We asked OctoClick's ML engineer, Mikhail Gulyaev, to share a bit about what’s under the hood of our CPC model.
🔍 Let's break it down!
How is the advertising auction structured?
Suppose there is space on a website where the webmaster places ads.
When a user visits the website, an auction starts, where various creatives participate. The creative with the highest bid wins—and this creative is shown to the user.
Previously, the winner was simply determined: the advertiser willing to pay the highest amount for 1000 impressions (CPM model) won. With the introduction of the CPC payment model, a new challenge arises, as creatives with different payment systems (CPM and CPC) can participate in the same auction. Since CPM implies payment for impressions, and CPC for clicks, direct comparison between them is impossible.
This is where the ML model comes in: it evaluates the click probability for each creative and recalculates the bids for the auction based on that.
When a user visits the website, an auction starts, where various creatives participate. The creative with the highest bid wins—and this creative is shown to the user.
Previously, the winner was simply determined: the advertiser willing to pay the highest amount for 1000 impressions (CPM model) won. With the introduction of the CPC payment model, a new challenge arises, as creatives with different payment systems (CPM and CPC) can participate in the same auction. Since CPM implies payment for impressions, and CPC for clicks, direct comparison between them is impossible.
This is where the ML model comes in: it evaluates the click probability for each creative and recalculates the bids for the auction based on that.
ML Model
During the search for the best option, many hypotheses were tested, and a large number of experiments were conducted.
The model’s parameters and feature set were carefully selected, including various user information, site characteristics, and creative features.
Innovative approaches from AdTech and ML fields were applied when creating the model, which allowed achieving good prediction accuracy and stability in the solution's operation.
The model’s parameters and feature set were carefully selected, including various user information, site characteristics, and creative features.
Innovative approaches from AdTech and ML fields were applied when creating the model, which allowed achieving good prediction accuracy and stability in the solution's operation.
How does the model work?
Data about the user, information about the site, creative, and many other features are input. The output is click probability predictions.
What’s more beneficial for the user: CPC or CPM?
It’s important to understand: if the click probability is too low, the ad won’t be shown to the user. This optimizes the advertiser’s spending, helping to avoid costs on ineffective impressions.
After a month of testing, we observed the following:
The average daily CTR for CPC creatives is more than 2.5 times higher than for CPM creatives, meaning the advertiser gets 2.5 times more clicks for the same number of impressions.
💡 Useful note:
If the price per click is set too low, the creative might have fewer impressions because it will lose out in the auctions to those with higher click bids.
After a month of testing, we observed the following:
The average daily CTR for CPC creatives is more than 2.5 times higher than for CPM creatives, meaning the advertiser gets 2.5 times more clicks for the same number of impressions.
💡 Useful note:
If the price per click is set too low, the creative might have fewer impressions because it will lose out in the auctions to those with higher click bids.
Conclusion
We continue to analyze the model’s performance, optimize algorithms, and implement improvements to make the advertising auction even more profitable and convenient for all participants.
🙏 Thank you for your trust! 🚀
🙏 Thank you for your trust! 🚀
