Today’s world of product development couldn’t be more different. It’s really easy to measure everything in your product – user behavior, feature adoption, conversion rates. With this comes the ability to accurately track how changes to the product impact all of these metrics, not just anecdotally but with cold, hard data.
The ability to make product decisions based on data has a lot of appeal. Tastes and convictions can differ, and nothing settles an argument better than simply trying different approaches and measuring the results. And while product management is always both an art and a science, making product decisions based on data tilts the balance toward the “science” side, which is appealing to many STEM-educated people involved in product development.
The ultimate form of this is the “data-driven” product manager, who bases product decisions on data as much as possible, and will try to set up quantifiable criteria for all product development. Improvement opportunities are identified based on where the metrics lag behind benchmarks. Changes to the product are evaluated based on how much they were able to move key performance indicators (KPIs) that were defined up front.
There are many good aspects of this approach. Instrumenting the product, collecting quantitative evidence, and setting up success criteria before launching product changes are all practices that I would highly recommend. However, going fully “data-driven” is a dead end, and here’s why.
The perils of Goodhart's Law
One of the fundamental flaws of the data-driven approach is that it relies on the unshaken belief in KPIs. After rigorous analysis of the data, you define good metrics to measure that reflect the critical drivers of your product’s customer and business value, and then you tune your product development engine to keep improving those KPIs. Sounds great, doesn’t it?
“Not so fast” calls out Charles Goodhart, wielding his law like a hammer:
When a measure becomes a target, it ceases to be a good measure.
What does this mean? It means that as soon as you make a measure a KPI and start trying to improve it, the measure is no longer useful. Let’s look at some examples. Let’s say your measure is daily active users, often used as a KPI because it is a proxy for how many users are getting value out of your product on a daily basis. Now, you send a bunch of push notifications, the number goes up. Did your product improve? Hardly. Did more people get value out of the product? Debatable.
The issue here is that most measures are only proxies, leading indicators that are hopefully correlated to some lagging indicator that we are actually interested in (like retention, some monetization event, or customer lifetime value).
It gets worse than that, however. Even if we take measures that aren’t proxies, Goodhart’s Law still applies. Let’s use revenue – not a proxy, just actual money flowing to the company, and clearly a necessary prerequisite for a thriving business. However, I can tell you a surefire way to grow revenue: sell dollar bills for ninety cents. Of course, no company is doing precisely that, but there are many that do something similar enough: spend more to acquire customers than they are getting back in customer lifetime value. If you spend a dollar on Facebook ads to acquire a customer with a lifetime value of ninety cents, then that is exactly equivalent to the dollar bill example. “Sure, we’re losing ten cents on every sale but we’ll make it up in volume” is an age-old joke that doesn’t lose its relevance. (For prominent real-life examples, look at ride-hailing companies. With billions of VC money poured into growth, it’s still not clear whether there is a sustainable business model there.)
Let’s make profits the KPI, then. Surely that must be it. Profits don’t lie. If I increase profits, then I must be doing something right? Sadly, no. Firing your whole engineering team will certainly increase profits in the short term… but does anyone believe that you can build a sustainable business that way?
In the end, taking into account Goodhart’s Law, you need to always be careful and qualitatively reason whether the changes you are observing are real, sustainable changes or whether the connection between the KPI and the long-term success driver of the business has been lost.
Baby steps to climb a hill
Another issue with the data-driven approach is that it favors incremental improvements. In general, it is far easier to find small tweaks that have a small positive impact on the KPI than it is to find big changes with big positive impacts. The reason for that is twofold: firstly, you can test small changes much more quickly. Google’s infamous “42 shades of blue” test is a good example. Hardly any effort required, just test different colors until you find one that slightly improves the conversion rate. Success! Donuts all around, we shipped a winning feature.
数据驱动方法的另一个问题是它倾向于增量改进。一般来说，找到对KPI有小积极影响的小调整要比找到有大积极影响的大更改容易得多。这样做的原因有两个：首先，您可以更快地测试小更改。谷歌臭名昭著的“42 shades of blue”测试就是一个很好的例子。几乎不需要任何努力，只要测试不同的颜色，直到你找到一个稍微提高转化率的颜色。成功！甜甜圈到处都是，我们发布了一个获胜的特征。
Secondly, these small incremental wins often lead to the current solution approaching a local maximum. You have optimized the heck out of the current solution, it runs like a well-oiled machine. If you now test a different solution, a more radical change, against it, it won’t be as optimized. If you only analyze how it performs, as measured by cold, hard data, it often won’t beat the optimized control experience – but that might just be because it’s not yet living up to its potential.
You can try teaching your product managers and product teams that failure is part of the job, that they should shoot for the moon and perhaps land among the stars. The truth is, though, that everyone likes to be successful. It’s natural to prefer running five small test out of which one or two were winners to running one big test which failed.
Moreover, because of the local maximum problem, if you are strictly data driven, you never know if you landed on a lower hill or just on the foot of a bigger hill. The data won’t tell you that. Which leads directly to the next point…
Solving the hard problems
The best products weren’t built by starting with a KPI and then trying to find ways to improve it. The best products were built by focusing on a problem and banging your head against the wall trying to solve it. Being data-driven can lead to prioritizing thinking opportunistically rather than strategically.
There are of course good reasons to act opportunistically. If you might unlock a big new sale or partnership by doing something that is ever so slightly adjacent to what you are doing currently, it might be wise to leap on that opportunity. However, the big question here should be whether that furthers the strategic goals of the company, not whether it drives up some KPI or another. Otherwise, it becomes a goose chase – an unfocused effort on doing whatever it takes to move KPIs, rather than being really, really good at solving the core problem that the product should be solving.
Perfect is the enemy of done
The last major issue with being data driven is that it can lead to slowing down the organization. If decisions aren't made until sufficient data is available, then decisions will be made later than they should be, on average. This is a bit counter-intuitive at first – after all, product analytics data is usually available in real time, whereas gathering more qualitative feedback from customers requires time to collect.
However, you can collect qualitative feedback much earlier in the development lifecycle than you can measure impact quantitatively. You can put a low fidelity prototype in front of a handful users before a line of code has been written. An A/B test, on the other hand, requires a reasonably functional implementation.
Data-driven product managers will scoff at the sample size of just doing a few prototype testing sessions. However, a lot of times even after five to ten interviews, you will see the findings converge. Moreover, often it will even take only a single conversation with a customer to uncover a flawed assumption underlying the current approach.
Do A/B testing and other quantitative validation approaches yield more “accurate”, reliable results? Perhaps. But in a world in which speed to market and pace of continuous innovation often determine who emerges the winner, and where the next hungry startup is always waiting to disrupt any incumbent who gets too complacent, you don’t have the luxury of waiting for perfect information.
The solution: Be data-informed, not data-driven
You shouldn’t take away from this that relying on data to make product decisions is inherently bad. In fact, it can be really useful, as stated at the top of this article. Accurate knowledge of how your product is being used is really valuable, so don’t disregard that. However, don’t let the data drive you. Become data-informed, instead. Instrument your product, use that data to inform areas of the product that don’t work as well as others. Use A/B testing if you are trying to optimize something, by all means – if you are trying to find the best-performing copy on your paywall, an A/B test is often the fastest and most reliable way to get results. Also, absolutely do define success measures for product changes and go back after the fact to validate that the changes did what you hypothesized.
On the other hand, also balance that quantitative perspective with a qualitative one. Ground your product decisions in mission and strategy, not just in how to achieve quantitative goals. Talk to customers. Test low-fidelity prototypes. And above all, question whether data and intuition (“product sense”) tell you the same thing. If something should move metrics and it doesn’t, or it shouldn’t move metrics and it does, ask yourself “why”. Better yet, don’t just ask yourself, but use user research to find out why. The answers to these questions will tell you more about your customers and your product than only looking at numbers ever would.