Network Health

Network Health is a way of thinking about the edges of a network and trying to understand what is the health of those nodes in terms of (a) the positive/negative feelings those nodes experience & (b) the positive/negative experiences that those nodes create for other nodes.

Network Health

In Patrick O’Shaughnessy’s excellent podcast episode with Zack Kanter they discuss the concept of Network Health. Zack first came across the idea during a call with Alfred Lin of Sequoia (formerly COO & CFO of Zappos).

In this post I produce a working definition of Network Health and attempt to explain how and why the concept fits into the world of business and investing.

Outline for this post:

  1. Back-to-Basics - What is a Network?
  2. What is Network Health?
  3. Why Does Network Health Matter?
  4. What Are Some Examples of Network Health?
  5. What are the Nuances Around Segmenting a Network?
  6. Summary

1: Back-to-Basics – What is a Network?

To paraphrase Zack, Alfred & Patrick:

Almost everything can be thought of as a network. You reading my post is a network of two nodes (you the reader and me the writer). Uber is a network of drivers and riders. Amazon is a vast network of engineers, customers, suppliers, third party vendors, etc. The internet is a network of connected computers all over the world (in fact the internet is often described as a network of networks). You and your spouse/significant other are a network of two, and if you have kids then the number of nodes in that network expands. Etc etc.

Below are some simple examples of networks and their nodes (please forgive my lack of artistic skills.! Also note I drew these when I used Substack vs Ghost):

2: What is Network Health?

To define what Network Health is, it is first easier to describe what Network Health is not.

Network Health is not the same thing as high-level metrics like revenue growth/number of users/gross merchandise value/average user time spent on an app, etc.

Most importantly, Network Health cannot be ascertained by looking at the average node in a network. In fact, Network Health is not even a single measure at all.

Instead it’s a way of thinking about the edges of a network (10th/90th percentiles etc) and trying to understand what is the health of those nodes in terms of both (a) the positive/negative experiences those nodes experience, as well as (b) the positive/negative experiences that those nodes create for other nodes in the network.

Here’s Daniel Ek, founder & CEO of Spotify on the topic:

The larger you become the more important it is to move away from averages and start looking at distinctly different segments instead. And ideally you shouldn’t even do that. You should just have truly personalized, twin sets and experiment on twins within your universe… at our scale today the reality is that we don’t have one user – we have many, many different users with very different needs and using our product very differently, too… we also have many different artists that are working with us and some have labels, some don’t. Some are so big that they’re almost entities in their own right… And so, health will be defined as slightly different things for all those different ones.

My (working) definition of Network Health is the following:

The experience of quality that each network node either feels itself or creates for others within the network.

This can’t be boiled down to a single number, nor can it be ascertained by looking at the ‘average’ node in a network segment.

3: Why Does Network Health Matter?

Network Health matters because it is a leading indicator of the durability of a network’s higher-level metrics like revenue growth, etc.

A network can experience an improvement in high level metrics such as growth in total revenues and simultaneously experience a deterioration in Network Health. For example, if Facebook doubled the number of ads shown in the news feed then revenues would likely increase in the short-term, but at the cost of degrading the user experience. Ultimately, this is not a sustainable practice.

Knowing how to measure the health of a network, and then actually measuring the health of that network, is extremely important when running a business. And when investing in a business it’s important to go a level beneath averages and high-level metrics and inspect what is going on at the network’s edges.

A friend of mine called Steven Rooda described the difference between Google Search and Bing: “[Bing is] probably just as good at delivering search results for 60-70% of total search results done. But Google is infinitely better at the long tail of random searches because of the health of its [edge] nodes.”

Google vs Bing is a good example of how Network Health (combined with massive economies of scale) can lead to dramatically different business outcomes over time.

4: What Are Some Examples of Network Health?

Zack Kanter mentions that when Airbnb thinks about Network Health they probably look at things like the percentage of listings that have never been booked. This is an important measure of Network Health to consider for two reasons:

  1. Those sellers are having a bad experience with the service.
  2. Those nodes are likely not a good value proposition for buyers. Potential buyers are only willing to spend a certain amount of their day browsing Airbnb’s website/app, therefore displaying low quality listings is a waste of their attention and detracts from their experience on the app.

Zack elaborates that when Uber thinks about Network Health they look at (at least) two things:

  1. The percentage of ‘zeros’. A ‘zero’ is a passenger opening the Uber app and having no available Uber drivers in their area. For example, if you open your app in Chapel Hill, North Carolina at 4am and there are no Ubers available, that’s an unhealthy node in the network. This can be solved by increasing driver density and incorporating surge rates, etc.
  2. The amount of time that drivers spend driving around without rides. This is important because if drivers aren’t making money then they may churn off the platform or start dual-driving for other services like Lyft/GrubHub. This can be solved by increasing passenger density i.e. bringing more Uber passengers onto the platform within that driver’s market.

Daniel Ek discussed two examples of network health measures at Spotify with Patrick on ILtB:

  1. The percentage of customers that Spotify helps discover at least one item of content each month that the customer absolutely loves. Ek elaborates that one of the most important jobs-to-be-done by Spotify is to help subscribers discover amazing content, and that if Spotify can do that then there’s a strong chance that the subscriber will remain a subscriber for a long time to come.
  2. The percentage of artists that are self-sufficient just through Spotify. This is important because part of Spotify’s mission statement is to make it possible for a million artists to live off their art.

Patrick elaborates: On the one hand Network Health is a simple concept to understand – every network has nodes and there are links between these nodes. You can typically define these things with relative ease. On the other hand, what really drives insight into the health & sustainability of a network is thinking a layer deeper about what would result in a high quality link – is it frequency/ease of connections made, is it node intensity, is it discovery?

Finally, this concept must be exercised with caution since it can easily go awry, particularly when trying to improve engagement metrics. For example, when Snapchat sends you an overload of notifications or when you log out of Facebook and they send you emails attempting to bring you back onto the platform. Those are examples of someone trying to improve the health of you as a node in a way that neither adds value to you nor improves the experience for you if you do come back onto the platform.

5: What are the Nuances Around Segmenting a Network?

There are many factors that go into how one might segment a network. For example, the degree of homogeneity between nodes can make a large difference into how one segments a network.

For Uber, each side of the network (i.e. drivers vs passengers) is very homogenous within their geographical market e.g. passengers in Chapel Hill, NC simply want to catch a ride, and drivers in that vicinity just want to make money from driving a passenger. Uber therefore probably considers geography to be the primary manner of segmenting their nodes. This isn’t to say that this is easy, or that there isn’t more to it than geography (for example there’s UberX/UberXL/Select/Black/POOL/SUV). My point is merely that the goal for driver nodes in the network is to drive passengers, where passengers are essentially commoditized (assuming they don’t throw up during the ride, etc), and the goal for passengers is to get from A to B, where drivers are essentially commoditized (assuming the driver’s car doesn’t reek from the last Uber Eats order the driver delivered, etc).

However, for The New York Times there are multiple types of subscribers, not only across products (e.g. Cooking vs Crosswords vs core News) but also within each product: some news subscribers subscribe to the print edition as well as digital and care about regular, on-time delivery of their paper, whereas other subscribers are digital-only and may only be interested in reading articles on foreign policy, etc. Some digital subscribers are part of the NYT’s promotional offering of $1/week for a year, whereas others have subscribed to the product for much longer and pay $17/month.

For the NYT, segmenting customers by tastes, preferences and tenure with the product is far more nuanced than it is for Uber largely because neither side of the NYT’s network is as commoditized.

6: Summary

Network Health is a critical measure of the quality & durability of the experience of all stakeholders, which is a critical input into long-term shareholder value.

Improving the health of a network requires first that you correctly identify the best measures of Network Health within each appropriate segment (or, ideally, for each individual node). Importantly, the average node within each segment is unlikely to tell you much about the health of that segment. Instead, consider the nodes at the edge of each segment and see what can be done to sustainably improve their health.

Disclosure: I have never worked at any of the companies discussed in this post. However, I am a shareholder of GOOG, FB, NYT & SPOT.