Why geo-contextual targeting could be the answer to cookieless advertising

Joshikaa R, Data scientist I, MiQ

The ad-tech industry has traditionally relied on user data for targeting. But with recent changes, including Apple limiting user tracking and Google announcing the deprecation of third-party cookies, the search is on for privacy-compliant targeting methods.

At MiQ, we needed to figure out how to best identify and reach relevant customers without using cookies or targeting users directly. It was important to also keep privacy and our target audience in mind. Our approach had to be data-driven, scalable, advertiser agnostic and at the same time, cost-effective.

Let’s explain two types of cookieless targeting before exploring one part of our cookieless solution in more detail:

Contextual targeting:

• is the placement of ads based on context using features like postcode, site domain, keywords, browser, OS, etc.
• gained huge popularity post-GDPR (according to recent research by exchangewire[1], 52% of UK and US marketers plan to up their investment in contextual targeting over the next two years)

Geo/location-based targeting:

• is the method of delivering ads to customers based on their geo-location
• can help advertisers reach customers with a message customized to their location

Why isn’t one targeting method enough?

Location information allows advertisers to reach customers with a message customized based on their geography. Contextual information is about the internet sites users visit along with browsing patterns and content they like. Since geo-postal codes are the most granular level of location information, you could consider targeting all users in that location. But that can lead to budget overspend because it targets more users than intended.

So, we looked at our target audience based on website viewing patterns. Using both of these strategies allows us to display the right ads based on the users’ online activities within a specific geography, giving a higher accuracy for targeting. We named this two-pronged approach geo-contextual targeting.

We’ve got a new method

The key is in exploring and identifying the data required. The data feed contains log-level details of auctions that MiQ managed to win in RTB (real-time bidding). For this example, we considered the following geo and contextual features:

• Geo postal code — Zip code where the ad gets served
• Site domain — Website where the ad inventory is available
• Device type — Type of device where the ad gets served. Six devices considered for the experiment include, desktops, mobile phones, tablets, TV, game consoles and set-top boxes

Keep in mind that some websites generate more traction from certain devices. For example, gaming-related websites typically have more audience visits through mobile phones than through desktops. Device type should also be taken into consideration along with zip codes and websites.

After the data had been finalized, the next step was to cluster and rank zip codes based on their importance.

• We formulated exposure scores for each zip code, device type combination based on the count of potential users falling under them.

Where i = user count and fc = unique feature combination

• The resulting score was used to discretize the feature combination into equal-sized buckets using a quantile-based discretization function. The number of buckets is a hyperparameter that can be optimised based on business requirements.

Site domains were categorized based on their content. For example, foxnews.com came under the “News” category. This gives an additional advantage to consider the type of content our audiences are into.

• We found the top site categories for each device type and calculated their support score using FP- growth, a pattern mining algorithm.

Sample output for device type — Desktop:

Now that we have the exposure score and top site domains in place, they are combined to produce the final data containing zip code, device type, and class along with the list of top site domains under each device type and category. The count of users under each feature combination is normalised to get the confidence score of that feature combination.

The ad will then be served to users falling under any of the feature combinations based on confidence scores.

This approach was tested in a live campaign, resulting in 67% lower CPA (cost per acquisition), 110% higher CVR (conversion rate) and 31% lower CPM (cost per thousand impressions) when compared to the baseline strategy.

What does it all mean?

With most of the countries around the globe moving towards a privacy secure internet and data localization, the programmatic advertising industry must build new solutions. This solution uses pattern mining but the next step will be to apply new machine learning techniques to make the targeting more precise and tailored to audiences.

References

1. The future of contextual targeting

Joshikaa is a data scientist at MiQ, having first joined the Bangalore office as an intern in December 2020. A self-described potterhead, she’s also keen on astrophysics and enjoys watching space documentaries in her free time.