Category Archives: Technology trends

Pareto/Power Law distributions: the needle in the haystack

I wrote previously about the prevalence of Pareto/Power Law distributions in product users’ behavior here.  Wow, that’s a lot of alliteration.

But the discussion stopped at only one dimension of data.  For example, a single dimension like Free versus Paid users of a Freemium product such as online backup.

The story gets really interesting when you consider multiple dimensions (aka variables) of data at once, each with its own Pareto characteristics.  The outcome can lead you to a some very interesting places.

In the first scenario, a small set of users in Dimension One (let’s say, Paid product users) also represents a small set of users in Dimension Two (let’s say, country of user origin).  This can mean that a tiny percentage (sometimes less than 1 percent!) of an entire user base represents almost all of the revenue or commercial value.

When this happens, it’s incredibly important to know who these users are; you’ll need to hang onto them for dear life to protect your revenue stream.  For example, you might cater to the specific needs of users from their country of origin.  Do you think users in China have different product needs than in France?  Probably.

In this scenario, you’ll also need to consider a revenue diversification strategy to protect your risks of relying on such a small segment.

Another scenario is that users in Dimension One (again, Paid users) don’t belong to the majority (or, “head”) of the distribution within Dimension Two (again, country of origin).  In which case, the implication is that country doesn’t matter in targeting your best (e.g. paying) users.

You can go astray in this scenario by looking at country of origin in isolation.  Maybe you have a huge pool of users from Germany.  The temptation would be to conclude “Germany is my most important market”.  Unless you knew that paid users didn’t cluster around a single country and that Germany was comprised of lots of free users.

What to conclude?

One: make sure you know if your most valuable user segment is much smaller than a Normal distribution would imply.  Most people think that their most important user segment is something like 10-20% of their base.  If 1% of your users drive the business, know who they are, find more like them, and don’t lose them.

Two: don’t let any one dimension of data drive your definition of user segments and internal decision-making.  If you hear sound bites inside your company like “German users are our most important”, that’s being too imprecise.  It generally takes 2-4 dimensions/variables to be precise about a user segment and to know how to best treat them (“Paid users with broadband PC connections in Germany are most important”).

Three: if you truly have 1% of your users driving the business, consider diversification strategies.  You’re carrying a lot of risk, but you also have 99% of your users from which another valuable segment can be found and served.

Last: as I argued in the prior post, it’s easy to dismiss the Pareto effect as only applying to obvious examples like Freemium for online consumers.  I’ve found the same patterns in other businesses.  In which case the gap between reality and perception is even wider!  Spend some time hunting down these patterns inside your company.  I promise you will be rewarded with new insights.

Our brains are wired somehow to think of everything in terms of a Normal Distribution, aka the “Bell Curve”.  It’s a trap that can kill a tech company.

The shape of the curve means that we think of populations of data (such as users) as being a somewhat homogeneous group if only we could compute the average.  For example, how many minutes per day “on average” a user spends on a website.   Or, the percentage of people “on average” who actively post on a social media platform.

The problem is that populations of people almost never behave in a normal distribution when online or using software products. Instead, the more prevalent pattern of behavior is a Power Law, or Pareto Distribution:

The Pareto distribution is also known as the “80/20 rule”.  Except that in online worlds, the ratio can be even closer to “95/5″.

Think of Freemium business models.  Generally, 2-8% of users consume a paid offering.  The rest use the free version.  Power Law/Pareto distribution, not Normal.

Think of participation in social media.  1% are active contributors, 10% are intermittent contributors and 90% consume but never post.  Power Law/Pareto distribution, not Normal.

These steep Pareto curves have profound meaning on making choices in running a technology company.

If you operate a Freemium business but don’t know which users are the 5% most likely to upgrade to the paid version, then you risk catering to the needs of the Bell Curve: a population of users that looks more like 50-60% of the whole.  Who don’t necessarily pay or monetize.

This is the trap. Chris Anderson touched on this in his book “Free”, by illustrating how the Power Law distribution drives monetization in Freemium business models.

There are other traps by thinking in Normal terms.  Beyond Freemium, the Power Law distribution of behavior still applies.

Take Enterprise business models.  Every user is a payor, of approximately the same fee.  Yet 2-10% of a user population is massively active versus the rest.   And with that 10% of users comes maybe 10-20% of the revenue.

Which is your most important segment? Are you trying to solve the problems of those 10% “power users”?  Or the needs of the rest?

An example: I managed a product that enabled monitoring of corporate networks and systems for the sake of spotting anomalies.  Anomalies which could indicate a security breach in progress, or the risk of one.

Some users spent a large percentage of their day performing the monitoring function for the company.  They were specialists who used the product intensively throughout the day.  These power users had distinct needs, such as the ability to mine and explore data in depth to spot anomalies for themselves.

The rest of the users were different.  They weren’t monitoring specialists.  The monitoring role was only one of many roles they played for their companies.  Thus, they wanted to spent the least amount of time possible in my product.  Instead, they expected the system to alert them automatically, and offer specific actions to take.

Two user populations.  Two very different sets of needs.   One “market”.

Knowing who your core audience is, and the nature of the Power Law distributions, is essential in setting priorities on which segments to serve.  And those that can trap you.

In this post, I’ve only been discussing Power Law in one dimension of meaning (free vs. paid, automated alerting vs. manual trend-spotting).  Some of the most interesting Big Data analytics findings come from combining multiple dimensions of meaning, each with its respective Power Law behavior (a simple example: free/paid combined with locale).  I’ll tackle that one in a future post….

I’ve been driving a Big Data initiative at work.  We use the Hadoop technology stack extensively.  The Hadoop logo looks like this:

This morning, I woke up and started making coffee.  As I do every morning, I placed a paper towel on the counter to catch my coffee spills.  Except this morning, the paper towel caught my eye:

The similarities are striking, no?  I mean, Hadoop is popular and all, but I didn’t realize it is now marketed to Tesco customers in Prague via paper towels.