Sramana Mitra: Very interesting. When you got a bunch of companies going, how were you pricing? Was it a subscription pricing model or was it a media buying pricing model where you were taking a percentage of the budget you were managing?
Daniel Nathan: We were doing arbitrage. We go a client and say, “We want to make you profitable. The ad spend that you’re going to do with us is going to be positive. What is the current CPI that makes you sure that you’re going to make money if you’re going to buy at that CPI?” Then they tell us, “We have an average return per user of $2.3.” Then we start at $2.
Starting at $2, we’re buying again on CPM and CPC completely taking the risk on our side. We were plugging our technology on their servers. Every time we’re sending users to their app, we were able to understand the kinds of users. In real time, we are able to understand how an audience that we bought for them works out. Before they even know this is good or bad, we were able to get the best quality for them.
Sramana Mitra: How do you do that?
Daniel Nathan: By plugging into their servers. The only way that it works for us is if they send us the post-install events. If they don’t send us the post-install events, we can’t work with them.
Sramana Mitra: I understand that you’re getting that data, but give me an example of a heuristic that helps you label somebody as a good user versus a bad user.
Daniel Nathan: Let’s say we are buying from CNN app and we find out that after we bought five million impressions, the hundred users who downloaded are very high quality. Let’s take Uber for example. They are looking for users to download the app and use the app. We understand that. Uber is sending us the first ride event programmatically.
What is the pattern of those hundred people? First, they were on CNN. Then we do probability. We go to the Facebook page of CNN. For example, Facebook will tell us, “75% of the CNN people are male and they were between 35 and 45.” We take that and now we know that the user has more probability to be a good user if the user is a male aged between 35 and 45, and if he’s reading news. If we buy users on ESPN, those might be great users.”
We take a bunch of data and then we try to understand how we can multiply this audience by 20, and we try to buy that audience and see if it works. This is all based on machine learning. The more data we aggregate, the smarter we are.
Sramana Mitra: You were helping your clients achieve profitable leads. Get back to the question of how you were pricing.
Daniel Nathan: The pricing is very simple. We ask how much they’re buying right now, and how much they feel comfortable to give us. We had no money and no cash and we had to make deals. We were just trying to get some deals. Whatever they gave us, it was great.
Sramana Mitra: As you were showing performance, they were giving you more budget. Today, what is the pricing model that you’ve settled into?
Daniel Nathan: We still keep that humility. When we meet a new client, we say, “We want you to be comfortable and profitable. How much can you spend in terms of cost per install?” After we have spent $5,000 or $10,000, we’ll say, “If you want a higher inventory from us, you will have to put a higher CPI because right now, we’re not profitable. We’re not able to run your ads anymore.”