Yandex Advertising Network (YAN) is Russia’s largest network with more than 20 000 verified partners and 4 billion qualified impressions per day. According to Yandex, as of June 2017 YAN has more than 65 million monthly users, most of whom don’t intersect with Yandex.Search audience. This consistently increasing audience and a variety of ad formats and placements make YAN an attractive acquisition channel. Many international advertisers have increased their investment in YAN in recent months, and Yandex support for said advertisers has followed.
Automated control of click and conversion quality in ad networks was introduced by the Yandex team more than a year ago, and has since been constantly improving. In a nutshell, the algorithm automatically applies negative coefficients to advertisers’ bids depending on conversion probability – advertisers are paying maximum set CPCs only for those clicks that have the highest chance to convert.
According to Yandex data, 40% of most valuable clicks have coefficients varying from 0.8 to 1. For 15% of clicks with low conversion probability the original bid is automatically decreased at least by half and in some cases by 10 times or more (coefficient of =<0.5), thus decreasing the acquisition costs for advertisers. Let’s look at a case that illustrates how these automatic negative bid corrections work.
Advertiser X buys wi-fi routers for 1000 RUR each and sells them for 1500 RUR each. Based on last month’s results, average conversion rate from ad networks for Advertiser X is 4%. Advertiser X calculates that in this case his maximum bid can be no more than 20 RUR = 4% (1500-1000) and sets it as his bid. There a number of websites in ad networks where wi-fi router ads may be appealing to some regular visitors however Yandex algorithm predicts that the conversion rate from these websites will be around 1%. For these few websites the system will decrease the bid to 5 RUR – Advertiser X will buy clicks cheaper and won’t miss the chance to sell another router.
To predict conversion probability, Yandex advertising quality team uses machine learning algorithms to come up with a formula that recognizes valuable session events on websites with Yandex.Metrica. Yandex team followed a sequence of steps to get this solution:
1. All conversions had to be categorized first. All goals created in Yandex.Metrica were analyzed and the algorithm was trained to recognize conversion goals from other upper funnel events (visiting Contacts or About pages, visiting 3 pages).
2. All user behavior patterns needed to be studied next. To determine what’s driving conversions, sessions were categorized as successful and unsuccessful.
3. Once characteristics of successful visits were identified, it was fairly easy to train the algorithm to predict which clicks would lead to conversions just by looking at each session. The conversion prediction formula takes into account a variety of factors: device type, targeting, user profile, website and ad content and many others. All these factors are anonymized.
After the new algorithm was introduced advertisers started receiving on average 5% more conversions for the same ad spend.
YAN campaigns are run in Yandex.Direct UI just as regular search campaigns. Text & Image Ads, Ads for mobile apps and Smart-banners (beta) campaign types can be served in Yandex Advertising Network.