Navigating the post-cookie era

In recent years, user-level identification used to facilitate tracking and targeting has faced increased scrutiny from legislators and technology giants alike. This has generated renewed interest in contextual targeting, which had previously lost prominence to audience targeting with the advent of programmatic and real-time advertising. The renaissance of contextual targeting is seeing a host of new approaches aiming to go beyond just the URL as the basis for targeting and to a multitude of attributes of the bid opportunity.

Understanding multidimensional targeting

A programmatic bid opportunity is rich with various attributes such as user ID, URL, device type, geolocation, time of day, day of week, ad format, and app ID to name a few. In cases where the User ID is missing, which is becoming more common, these additional attributes can help make the most informed decision as to the best use of each bid opportunity.

Approaches employing machine learning are used to create models for specific marketing goals. These goals could be reaching a particular audience when user IDs aren’t available or driving specific actions like conversions. To train these supervised machine learning models, we need a 'seed' – an initial set of data such as a group of users who've converted, or certain page view patterns that lead to conversions.

We have traditionally leveraged cookies to seed the algorithms, but how will this work in a post-cookie world? Luckily, cookies aren’t the only form of user ID available, and its upcoming deprecation has led to the creation of alternate IDs such as Unified ID 2.0, ID5, RampID and Panorama ID. This enables the identification of behavioral patterns that distinguish converters from the rest of the population. Therefore, even if a user ID isn’t available, the model can identify the combination of attributes that commonly characterize converting users and increase the probability of reaching them. In other words: We are taking a probabilistic approach to user identification in the absence of a ‘hard’ user ID.

For instance, the algorithm may observe that successful conversions often come from users accessing certain websites and apps on specific devices from set locations, mostly on weekdays, between 8-9 am and 7-9 pm. Spotting these combined attributes on a bid opportunity can significantly boost our chances of reaching the right audience from a 1% probability to a promising 25%. Sure, it's not as solid as user ID-based targeting, but it's a big step up when user IDs are missing. The strength of this approach is its flexibility: it's not relegated to any one attribute but works with the information that is available to get the best results.

If we don't have any trackable users to seed with, we can shift our focus to page view patterns leading up to conversions. The algorithm can be trained to learn from these sequences, which often start outside the advertiser's site and end in a conversion. Leveraging session or 1st-party cookies or other non-third-party cookie IDs, we can pick up trends. For instance, the algorithm may find that conversions often start from a 300X250 pixel video banner on certain websites, accessed via a specific device from particular locations, typically on weekends between 9am and 12pm. Spotting these attributes in a bid opportunity could bump the likelihood of conversion from a scant 0.01% to a more hopeful 0.05%. Again, we're not tied to any one attribute, we're all about making the best prediction based on whatever data we have on hand.

While these two cases are presented separately for clarity, they can actually be combined into a single, comprehensive model to optimize for the right user and the right moments to boost chances of achieving marketing goals.

Challenges and solutions in multidimensional targeting

Multidimensional targeting requires a comprehensive integration between the targeting and the activation platform providers. For a simple contextual integration, the activation platform shares the URL of the bid opportunity with the targeting provider, which then provides relevant contextual targets. This information can be stored and reused for future impressions from the same URL within a certain timeframe, say 4 hours. However, multidimensional targeting is more complex, as it needs the platform to provide multiple attributes for each bid opportunity, which doesn't allow for caching. So far, two methods have been developed to tackle this challenge.

Bring Your Own Algorithm (BYOA) is a method that's been favored by DSPs in recent years. It lets the targeting provider use their own algorithm in the DSP's infrastructure, using a subset of the bid opportunity attributes. However, this approach is limited by attributes availability and hardware, which puts hard limits on the complexity of the algorithm.

True real-time integrations pioneered by IPONWEB's MediaGrid and Xandr's Real-Time Data Partner integration, is another solution. It requires the targeting provider to respond to a bid request within 10 milliseconds, as the response becomes part of the bid opportunity before it's sent to DSPs for bidding. This method comes at a significant cost, as the targeting provider must establish a hardware presence near or in each activation partner's data centers to meet these extreme response time requirements. The advantage is that the provider is not dependent on the hardware resources offered by the activation platform, which gives them full control over algorithm complexity.
‍

Our approach to multidimensional targeting

At Fyllo, we have made two foundational investments in multidimensional targeting. The first was to create a global network of data centers near the main activation platform hubs, enabling us to fulfill the response time requirements of current and future multidimensional targeting demands.
Our second investment is a feature known as Contextual Audience Extension. It uses a seed audience (composed from e.g., brand first-party data, panel data, purchase data, and so on) to train a supervised machine learning model. This model predicts the overrepresentation of the seed audience on a specific page, based on its semantic characteristics. Essentially, it creates a heat map of the entire addressable internet, with pages featuring a high overrepresentation of your audience shown as ‘hot', and those with low overrepresentation shown as 'cold'. You can now choose how ‘hot’ a page needs to be in order to be in target for your campaign.

Contextual Audience Extension serves as a vital first step towards multidimensional targeting, with applicability across all activation platforms currently supporting contextual targeting.

As more of our programmatic partners embrace true multidimensional targeting, our algorithms will gain enhanced prediction capabilities and resilience by reducing reliance on single attributes.

Join us on this journey as we endeavor to create net new forms of contextual targeting, taking a much more holistic, multidimensional approach.

Contact us today to learn more!