Walking the RTB Talk
By Chris Back
Realtime bidding (RTB) as part of a DSP’s offering has gone from being a distinguishing capability that sets a DSP apart to functionality that is now considered table-stakes. Despite this shift in perception about RTB, there is still significant confusion about the precise advantages of RTB and why it all matters.
A DSP can use RTB to solve three problems: reach, precision and effectiveness.
Before exploring the details of how RTB solves these problems, I’ll explain the the mechanics of buying a single ad impression via RTB.
To start the entire process, a publisher must put an ad exchange’s ad tag on their site. At its simplest, the ad tag identifies the publisher, site and ad size, but may also include some information on brand restrictions to avoid channel conflict, or constraints on creative format, among other things.
When the page containing the ad tag is viewed by a user, the ad exchange receives this ad request and layers in its own data including publisher information, anonymous user ids, and geographical data derived from geo-ip lookup.
The ad exchange then goes through its own process of determining which of its participating buyers are eligible to purchase the impression (see earlier post, Drinking From the Fire Hose) and sends a formatted message, called a “bid request,” to each potential buyer participating in the exchange.
Buyers have a brief period of time (on the order of 25 ms) to decide if they are going to bid, and, if they are, to determine the constraints of their bid and send it back to the exchange. Once all the buyers have either responded or timed out, a winner is determined by the ad exchange and the ad tag of the winner is delivered to the page.
The amazing thing is that all of this processing happens in the brief period of time between when a consumer clicks on a link for a page and when that page renders in their browser. The timeout value, or period of time an ad exchange allows between a bid request and a response, varies from as low as 50ms to as high as 120ms, including any network latency. To keep things in perspective, a human blink clocks in at around 350ms. At OwnerIQ we manage our average internal bid processing to achieve below 25ms average processing times. Thus, at peak hours, we’ll make almost 6,000 ad buying decisions in “the blink of an eye”.
Solving Advertiser Problems
Given the volume of ad opportunities flowing through the ad exchanges and the ability to bid on each one, a DSP with RTB capability can solve the first problem: reach. Rules within the bidding software can be configured to look for a predetermined list of criteria associated with the ad opportunity, such as domains, classes of sites, or even browser type and return a bid with a CPM rate defined at the campaign level.
There are some obvious problems to the simple buying model just described. First, if everyone is evaluating an impression against the same data points, the bidding model and therefore the DSP itself is quickly commoditized. For a DSP to truly thrive and provide value in the RTB space there are two obvious next steps: adding value to the inventory through proprietary knowledge (adding precision) and optimizing campaign performance based on past observations (increasing effectiveness).
Adding value to inventory is no easy task. At OwnerIQ we have dedicated significant resources to learning more about the user. We gather this user data in three ways: 1) our own property, ManualsOnline; 2) direct relationships with publishers at every stage in the purchase cycle, such as shopping comparison and even product recycling; and 3) direct relationships with manufacturers and retailers to tag consumers during the research and purchase process. This robust network of data sources allows us to determine previous purchase behavior and intelligently predict future buying intent. And, like most other RTB participants, we can augment our proprietary data with information from third-party data providers.
As difficult as it is to establish the relationships and support the technology to gather that data, it’s just as hard to make it actionable in the tight time constraints of real-time bidding. Now, as part of that 25ms response time, we will need to pair the exchange data with our value add information about the user. We are gathering about 5 million data points per day against almost 1 million unique users. A DSP needs to find a mechanism for keeping that data readily available inside its bidding system, ideally in memory (the average random access time for the fastest hard disks is about 5ms). The ability to keep data of that size instantaneously available is beyond the capabilities of standard database systems. Doing that at scale requires very fast memory-based database systems backed by redundant on-disk storage, technology that is still in its own infancy stages. If you are interested in more information, take a look at the good work being done by the Apache Cassandra team.
At this point, we are finally in a position to offer the last promise of RTB, campaign effectiveness. This is also where the process becomes much more computationally complex. When a bid request is received, we need to evaluate every campaign’s creatives against all data points available to us. Each running campaign may value an impression differently based on OwnerIQ user information, how we have classified the site, number of ads on page, the browser the user is using, the user’s geographic location, and frequency constraints to name just a few of the variables. The decisioning isn’t just buy or no-buy, but determining a true value to the advertiser to show a specific creative to that particular user on that page at that particular time.
To accomplish that computational feat, OwnerIQ must re-evaluate our logic throughout the day for each campaign. Our bidding logic evaluations don’t determine the final CPM for a campaign, but a creative specific pricing algorithm. In essence, we are doing some of the pricing computation early, and then allowing our bidding software to plug the final few datums into the formula at the very last millisecond. All so we can pick the right creative and price within our 25ms.
It’s still very early in the RTB world. The myriad of ways to make ad buying decisions in RTB is growing, and I expect it will become more complex as software engineers start applying intelligent out-of-band data processing that helps make better pricing algorithm updates. While today using RTB as a means to simply achieve reach might be enough, that won’t satisfy a savvy media planner a year from now. The industry leaders will be able to plug multiple targeting parameters into unique, campaign-specific pricing algorithms in real-time to deliver on the promise of precision and effectiveness in addition to reach.