作者：Dmitri Williams

之前我写过一系列关于游戏分析和预测的内容，而现在我想回到一个越来越重要，但是经常被游戏开发者所忽视的主题：用户获取及其归类。

因为与广告相关，所以关于用户归类模式的争议通常都是发生于首席营销官与首席财务官之间。首席财务官想要削减市场营销预算，而首席营销官则尝试做出反击并使用参数去证实公司的广告开支的重要性。任何曾在动视会上解释过数据的人都知道这些数值的精准性有多重要以及你如何证实它们有多重要。

Advertising(from jonrpatrick)

市场营销预算就像一头反复无常的野兽。然而随着广告模式越来越受到数据的影响，市场营销团队外部人员也应该开始理解这一归类以及它对于游戏的含义。

让我们先从前面开始。当用户点击广告而进入你的游戏时，归类将给予应得的信任。它能够呈现出销售，新用户以及其它结果并明确这些用户是来自哪里，从而让市场营销者能够决定哪些资源对他们来说更加有价值。当前关于归类模式的争论是，划分所有资源的最佳方式是什么，但是所有归类的要点是着眼于不同广告资源并传达出“比起资源B，资源A能够提供更大的投资回报率（ROI），所以我们应该将广告放置于资源A。”你的ROI总是基于终生价值与获取成本之间的一些衡量。

现在我们同时也要考虑资源A可能是广告发行商的事实，但基于另外一个层面，它也有可能是一个真正具有创造性的广告。如果你着眼于发行商，你便可以只是看ROI，但我们发现效果方面具有真正的区别，这甚至是在我们添加自己的社交功能之前的事。如果你站在发行商角度来看，你便会考虑特殊创造性或格式（视频，广告等等）的ROI。这是侧重于信息传达而不是中间商。不过两者都是非常重要的。

通过创造这个真正的例子，优秀的归类模式将告诉你是否应该在Facebook或Google Search上放置更多广告—-基于哪种广告能够吸引更多玩家并赚到更多钱。随着网络广告的崛起以及它们与在线游戏间的强大联系，这变得比之前更加重要了。

当前归类模式的最大问题便是它们并不是非常精确。今天的归类总是包含许多估计而缺少足够的数据。在最糟糕的情况下，它们只能发挥一定提醒作用，就像：50%是可行的，但却不知道这50%是指什么。

归类广告的一种有效方法便是通过多点触控数据，这意味着我们将把决策制定过程分为多个步骤。想想你是如何在网上找到东西的：你将遵循一些链接并看到多种广告，然而根据来自朋友那的建议以及在亚马逊上的评论而做出购买决定。多点触控归类是将整个过程当成触控点系列，然后在整个过程中将购买信用分到每个资源上。

你有可能猜到存在多种这样的模式，并且许多关于这点的争论都是非常精确的。让我们先从简单的开始，如最后点击模式。这看似是一个非常复杂的问题的一种简单的解决方法。不管用户最后点击的是什么，都是他们将其带入进来。这是对于大多数归类公司的默认情况，所以如果你正在使用这一模式并且不曾提出疑问，你便很有可能得到这样的结果。这并没有什么错，但它可能太过简单了—-在制定决策的过程中存在许多步骤和影响，然后最后一个资源才能够得到用户的信任。基于低回报的方法将能够通过播撒种子而创造长期的销售流。

最后一次点击的反面便是首次点击。在整个链条中，这是你第一次听到一件产品的地方：来自朋友的推荐或者你看到的首个广告。因为意识生成理论，这种方法得到了广泛的使用，但再一次地，它也未完全考虑过程中的其它所有步骤。

这同样也是一种线性归类，即同等归类到链条中的每个步骤。因为它考虑到了多个资源，所以这是一种不错的方法。然而它并未考虑到不同的广告资源可能会得到来自用户的不同回应（与兴趣）。

随着归类的不断发展，这些模式也变得越来越高级。我们可以基于变得更完善的理论去考虑模式。时间损耗也开始将认知和决策规程纳入考虑中，并明确玩家兴趣会随着时间的发展而加深。首次点击可能未能分配到足够的信任，但它却会随着每一步骤而不断增长。然而这一模式可能会高估了最后一次点击。

另外一种更复杂的模式便是基于位置的分类。首次点击和最后一次点击拥有大多数信任，而剩下的信任则平均分配到每一个接触点上。再一次地：所有的价值最终都是随意分配的。

最后存在一种受数据驱动的方法。机械学习模式可以在接触流中采取不同序列，并且基于能够真正创造出最佳结果的模式而不是更重要的理论。基于大多数机械学习模式方法，结果有时候会更难理解，但它们却是最精确的。它们同时也会要求一些较稀有的技能集。我的直觉便是，这是一种长期结果，但它要求你投入一些时间去获得必要的工具与知识。

所以解决方法是什么？你可能会说适当的策略便是将广告覆盖整个网络，但有人会说你未能给予那些有效的资源更多报酬—-如果你正在为广告制定价格，也有人会说你的广告并未被低估。如果未能精确地估算广告的价值，那么点击进入的行为便会被低估，广告的价值也会被削弱，从而使你不能得到精确的ROI报告。更糟糕的是，这一过程可能会往另一个方向发展，你可能会全部买下一些被高估的广告并向公司报告预算，并因此惹恼了首席财务官。有时候你会听到“优化”这一词，一般情况下这指的是：确保你能够考虑到手上所拥有的信息和工具而做出最有效的消费。

这一系列文章并不是如何创造出最划算的广告的速成课，这只是帮助你独自找出解决方法的一些内容。在之后的文章中我们将进行更深入的探讨。

（本文为游戏邦/gamerboom.com编译，拒绝任何不保留版权的转发，如需转载请联系：游戏邦）

User Acquisition and Attribution: The Basics (Part 1)

by Dmitri Williams

Previously, I’ve written series on game analytics and prediction, and now I want to turn to a topic that is increasingly important, yet often misunderstood by game developers: user acquisition and attribution.

As it relates to advertising, the debate on user attribution models is usually one that takes place between the CMO and the CFO. The CFO wants to cut the marketing budget, and the CMO tries to fight back and use metrics to justify the company’s ad spend. Anyone who’s had to explain their numbers in a board meeting knows just how important it is that those numbers are accurate, and how important it is that you can prove them out.

Marketing budgets are volatile beasts. However, as advertising models become more data-driven, it’s important that people outside the marketing team also understand attribution and its implications for games.

So, let’s start with the very beginning. As users click on ads to reach your game, attribution gives credit where credit is due. It takes sales, leads, new customers, and other favorable outcomes and figures out where those customers came from, which allows marketers to determine what sources are more valuable to them. The current debate for attribution models is what’s the best way to divvy up the credit among the sources, but the point of all attribution is to look at different ad sources and say, “Source A gave us a bigger return on investment (ROI) than Source B, so we should probably be focusing on placing ads in Source A.” Your ROI is always based on some measure of lifetime value compared with your cost of acquisition (CoA, or CPI for cost per install).

Now also consider the fact that Source A may be an ad publisher, but at another level of detail it can be an actual creative ad. If you’re looking at publishers, that’s a fairly gross-level way to look at ROI, but we consistently find that there are in fact real differences in performance, and that’s even before we add our own social special sauce. If you’re looking within a publisher, you’re considering the ROI of the particular creative or format (video, banner, etc.). That’s looking at messaging rather than reseller. Both are equally important.

To make this a real-life example, a good attribution model should be able to tell you whether to place more ads on Facebook or Google Search — based on what ads are bringing in the most players and money. With the rise in online ads and their strong links to online games, this is more important than ever.

The big problem with current attribution models is that they’re not very precise. Attribution today involves a lot of guesstimation and not a lot of data. In the worst case, they are a reminder of the old joke in advertising: 50% works, but we don’t know which 50%.

A good way of attributing ads is through multi-touch data, which means we’re looking at the decision making process as multiple steps. Think about how you find things online: You follow link chains and see multiple advertisements and get recommendations from friends and read reviews on Amazon before deciding to buy. Multi-touch attribution looks at the whole process as a series of touch points, then distributes credit for the buy to each source along the process.

As you might have guessed, there are many models for this, and a lot of debate on which is the most accurate. Let’s start with the simple ones, like last-click modeling. It seems like a simple solution to a very complex problem: Whatever the user last clicked on is what is attributed to bringing them in. This is the typical default of most attribution companies, so if you’re using one and never asked, there’s a good chance that this is what you’re getting. It’s not flat-out wrong, but it may be too simple — there are many steps and influences in the decision-making process, and yet only the last one of those sources is what gets the credit for bringing the customer in. The approach under-rewards campaigns that sow seeds and build up steam in a longer sales cycle.

The opposite of last click is first click. In the chain, this is the first place you’ve heard about a product: the recommendation from a friend or the first advertisement you saw. It’s widely used because of the awareness generator theory (what initially built awareness is what brings the customer in), but again, it doesn’t take into account all of the other steps in the process.

There’s also linear attribution, which attributes to each of the steps in the chain equally. That’s good because it takes into account multiple sources. However, it doesn’t take into account that different ad sources might elicit different responses (and therefore interest) from users.

As attribution has evolved, models have become a more advanced as well. We can start considering models that weight based on theories that are better than first, last or all-equal. Time decay starts to take into account cognition and the decision making process, and says that interest (and therefore attribution) grows as time goes on. The first click doesn’t have too much credit assigned, but it grows with each step in the chain. However, this model might overvalue the last click that “seals the deal.”

Another slightly more complex model is position-based attribution. The first and the last click are assigned the most credit, and the rest of the credit is distributed evenly on all the touchpoints in between. Again, better: but all the values are ultimately arbitrary.

Lastly, there’s a data-driven approach. A machine learning model can take the different sequences in the touch stream and start to base attribution not on your theory of which matters, but on which pattern actually yielded the best results. As with most machine learning model approaches, the results are sometimes harder to understand, but they tend to be the most accurate. They also require rare skillsets to execute against. My intuition is that this is the long-term future of the field, but it will take some time for the tools and the expertise to catch up.

So, what’s the solution? You might say the right strategy is just to blanket the web in ads, but who’s to say you’re not overpaying for all the sources that did work — and, if you’re pricing ads, who’s to say your ads aren’t undervalued? Without an accurate gauge of advertising value, click-throughs can be underreported, and the ad will be devalued, leading to inaccurate ROI reporting. Even worse, the process can go the other way, and you could be buying up overvalued ads and sending your company over budget, much to the chagrin of that CFO. You may sometimes hear the term “optimizing,” and this is what it means in the most generic sense: making sure you are spending efficiently given the information and tools at hand.

Now, this series isn’t meant to be a crash course in how to get the most bang for your advertising buck; it’s simply a look at the ways you can figure that out on your own. And the next article will delve into that more.(source:gamasutra)