Google has over the years expanded its capabilities in the Machine Learning space for ad targeting and optimization. There are a slew of ML features that Google has in regards to its ad management platforms of Google Adwords and Double Click Bid Manager. These include but are not limited to Smart Bidding (Auto Bidding), Dynamic Search Ads, Universal App Campaigns, Google Attribution, Smart Display, Expected CTR (Predicted CTR) and Meet or Beat Goal LI Strategies.
An Analogy to the Financial Markets
So for those of us who are not familiar with the digital ad landscape, Google has the most premium ad inventory available to it through Google Search (Search Ads that you see on Google), Video Ads (YouTube) and AdSense (Many display ads that you see on websites like news-sites, utility sites, etc). Google makes money on each click or ad impression served to the user depending on whether it’s a search inventory or a display inventory. To manage this inventory Google provides ad management interfaces to advertisers and agencies. The analogy of Google in the financial space is the stock exchange say NSE or a NASDAQ. The agency and ad management platform would be analogous to stock trading desk like a Karvy account or a Fidelity account. The ad management platform are where advertiser or agencies bid for available inventory by placing a price against the targeted inventory and like the stock market, the advertiser with the highest bid usually wins with a slight provision for accounting for ad relevancy through measures like Quality Score. Like the exchange might have more information on the asset being traded, Google has a lot of information on the inventory , i.e, the person searching the query for ad targeting or visiting a site where there is a possibility of serving the ad. However, a key difference in the case of Google Search Ads (only available through Google Adwords) is that Google owns the platform used to trade and buy the inventory as well . And Secondly Google unlike the Stock Exchange owns the asset (the inventory) . Regardless, these interfaces were initially meant to work very similar to the standard trading desks, in a manner that the control of what to buy and the price at which to buy the ad inventory was with the trader (advertiser).
Traditional Optimization Methods
So as a trader (advertiser) would rely on host of information which is non personal information of the user to optimize on what targeting or keywords to bid for and which inventory to let go. The advertiser would run dimensional analysis as to which keywords perform well and give best returns, what times of day or days of week user are most likely to take desired action on their website or optimize for geographies that convert the best. This was pretty similar to stock valuation or intra-day techniques used in stock market which are based on Publically Available Information. Post running these analysis the advertiser can change bidding strategy to maximize his returns very similar to how financial investors change their portfolio of stocks or financial instruments.
New Age Optimization and Lack of Clarity and Control
However, in recent times, Google has introduced machine learning techniques like Smart Bidding which take the control of auction and bid out of the advertiser’s hands. These techniques automatically makes changes to advertiser’s bids and targeting in order to optimize for return. These techniques, Google claims relies on billions of data signals to calculate the likelihood of a conversion, based on the performance targets that are set by advertiser. In recent times, these bidding techniques have worked tremendously better than the standard advertiser controlled bidding techniques like Manual CPC or Fixed Bid techniques with bid adjustments on working dimensions. One of the reason for outperformance is that these techniques involve Machine learning algorithms which apart from using standard signals which were used by advertiser, includes key data on the user that is only available to Google like the historical data of the user, contextual signals given by user like past searches or sites visited, and the likelihood of the user to click on an advertisement in a particular environment.
Few Signals used in Smart Bidding techniques
Showing an ad to a user which is apt as per the user’s online behavior works wonders in terms of getting more clicks and return but is the use of behavioral data and automated profiling of the user a breach of privacy? We believe that If the user does not intent to make his/her behavioral information public, the analogy of this technique in standard financial markets would tantamount to Insider Trading as the information being used to optimize bids and win auctions is personal information or non-public information on the user. The concept of Google being a stock exchange and the trading desk was itself difficult to comprehend, now with Google using these techniques trying to be an investment banker makes the process ludicrous. The conflict of interests and compromise of end users privacy in this model are unimaginable as there is complete lack of an auditor between the stock market, trading desk and the investment banker (all are Google here).
Here comes GDPR to the rescue
Interestingly Article 22 of the GPDR, is a right step in protecting interests of the user in this regard. Article 22 reads as follow
“The data subject shall have the right not to be subject to a decision based solely on automated processing, including profiling, which produces legal effects concerning him or her or similarly significantly affects him or her”
In lieu of the above, Google has already come out with guidelines for advertisers who intend to use interest-based retargeting or re-targeting which profile users in interest groups based on his/her online behavior. These guidelines make it clear for advertisers to obtain consent and clearly disclose to the user the details when advertiser is inclined to use this technique. However, Google has not rolled out any documentation or issued clarification w.r.t its machine learning techniques like Smart Bidding and not detailed the compliance efforts that would be needed to continue to use these automation strategies. Secondly, consent being a very tough nut to crack, the efficacy of these techniques would get drastically reduced if behavioral data of the user is taken out from these ML techniques. Thirdly, Google has not made public what elements and categories of data are used in techniques such as smart bidding which Google claims uses more than 1 billion data signals to optimize.
The probable path to tread for Advertiser and Agencies
In Summary, we would recommend advertisers who act as data controller under GDPR for running ads through Google platforms get clarity from Google for these use cases. This would impact all advertisers who intend to use the automated methods for targeting EU subjects like Smart Bidding, Dynamic Search Ads, Universal App Campaign, Smart Display or other similar optimization techniques available. The other alternative is to switch back to standard bidding practices before May 25th. The biggest casualty in the whole landscape would be App Install Campaigns since outside UAC only option available to advertisers in Google eco-system is App Install through Double Click Bid Manager. However with fate of programmatic (DBM and DS) itself in question due to GDPR, Google might want to re-introduce standard app install campaigns in the Adwords interface once again. And if Google modifies its current automation algorithms to not use personal or behavioral information going forward, the future of optimization might lie in third-party bid optimization forecasting models made by other vendors dedicated to this effort.