March 30, 2017
About the author : Joni holds a PhD in marketing. He is currently working as a postdoctoral researcher at Qatar Computing Research Institute and Turku School of Economics. Contact: joolsa (at) utu.fi
Controlling ad quality in programmatic buying
Highway to ad quality.
Ad quality is an issue in programmatic buying where ad exchange takes place via computer systems. In traditional ad exchange, there’s a human supervising the quality of advertising, but in a programmatic system it’s possible to receive spammy, illegal, or otherwise undesirable advertising without publishers (ad sellers) being aware of it. Likewise, the quality of performance such as clicks, likes or even impressions might be compromised by fraudulent bot behavior.
In the lack of humans, how to control for quality? Well, some ways include:
- bot detection — this is what Google uses to filter invalid clicks likely caused by bots. It includes i.a. detecting anomalies in click behavior. Facebook, too, has mechanisms for detecting bots. How well these systems function should be from time to time audited by neutral 3rd parties due to the inherent problem of moral hazard by ad platforms.
- performance-adjusted pricing and visibility — again, used by Google and Facebook in Quality Score and Relevance Score, respectively. What works cannot be wrong, essentially. The ads with the best response get the most views for the less money. However, this does not directly solve the problem of removing undesirable ads from the system.
- reporting — again, both Facebook and Google enable reporting of ads by end users. This shows to advertisers as negative feedback – once negative feedback reaches a certain threshold, the ad stops showing. It is in a way crowdsourcing the quality control to the end users.
- algorithmic analysis of ad content — for example, Facebook is able to detect nudity in the pictures and consequently disqualify them. This is among the best methods, albeit technically demanding, because machine can treat many millions of ad content units in batches. With constantly developing machine learning solutions the accuracy of automatic detection of undesirable content approaches human classifiers.
- finally, we can have human fail-safe as a “plan B”. Again, both Facebook and Google use manual detection of click-fraud but also in treatment of advertisers’ complaints over refused ads. However, the solution is expensive and does not scale over millions of ad units, so it can be seen as a backup at best.
There – I believe these are the most common ways to control ad quality in modern programmatic advertising platforms. If you have anything
to add, please share it in the comments!
EDIT: Came across with another quality control mechanism: private exchanges. They effectively limit the number of participating advertisers making it manageable for a small number of humans to verify the ads. The whole point of the problem is that this works for a handful or so ads, but when there are millions of ad units, humans cannot be used as the primary solution.