ATS London 2017: The Advertiser Strikes Back - Keynote from Dr Boris Mouzykantskii, IPONWEB

Thanks for bearing with me for another year. I'm Boris Mouzykantskii. I want to talk about advertisers. In a 20-second summary: as an industry we somehow failed advertisers. They have the money, and that's a big problem for them and probably an even bigger problem for us. I run IPONWEB, an engineering company doing machine learning—or mathematics now. I struggled to find a metaphor for the advertiser: extremely powerful but confused, like Luke Skywalker, vulnerable to bad advice and constantly searching for the right path. The reason they're so confused is that it's difficult to understand where advertising works for them, or how well it works. You can measure spend and sales, but you don't know if a consumer would have bought without an ad. That core confusion is a very difficult problem—some think it's scientific, some think it's technological. Two paths: treat it as art and do what feels right—buy a Super Bowl ad, trust agencies—or take the more difficult path of measurement and attribution. Invent a proxy signal (click, conversion) and try to equate success in driving that proxy with business success. The first path is often called brand advertising, but both paths promote the brand while seeking performance. Brand advertising used to be simple: create great ads and trust an agency. But something went wrong—we learned that trusting them wasn't a good idea after all. Don't blame agencies; look at what we as an industry did to advertisers on the technical topic of auctions. In 2009 we said second-price auctions meant they could trust us with their true value. Over time we took advantage of that knowledge. Now, in 2017, we tell them we can't keep stealing that information, so let's be transparent with first-price auctions. Now the advertiser must do price discovery themselves—an interesting lift after eight years of trust erosion. Technology can help rebuild trust in brand advertising through validation and transparency. In performance advertising, the challenge is that optimising a proxy signal with machine learning inevitably finds pockets where the proxy no longer correlates with success—bad UX, incentivized clicks, or outright fraud. Fraud isn't black and white; there's a constant battle. The industry consolidating around a few vendors sometimes helps the bad guys because they only need to beat a few filters. At the end of the day, when we game the proxy, everyone wins except the advertiser—publishers get more traffic, the ad industry gets more tax. Advertisers respond by taking their budgets away. Short-term that's understandable, but long-term it's not sustainable because internet media consumption grows. They need to bring budgets back. Technology can help by cleaning up the supply path: duplicate bids, resellers, mispriced inventory, bogus data. Then advertisers need to model the probability of winning in first-price auctions—a difficult task that we landed on their doorstep. They need to use verification methodology but remember it doesn't catch those who over-optimize your metrics. The solution: stop using common proxy signals (clicks, installs, simple registrations) and instead drive machine learning loops through custom proxy signals unique to you. For example, instead of a site visit, use interaction with a widget. That's difficult for others to game and better correlates with success. Also, incrementality testing is the ultimate proof—use a holdout group to measure true incremental sales. Typically, incremental value is low most of the time and peaks during external events like back-to-school or holidays. Advertising can't create the urge to buy; it ensures the customer picks you over the competitor when the urge arises. The standard 30-day attribution window misses that. Summing up: advertisers need to invest in true price discovery, use verification systems, optimise toward custom success signals, and understand incrementality. Once we do all that, we bring balance to the ecosystem and can help the industry. Thank you.