This is another old post from Tumblr.
In my last post, my argument was essentially: The difference of CPM/CPC/CPA is risk transfer and the fair price can be determined by looking at the mean and standard deviation of the three risks that are being shared between the advertiser and the publisher. The three risks being: click through rate (CTR), conversion rate (CVR), and average order value (AOV).
I was thinking option pricing may work to get to the right price but upon further thinking, this looks more like a job for crystal ball or some statistical simulation tool. So, if you have all the data points, you can plug them in and boom, you should have an idea of fair relative price for the each model. If you are bidding on an impression at a certain price, how much should you be willing to pay in CPA? This equation can be expressed like this:
This assumes that whatever bid price you have for an impression is the right price. In a perfectly efficient market where the bid price is always right, this makes sense. But we live in a imperfect world and advertising is especially an inefficient market, so this should be the other way around. Our bid price for an impression should be driven by how much % of sales for this certain product you are willing to give up in order to show an ad in front of this person. So the equation that we are actually trying to solve should be:
If I am making 5% gross profit margin on a speaker, I am only willing to give up 0.5% of the price. But if I sell my own brand of super expensive shoes at the GP margin of 70%, I may be willing to pay 10% of sales. So the % of sales you are willing to give up is purely an internal decision driven by the nature of the product and your risk adversity. Let’s say for this one user looking at this blog, I know the likelihood of him clicking on the ad is, the likelihood of him buying something is, and know how much he will spend, plus I know I am willing to pay 10% of sales to get this guy to purchase my product, then I know exactly how much this impression is worth to me and how much I should bid.
The only problem is, we don’t know the exact CTR and CVR and have no idea on their standard deviation. This is where big data comes in. If we can identify this user and look at his past behavior, we can make an intelligent guesstimate of the variables. We may also take context into consideration. If you are looking at a funeral website, you probably won’t be clicking on an ad for a sports car. But what is the problem we are trying to solve? It’s not just that the more data the better. We need to predict this user’s behavior as accurately as possible.
So the point of this post boils down to this. The holy grail of online advertisement is predicting intrinsic demand.
That means I want to show you an ad with a product that:
- You don’t already have
- You don’t know you want (because you would be on Google or Amazon searching for that product if you wanted it now)
- You really really want
I think in the online advertisement market, we are finally trying to figure out #3. Retargeter do this by saying, “hey you almost bought this product, maybe you would actually buy it if I give you free shipping”. The problem with that is scale. You don’t abandon shopping carts everyday so the amount of ads you can serve with this method is limited or you will be annoyingly repetitive. So, in order to understand what you really really want that you don’t know already, we want to understand who you are and your taste.
Let’s say I sell leather pants (which I don’t and have no plans to). There is a user looking at a blog and I have an opportunity to show an ad.
Here is the user profile:
- stable job
- married with children
I’m quite indifferent about this guy. Not willing to pay much at all.
But what if we also find out he loves metal, Harley Davidson’s, and wears leather jackets? This is a guy I want my ad to be shown to and I am willing to pay for it. So the more specific information I have about this guy and the more match I see with my product, the more I am willing to pay.
The race for the new new thing in online advertisement is, who can predict a user’s taste and match that impression to the advertiser with a product that maximizes his desire. At this point, advertisements should become less of an annoyance and more of a content. Amazon has a whole bunch of information on what you’ve bought before and can infer what you might like. But the cool thing about Pinterest is that it can connect a whole bunch of seemingly random products that you like, or someone similar to your taste likes and create a taste map. This can be used to understand the user but also to understand the product.
With the granularity of decision making reaching the logical extremes of “one impression” to “one product”, it looks like the field is finally set to start the data collection & algorithm battle. There is so much inefficiencies for companies to profit from and it will be interesting to see if giants like Google/Amazon/eBay will figure it out first or a new comer like AppNexus/Pinterest will come out on top. I’m excited to see Pinterest taking a stab at this with their unique asset.