Media Dojo

Publishers have an inventory and supply chain management problem similar to airlines and their seat pricing. The value of the asset changes rapidly, changes according to who wants it, changes according to how it’s bundled with other offers. It’s a whirl of activity right up to the point when the plane takes off or the ad impression is served whereupon the value of the asset plunges to zero. It’s an everyday challenge to practice effective yield management for both online ad spaces and airline seats.

Savvy operators in both industries have long known what they wanted to do to crack this problem: take raw data in the form of server logs or ticket price searches, refine it into information that can be combined with other information in an analytics package, process that information into insight about opportunities to maximize revenue at the margin, and then take action based on that insight. It’s a relatively straight forward process that, technically speaking, requires a shitload of data, storage, processing, and reporting infrastructure. Thus, the “cost-effective” constraint raises its ugly head again.

However, cloud computing is starting to change that dynamic. I spoke with Tom Shields, CEO of YieldEx which is a cloud-based provider of tools for publishers to manage and optimize their online ad inventory, including forecasting of overlapping inventory and best possible campaign allocations. YieldEx became the second company to win Amazon’s Web Services Start-up Challenge, receiving $50,000 in cash, $50,000 in AWS credits and eventual direct investment into the company. For a quick overview, access the MD Tear Sheet. MD-TearSheet-YieldEx

Media Dojo: Why the focus on cloud computing?

Tom Shields: We ran into a number of problems during the 90s and into 2000 that were essentially unsolvable with the compute resources and capabilities we had at the time. The volumes of data we needed to process, the types of analytics that people wanted to do, the kinds of information that people were trying to get out of that data was pretty hard obtain cost-effectively. But we knew what we wanted to do. It’s a combination of yield management and plain old supply chain management for publishing. So what YieldEx is trying to do is apply computer science to and new capabilities offered by cloud computing to advance the state of the art.

MD: How are ad spaces different from airline seats when it comes to yield management?

TS: There’s a couple of things that make advertising impressions unique. Like airline tickets they disappear and the value goes to zero. If you have a physical object on a shelf, it may decline in value over time but it doesn’t disappear. The second interesting thing that so far seems unique to advertising impressions is that impressions can be characterized in many different ways. Those different ways have different values to different people.

So an impression coming in may be a male 25-34 on west coast looking at sports content. You might have some marketer wanting to buy sports content opportunities, others want to buy males; still others want to buy west coast. And these marketers may have different values attached to each one of those things. Part of the challenge here is that what one person considers premium inventory, another considers junk, what works for one marketer might not work for another. So there’s an optimization challenge to figure out how do I as a publisher maximize my revenue given this set of X impressions that expire at a given Y rate and given Z demand from all these different marketer audiences. It ends up being very complex and challenging, especially when you have 10 million ad impressions being served every month and you’ve got 10 thousand potential ad placements. The cloud allows us to economically crunch that kind of data unlike before.

MD: How does this work in practice?

TS: We take in a day’s worth of our publishing clients’ server log files. We process those log files overnight, turning them into clean information sets that you can run queries against in order to surface an opportunity to optimize your ad inventory. We need to be able to start up a hundred instances of significant sized virtual machines, and run them all in parallel for a relatively short period of time like an hour, 30min even. Then we need to shut them all down again and stop paying for them. That’s why our application is particularly well suited to the cloud.

In terms of the amount of server log data we need to process, it ranges from a couple of GB to a couple of hundred GB per customer per night. For some of the larger publishers, their current reporting infrastructure takes them 22 hours to process one day’s worth of server log data into information they can sell against. If you take that same data and put it in a cloud based parallel infrastructure, you can process it sometimes in less than an hour. It’s a huge difference.

MD: Is the nature of impression data changing online? How do you optimize that?

TS: Along with the Facebooks and other social platforms generating mountains of impression data, publishers are going to create a lot of that stuff anyway. How they obtain that data and use it for their own benefit and potentially sell data as well as impressions is probably where we’re going over the next 3-5 years. Publishers will probably also be combining their impression data with 3rd party impression data to get better demographics so that they can generate profile information that they can turn around and sell more effectively than an ad network might be able to.

Then there’s going to be this whole concept of measuring performance. What’s interesting is that for a number of marketing objectives that brand guys typically talk about, clicks are really a terrible measure for those things. For the publishers, the trick going to be about attribution tracking. It’s knowing, “sure that guy went on Google and did a search on the product but he saw five banner ads for that same thing before he searched and wouldn’t have known what to search for without getting the awareness from those display ads”. How does a publisher get credit for that? He won’t unless he can prove his content had an impact. And that requires data.

MD: Why do clicks rule? Should they rule for branding?

TS: It’s easy to optimize stuff you can measure. If you look at optimizing around clicks, it’s actually pretty easy because you can measure the clicks right away. You can promote your site because you can see where people are clicking and where they’re not. Without minimizing the scaling challenges involved, it’s a pretty straight forward problem. If you have feedback, you can optimize rather quickly.

The problem is that it’s hard to get feedback from activities that move higher in the funnel. Take brand awareness. How do I get feedback on brand awareness I can quantitatively measure automatically in software? If I can’t measure it quantitatively, it will be much more difficult to modify the format or delivery of the ad to optimize it quickly. If you can start measuring those things, you can start optimizing them very quick. But part of the problem is that there is this time delay between measurement and feedback because you can’t measure the impact of brand awareness until much later.

MD: Last question, do you see cloud computing as a means to collapse this float between measurement, however defined, and the ability to take action?

TS: Cloud computing offers only the ability to ingest, process and report a lot of data under far better economics. It’s no replacement for common sense on the publisher side. Whether you use cloud systems or what ever the key to optimizing your inventory is to learn what works with the customer, then go after it and earn money. How fast can you run your learn cycle so that you can maximize your earn cycle.

For marketers, learning is the overhead in the overall media buy. You’re taking your hit via lower effectiveness during the learning period of the campaign cycle in order to maximize the effectiveness of the earning period. So the shorter and more effective you can make the learning period, the less overhead you have, the more efficient the overall media buy becomes.

2 Responses to “Cloud-based SCM for Ads: MD talks with YieldEx”