Within the realm of programmatic media-buying, real-time bidding is one of the most popular and well-known types. Many people associate this form of advertising with instantaneous action, and it is true that a media-buying decision is made within the span of 200 milliseconds, but split-second timing does not define “real-time.”
Real-time is a relative term. While it is commonly made analogous to actual elapsed time, it is actually based on events as well as timing. The computer industry coined the term as a way to describe a constraint in which an event creates an operational deadline for a system response to occur. In other words, real-time depends on the process at hand. Within a window of opportunity, an action needs to take place while it is still relevant.
Let’s pretend that you are driving home from a party. As you drive, you are bothered by a comment that was said to you at the party. It was put forth jokingly, but it left you offended, you suddenly think of a witty return to the statement made to you. You say it out loud, mockingly, to your car windshield, as though the person at the party can still hear you. With this, you feel much better. Your face relaxes, and your attention turns elsewhere.
In this scenario, your rejoinder in the car was not made in real-time. You had left the party, and saying it out loud was no longer effective; the offender was no longer with you. If you had uttered the comment back while he or she was still standing in front of you, then your response would have occurred in real-time.
So, real-time is a time window during which an effective response can influence an event while it unfolds. For conversations, it may be seconds. For international affairs, it may be days or months, depending on the event. For self-driving cars of the future, it may be milliseconds, as it is with programmatic media-buying.
Looking at programmatic specifically, real-time can apply to more than just the bidding process. It can also pertain to performance feedback, optimization, and audience targeting. If a campaign is mid-flight and a partner’s machine-learning system discovers a whole new intender audience that was previously un-tapped, then it can adjust the targeting of the campaign to serve ads to these consumers going forward. The targeting alterations were made in real-time only if they were made while they still affected campaign performance. Given the currently observed dynamism of audience response, real-time for audience targeting may be approximately 60 minutes.
If you look at a programmatic campaign in terms of two phases – the bidding phase and then the feedback phase that optimizes performance – then this progressive definition of “real-time” applies most to the feedback phase. Machine learning systems apply information about how a consumer engages (or does not engage) with an ad once it is served. The time that it takes for this to occur varies from one system to the next; some may update every 24 hours, others every 15 minutes. As long as optimization occurs while the newly detected targeting is still the right targeting, then it occurs in “real-time.”
Audience relationships to products differ in how quickly they change. Some may be steady for days, while others may change within the hour. However, the system that optimizes every fifteen minutes is ultimately much more effective at moving away from consumer personas that have stopped engaging with an ad, and at identifying top performing personas and scaling out this target audience to include new audiences that are performing better. This is true for two reasons. First, slower systems may miss a chance altogether. Second, faster systems get in early and capture more of each window of opportunity.
To uncover differences from one programmatic partner to the next, a brand or advertiser might dig deeper when a learning system is described as working in real-time. Every real-time bidding system must meet the 200-millisecond deadline for returning a bid to each opportunity. However, only one aspect of campaign performance benefits by simply participating in the auction. To get to a clear view of how the system delivers on audience, time, geo, and other optimizations, it is better to ask about the specific amount of time that it takes for incoming data to be processed and for a campaign to be optimized by the system.
In the end, “real-time” is an example of a term that is commonly used throughout the programmatic world but whose meaning is rarely focused on. This applies to other terms like “algorithm,” “transparency,” “second party data,” “premium programmatic,” et cetera. As an industry, especially one that is evolving so quickly, we need to make a concerted effort to define these terms, and avoid using them in a situation where they can be ambiguous.