first-party data set, use those insights to build a new segment, and then target that segment. This can take a lot of paths, including targeting lapsed customers, trying to upsell existing customers, or identifying new customers.
This latter use case is where I want to drill down. Most of the time, marketers start out with an image of what their new customers should look like. It’s very likely that they look like the existing customer base. So, as a model is being constructed, the marketer and/or data scientist will feel inclined to make sure that the result matches their preconceived notions.
This isn’t necessarily a bad thing, but it’s also a huge missed opportunity to use AI tools to their fullest potential. Beyond speed and efficiency, AI tools can surface connections and insights that run counter to what humans might think about their target audience. Because machine learning is looking at the data without bias, it may find net new audiences that don’t look like current customers.
Simply looking for audience models that confirm what you already know isn’t going to result in poor performance. But using models that go outside the expectations has a strong chance of widening a brand’s customer base and increasing the ROI, exposing a product or offering to a completely new audience.
That is likely the true value of using AI to build models and find audiences across media.
