Tackling the cookieless era with MiQ’s Cookieless Domain Recommender
Manish Pathak, Team lead data science, MiQ, Divyaprabha M, Data scientist II, MiQ
Picture this: It’s 2023 and you are configuring a new campaign. However, the client wants you to reduce dependence on cookie data. What do you do?
The cookieless data question has been plaguing many traders and advertisers, especially since the big push on GDPR took place. Furthermore, Google’s announcement and subsequent delay of cookie deprecation have left many on edge.
Therefore, the writing is on the wall, cookies for the purpose of tracking user activity are on their way out. We at MiQ have invested in exploring ways to reduce the impact of this eventuality and innovated to stay ahead of the curve.
The Cookieless Domain Recommender leverages SimilarWeb’s dataset and DSP reports to come up with domains to target in an advertising campaign. This solution generates a list based on the geography and the advertiser’s website.
The picture now is a lot clearer!
What we set out to achieve?
Our goal with this recommender is to generate a list of domains that can be used in campaigns without any prior data. To accomplish this, we have utilized SimilarWeb APIs, which provide valuable engagement metrics and insights into website traffic. These metrics help advertisers tackle important industry challenges.
Some of the highlights of our solution include:
- DSP agnostic: Our Solution is compatible with any DSP
- Cookieless: We do not rely on any third-party cookies
- Initial recommendations are day zero i.e., there is no need for the campaign to be live to generate the recommendation.
- Focus on delivery and KPI goals: We prioritize campaign delivery while ensuring Key Performance Indicator (KPI) goals are maintained.
How did we select our final approach?
We shortlisted 14 approaches that can potentially fulfill our requirements. These approaches covered different Similarweb API data sources, calculation methods, and metrics. In total, we conducted over 200 tests across Xandr and TTD campaigns and chose the approach that gave us the most consistent results across our tests.
We utilized historical data obtained from DSP platforms spanning 90 days for the advertisers that were chosen to validate this approach.
The best approach that we chose had the following differentiating factors:
- It leverages collective insights derived from MiQ’s historical network level performance data and Similarweb’s advertiser-specific contextual web traffic insights to generate compelling recommendations
- It accounted for 14% of the IO delivery on average while containing less than 300 domains, each beating the baseline expectation of 10%.
- The minimum delivery share of our recommended domains was 7% while the maximum was 38%
- Overlap between recommended domains vs domains where the advertiser had previously delivered averaged around 10%
Please note that advertiser-specific data was only used to validate our recommendation. The solution does not use advertiser-specific data anywhere in the recommendation process and can therefore serve zero-day requests.
What do real-life test results look like?
We carried out our initial analysis on 40+ line items, and here are our observations:
- 42% of line items with our recommendation had an average 35% lower CPM than the rest of the insertion order. These line items also demonstrated superior delivery performance.
- In another 31% of the cases, we observed partial success, where either superior delivery or a better CPM was achieved.
How does it work?
Now let’s go deeper into the solution, below is a flowchart of the workflow.
Propagated Score Calculation
We start with generating a graph of over 10,000 websites that are connected. Each connection gets an affinity score and an overlap score from a similar web API. The entry point to this graph is our advertiser’s website.
The affinity score measures the relationship strength between websites/domains, indicating the likelihood of shared visitors. The specific methodology behind calculating the affinity score, as employed by SimilarWeb, is not publicly disclosed.
The overlap score indicates the degree of audience overlap or shared visitors between the analyzed domains.
The next step is to calculate propagated overlap and propagated affinity for every domain in the graph. The intuition behind this approach is that propagated scores balance how many links a website has with other websites and the importance of those links. It resembles the well-known PageRank algorithm, where each domain is assigned a weight that can be distributed throughout the graph.
So in the end, we will have a pool of >10k domains with propagated affinity, propagated overlap, and frequency scores, this forms our raw data.
Ranking
After the previous step, we join this pool of domains and their propagated scores with the DSP reports to get their delivery for the last 90 days and remove domains from the pool we know nothing about. Next, goal programming is applied to maximize propagated_overlap, propagated_affinity, and frequency of domains.
This approach assigns a score to each of the domains. We can then select the top high-scoring N domains based on business needs.
Here’s the snapshot of the final aggregated data.
What’s next?
Similarweb domain recommendation is a cookieless domain recommendation approach. This works across DSPs and is extensively tested and adopted by traders worldwide.
As we continue to enhance its efficiency, we are focusing on the following key advancements:
- Recommending a price or bid value to be associated with the recommended set of domains.
- Providing recommendations at the vertical level as an alternative solution when the third-party API lacks data for a specific site domain.
References:
Manish leads the Data Science team at MiQ’s Bangalore office. When he’s not working, you’ll find him enjoying movies, playing basketball, and crushing data science competitions. Divya is a Potterhead and proud of it! She also loves reading and sketching portraits.
