Following the previous post about what computers can and can’t do, let’s go to the specifics. Fortunately during the last years it has become clear that Data Science has huge capabilities for changing the way we are doing business.
One of those cases is programmatic buying. The programmatic platforms have completely changed the way brands invest their media budget and optimize their results.
However it is really important to understand that’s behind the scenes in order to really understand what can and what can’t expect from Programmatic buying.
- What’s programmatic buying: Programmatic buying refers to the technology that allow investing the media budgets through real time bidding (RTB).
- Real Time Bidding: RTB is a means by which advertising inventory is bought and sold on a per-impression basis, via programmatic instantaneous auction. Which means that advertising buyers bid on an impression and, if the bid is won, the buyer’s ad is instantly displayed on the publisher’s site.
So far so good, however this is not why Programmatic Buying is getting all its attention. What’s the value of advertising? Converting people into clients, rights? So let’s split that in two options.
- Converting people into clients today: The only way to measure this is by sold products or services.
- Converting people into clients tomorrow: We normally talk about branding to all those actions that are not related with an immediate conversion. But can we say that a branding action will generate a future client? We don’t know. Actually don’t know what is going to happen in the future, what we can do is predicting the future with an acceptable level of risk. In order to do so, we need to have a metric that can be related to conversion and branding is definitely not that metric. Brand tracking research is oriented to identify Recall Levels. What we should measure to understand the potencial future sales is Purchase Intention instead. PI measures measures how willing will be a person to purchase our products.
In both cases we have to be interested in how much does a new client cost. Compare that with the margin (price minus cost) and if the result is positive in a percentage that we are expecting from our business (15%, 20%, etc) voila. If not, we have to work in up-sale and cross-sale activities in order to make that specific client repay the acquisition cost and generate the right margin. So CPM or CPC are metrics that tell us really not much about performance unless we convert that into acquisition cost. So, let’s do that.
Acquisition cost based on CPM: (Impressions/1000)n x CPMn /(Sales conversions)n
Acquisition cost based on CPC: (Total Clicks)n x CPCn / (Sales conversions)n
where n=period of time
So basically what we need from any platform that uses technology to invest our media dollars is that they can work on any of the above variables to make the acquisition cost lower and the quantity higher.
So, how does the Programmatic Buying platforms works and how they can assure you that they are achieving that goal?
Well, let’s start from the beginning. How does a Programmatic Buying platform work?
- User views web page.
- Website request ad from ad exchange.
- Ad exchange request bids.
- Advertisers respond in less than 100ms.
- Ad exchange returns winning ad to website.
So basically on one side we have a technology “deciding” how much paying for a specific ad, on the offer side, we have a technology “deciding” who is paying more and (ideally) who is more relevant for a specific ad, and all that must happen in microseconds. But is everything happening really in microseconds? Well, actually is not. We all love talking about real time and even when in this case the reaction is in real time, but that is taking for generating the output, is not. If you want a real time answer you need to have the data pre processed and stored in an small table, ready to receive a query. What is not possible is running the algorithms using the current data in real time and generate an answer in 100ms.
Let’s see some technical challenges on RTB:
- Real time strategy: The ideal bidding algorithms should adapt the bidding strategy based on what’s happening right now and not in variables that were processed the last day or the day before. But that would be like tying your shoes while running.
- Tight algorithm response time: You need an infrastructure that allow you to process the CPM against the clicks, generates the CTR and generate an optimal bid value in less 30ms so the bidder can submits the ad exchange on time.
- High dimensionality of feature space: The quantity of variables that has to be process to identify if you want to pay more or not for specific ad (remember that what matter is the acquisition cost) it’s really extensive: Time ranges, day, ad size, ad placement, publisher and app/website among many others. Additionally we will be able to integrate information via other platforms using a DMP: Location (city, state, country), device type/model, attitudinal information, connection type, browser, OS etc. You may also have access to first-party or 3rd party user demographics or interest data. If you combine all the features with their metrics you will quickly realize that we are talking about terabytes of data to be process in a very short period of time.
So here the problem is definitely quantitative and not qualitative. In this case we are not dealing with the problem we’ve mentioned in the previous post about the generation of artificial conscience because as we said, in this case the variables can take a limited amount of values (even when they can be a lot).
So the solutions can be address by:
- Processing power: There’s not much we can do on this side more than using scalable cloud infrastructure on demand.
- In this case we can definitely work on composed algorithms that can work with packages of data. That reduces the required processing power a lot. Instead of working with each feature against a conversion event (Imagine the combinatory between features, metrics and results can generate can required millions of unique processes) we can work with packets of features and metrics against the conversion event. Reducing the processing power and allowing taking real time data.
So RTB is here to stay and I’m sure we will be solving the technical problems one by one. My main concern is about transparency. In this type of business it’s very easy, and then very risky, influencing business decisions. Just imagine the lift in revenues that can be generate if just one day we do “some tuning” to the algorithm and the first image in my mind is the people from finance saying, let’s do it, just tune the damn algorithm a little bit, but you know, that wont ever, ever happen…