This article builds on previous articles I've written and talks I’ve given. In particular, it can be viewed as a follow up to my recent article on The Future of IAP Monetization. The goal of that article was to convince you that the future of IAP monetization involves highly differentiated offers and offer types. This article is an extended discussion of the recent uproar around randomized price-points in Zynga’s CSR 2 and a survey of how prevalent sophisticated data-driven merchandising techniques already are in mobile gaming.
A few weeks ago, Zynga fumbled a pricing experiment in CSR Racing 2 (a highly profitable game which is a cornerstone of their recent earnings). From the news reports, it seems like Zynga took a highly anticipated, highly branded, piece of downloadable content (DLC) and charged different users dramatically different prices. As Kotaku summarized it:
Letty’s Sting Ray from The Fate of the Furious showed up as a premium microtransaction in CSR Racing 2 earlier this week. It cost $4.99 to unlock. Or $14.99. Or $34.99. According to Zynga, prices for the premium vehicle were set to random as part of a feature test.
There was a huge uproar, leading to a public apology (which was also criticized by some members of the community as not going far enough)
This was interesting to me (and to the entire team at Scientific Revenue) because it was a rare public glimpse behind the pricing curtain at a major gaming company.
Before continuing, I want to state three things very clearly:
- First, Zynga is not a Scientific Revenue customer. I do not have first-hand knowledge of what Zynga does, or does not, do.
- Second, dynamic pricing is what we do at Scientific Revenue. I founded the company because I firmly believe that using machine learning to optimize what you offer players, and how much you charge for each offer, is both natural and inevitable.
- Third, the Scientific Revenue point of view on randomized trials is simple: they’re an important, and indispensable, part of the pricing toolkit. They have been for a long time, and they will continue to be so for the foreseeable future—by far the best way to tease out important covariates for causal dependencies is by using randomization to build training data sets (for a longer, English version, of that sentence, see this classic article by Burtles or this recent book by Ogden).
Back to Zynga. Since Scientific Revenue is a dynamic pricing company, most of whose customers are mobile games, we got calls. Lots of them. Lots of good conversations, and lots of piercing questions from customers, potential customers and journalists alike.
And one thing occurred to us as we answered the calls – there was a lot more surprise than we expected. Game players, seasoned executive producers on hit games, and reporters who cover games for a living, alike were surprised that game companies are doing large-scale sophisticated retailing and merchandising.
Intriguingly enough, the vast majority of people know that there is some differential pricing and bundling going on. For example,
- Very few people are surprised to find out that prices in developing nations are often lower than prices in the developed world. If you’re charging $4.99 for a smallish coin pack, you’ve priced yourself out of India entirely. This is why Apple rolls out new alternative pricing tiers from time to time, and why Google introduced pricing templates.
- Almost everyone also accepts the idea of volume discounts. If you’re willing to buy a bigger bundle of coins, you almost always get a better unit price (see the Amazon screenshot below for a beautiful, and publicly accessible, example).
- Similarly, the idea of a “starter pack” or a “no brainer first offer” is equally well accepted. And in-game sales have a long and glorious tradition.
Figure 1. Amazon Coins offer Volume Discounts on Amazon.com
And, when you probe further, most gamers are not surprised to find out that flash sales are often somewhat dynamic (which users get flash sales, and what is on sale, is often decided on a per-user or per-segment basis. And is often randomized as well).
But people seem to draw a fuzzy line around the above practices. When game companies show signs of doing more sophisticated marketing and pricing, and of offering personalized or behavior-based pricing to individual consumers, people are surprised, shocked, and angered.
As I said, we were surprised by the surprise. Because, at this point, every consumer good, whether an in-app-purchase or a stapler, is dynamically bundled and priced. It’s a fact of modern life, and video games are no different.
To see this, let’s take a look at the Top 10 Grossing Games on Google Play.
Figure 2. Top 10 Grossing Games on Google Play.
There are some interesting things in that list of games. First and foremost, as has been noticed elsewhere, the rankings are fairly stable: King, Supercell, and MZ have 5 of the top 10 games (and 6 of the top 11) and have been in the top ranks for years. While some of the games are new, there’s been a lot of consistency.
Second, Supercell, King and MZ are doing very sophisticated and data-driven pricing and merchandising (and possibly the others too; I haven’t played them enough to know. But I certainly suspect that Playtika, one of the world’s leading social casinos, is leveraging machine learning and dynamic pricing).
To start with, let's look at Supercell. In Clash Royale. Supercell has rolled out an incredibly sophisticated dynamic offer system. Players get one-time and personalized offers based on player behaviors and habits, the loot boxes change percentages as you level up, cards are kept deliberately scarce, and so on.
For example, you might think that a giant chest is a simple thing – it’s a loot box that gives you a large number of randomly chosen common and rare cards, and it costs 490 gems. But it turns out that there are many different giant chests. And which one you get when you buy a giant chest is very much a function of your previous behavior and history.
To see this, take a look at Figure 3. On the left-hand side is an “Arena 3” giant chest—it’s the giant chest available to players who are actually playing in Arena 3 (only a subset of the total card set is available to low level players). But it's being offered to an Arena 10 player who had previously purchased but hadn’t purchased anything recently (and the offer bundles an Arena 3 chest, some gold and some money in what was intended to be a compelling value).
On the right-hand side is an Arena 10 chest. But it’s not the standard Arena 10 chest—it has the card probabilities altered to guarantee at least 106 bats (instead of the standard card distribution used in Arena 10 chests). This second chest was offered to a player who was extremely close to leveling up bats already.
Figure 3. Two different giant chests.
That’s personalized pricing and offer management, done with a high degree of sophistication.
Similarly, for years there have been rumors that King personalizes Candy Crush Saga by changing level difficulties when players are blocked. That may or may not be true. But it is true that King has experimented significantly with altering the sticker prices of currency packs and bundles and has run large-scale randomized trials in production with millions of users (much like Zynga did).
In fact, King took randomized trials and pricing theory seriously enough that they worked with Steven Levitt of Freakonomics fame, and published scientific articles about what they learned like this one in the Proceedings of the National Academy of Sciences.
And, of course, MZ is famous for dynamic bundling and running machine learning at scale. A quick google search finds literally hundreds of articles about the offer system inside Game of War. For example, Figure 4 is a screenshot from a 2015 article, How Game of War Makes Money discussing how MZ designed the monetization system.
Figure 4. Three different bundles from Game of War.
Note also that MZ has tried literally dozens of variants of the starter pack in Game of War (and, I would assume, tested them all in randomized trials).
Figure 5. Starter Packs in Game of War.
In fact, MZ is so good at machine learning at scale that they’ve launched data platforms to resell their expertise to other game companies.
Many large games, including at least 5 of the top 10 mobile games, already utilize sophisticated marketing platforms involving machine learning and personalization in order to customize offers for and prices for their players. While it’s hard to say exactly what a given game does or doesn’t do, it’s clear that many of the top games offer different bundles to different players, at different prices.
It’s also clear that the trend is accelerating. The number of data scientists being hired, and that fact that they’re being hired for data-driven marketing, is staggering. Here, for example, is the description of a position currently open at Amazon.
We're looking for a Principal Product Manager to define and execute a vision for dynamic pricing and personalized rewards, powered by machine learning, and delivered through Amazon Coins.
Amazon Coins (amzn.com/coins) is a virtual currency for use in the Appstore that blends one part pricing and one part rewards. Through Coins, Amazon has found a way to drive unprecedented levels of engagement with the industry’s best customers, all while generating valuable data about our ecosystem. And thanks to the performance of top grossing mobile games, mobile gaming has quickly become an industry rich with insights.
As a final point: dynamic pricing is often viewed as a morally questionable activity (as in the Kotaku article quoted in the beginning of this article). I strongly disagree with that as well. The fact of the matter is that digital goods have a close to zero marginal cost of production, and that means that there is a lot of room to discount in-app-purchases for players who wouldn’t otherwise purchase the goods. -- in Scientific Revenue’s experience, we lower prices more than we raise them (but we do so in a targeted fashion).
The Office of the White House CTO put it best when they wrote:
Economists have studied differential pricing for many years, and while big data seems poised to revolutionize pricing in practice, it has not altered the underlying principles. Perhaps surprisingly, those principles suggest that differential pricing is often good for both firms and their customers. When prices reflect a buyer’s ability to pay, sellers can often serve customers who would otherwise get priced out of the market, as with need-based financial aid for college students. Price differences can also reflect the cost or risk of serving different customers, which can discourage inappropriate risk-taking and expand the size of the market.