Advertising is based around a simple to understand but hard to solve problem: what is the best way to reach the right audience at the right time?
How advertisers have gone about solving this problem has changed over the years.
In the ‘Mad Men’ era, it revolved around circulation auditing agencies, TV ratings agencies, other offline evaluation methods, and deals made over steak dinners.
In the 2000s and 2010s, as advertising moved online, tracking users via cookies was the predominant method of creating and targeting audiences.
Today, the advertising industry is experiencing a major disruption and change brought on by 2 major factors:
1. The rise of privacy regulations and commercial pressures that are causing cookies to go extinct.
2. The rise of large language models (like GPT-4) that are creating massive opportunities and tools that were inconceivable even 12 months ago.
At Pontiac, we’ve partnered with Michael Misiewicz at MContextual to help develop a state of the art contextual engine that solves a piece of the privacy-friendly advertising puzzle.
Michael and our CTO Erik Thorson partnered to write up an overview of our approach to solving this tricky issue of how to evolve advertising to be more privacy friendly, while still being scalable and effective.
The Problem
The current state of digital marketing is built on personalized ads, remarketing, and user segmentation. These core elements are made possible by the third-party cookie and user specific ids. In the evolving age of privacy compliance and data regulation, user tracking and the practices that enable it are no longer a viable option. This capability elimination leaves the marketing industry in an unusual predicament: clients expect informed, efficient targeting, yet the traditional tools are no longer available. Many in the industry are trying to find work-arounds such as first party shared cookies or device fingerprinting. Ultimately these methods are unlikely to succeed because they are still tied to identifying a single user and mapping their behavior: a privacy no-no.
Marketers are now asking:
“What else is out there and how the industry can approach this issue in a novel way? How do we leverage the data and tools that we have in the modern era to accurately target populations based on their characteristics and interests as accurately and efficiently as possible while still allowing for complete privacy compliance?”
If user lists, device graphs and third-party data are no longer available, the industry will require reliable tools and methodologies to efficiently and accurately find audiences and allocate media budgets. To do this, marketers must increasingly rely on statistics and probability gradients. We must evolve from “How can I target my audience precisely?” to “How can I maximize the probability that my ad will reach my intended consumer?” To do this, marketers must start leveraging tools that collect as much privacy compliant data from as many sources as possible and mine this data for significant signals that correlate to better performance and audience reach.
The question remains: “How do I find the right audience at the right time?” Cookies helped paint a comprehensive picture of who an online user was by tracking them from site to site and inferring their interests through content engagement, shopping behaviors, and brand affinities. These means are no longer available, but fortunately, there is more than one way to paint a picture. By using modern machine learning techniques across an increasingly large volume of data, advertisers can use proxy variables as a means of assessing audience interest and intent. The extinction of cookie based targeting means that the advertising industry is losing a potent tool, but a migration to contextual and advanced demographics-based targeting is actually a shifting of the means, rather than the ends.
The Opportunity
Whether a function of migration or adaptation, people of similar mindsets, demographics, and socioeconomics tend to gravitate together geographically. Six times over the last 100 years, researchers have conducted similar studies that demonstrate how people in certain regions of the United States share a vocabulary that is distinct from other parts of the country. These lexical sets carry further than simply the words they use; in many ways they define communities and culture of the people tied to them. Put people from Boston, Chicago, and Sacramento in a room and have them share what they call “athletic shoes” and mayhem will ensue.
Nielsen has built a massive and lucrative TV industry benchmark for measuring audiences using geographic groupings since the days of broadcast radio. The implementations and learnings from these first party studies and tools could prove to be powerful indicators of opportunity in the current marketing challenges. Could combining modern machine learning tools and geographic and contextual data sources unlock a whole that is greater than the sum of its parts?
Two of the most prolific and ubiquitous internet channels have fundamentally changed the advertising industry: Search and Social. The two leaders in these channels – Facebook and Google – have already demonstrated the viability of non-cookie-based advertising methods. In Facebook’s case, audience segmentation is based on social activity and engagement. Facebook uses social data to who you follow on Instagram as a proxy of a user’s intent, and therefore, membership in an audience, which they sell to advertisers. In Google’s case, AdWords on search can assemble audiences by inferring a user’s intent from what they literally type into a search box. These 2 companies are among the largest and most successful in the world.
Contextual and geographic targeting at an aggregate level are completely privacy compliant and can be accomplished without assigning a personally identifiable id of any kind.
Geographic Targeting
Marketers implementing geographic targeting have traditionally focused proximity within DMAs, counties, zip codes, or even lat/lon where a product or service is physically located. An example of this type of geotargeting is promoting a local business such as a car dealership to the surrounding population. The most likely potential customers are within a radius of that dealership which makes it practical for them to travel to the dealership.
A more advanced use of geotargeting is unlocked once we dive deeper into the nuances of geo targeting. There is far more to a location (and the population within) than just proximity to a business. For example, certain zip codes contain very heterogenous populations such as college campuses and areas surrounding military bases where the population can be extremely specific and different than their surrounding communities. ZIP codes can be identifiers for many demographic, socioeconomic, and cultural groups, enabling accurate audience reach for any businesses, not just local brick and mortar. For marketers, these geographic indicators offer unique targeting opportunities for nearly every advertising campaign.
Contextual Targeting
Contextual targeting and modeling has traditionally been based on the content that a user is reading. Media buyers target keywords which relate to the product or service that they are advertising. For example, a user is reading an article about planning a ski vacation and they receive an advertisement would be more likely to respond positively to an ad for jackets and outerwear rather than say cosmetics or pet food. The implicit assumption is that the content you are viewing has a relationship to the user’s state of mind and their interests.
With the advent of neural nets and large language models, contextual targeting and modeling has progressed beyond keywords to a more comprehensive understanding of the actual meaning of the content. For example, the word “squash” can refer to a vegetable, a verb, or a sport. The context in which that word is found is as important as the word itself. Modern models can make that differentiation and find pages that are truly relevant to the advertiser’s target content in a more “human”, comprehensive way.
Geography + Contextual (+much more)
Using geography and contextual on their own has been a successful strategy for many campaigns since the dawn of advertising; but even more potential for success lies in the overlap and correlation between them. The latest machine learning frameworks have become so pervasive and successful across our society because they offer the ability to ingest extremely large, diverse data sets and make sense of them. The insights that they gather are unique and intriguing to the point that the term “AI” has become universally recognized. If the data is properly prepared more data eventually leads to more predictive models which are able to recognize patterns and variables that are not observable humans. By building large, diverse feature sets that incorporate many geo specific data points and building models off of them advertisers can potentially reveal characteristics and correlations in populations that are more indicative of potential customers than 3rd party, cookie based data.
To that end, user interest segments could potentially be replaced by geo segments to achieve the marketing goal of hitting an interest specific population while providing complete privacy to the consumer.
The full whitepaper will be released soon, so check Pontiac Intelligence for the download link.