Renewing Relevancy and Driving Paid Content Monetization with User Knowledge, AI and Machine Learning
Legacy media company Mittmedia is renewing local relevancy and driving customer retention in the subscription business using ux, machine learning and AI. The Mittmedia digital products are autonomously personalised, delivering the right content to the right customer at the right time. In this slide based blog post we´ll try to fly over the basics of the hard and creative work done by the excellent development organisation at Mittmedia.
Above are some simple facts on Mittmedia as a local media company.
The last two bullet points might be the most important ones.
We have a mission to contribute to, and uphold, local democracy by producing and distributing journalism and information.
That mission drives us forward in transforming our business from printoriented to a completely digital one. That mission is also the basic logic behind todays presentation. Because, to be able to fulfill that mission, we need to renew ourselves.
And the first step in doing so is to transform the idea of what kind of media company we are.
Mittmedia needs to go from legacy media company to become a partner for all of the information local people think is relevant to them.
We´ll continue to stay true to the core of unbiased local journalism. But we must constantly add new relevant information to production and distribution as our market demands it.
The journey from a legacy media company to an information partner is in our opinion the only way to go if we want to continue to relevant in a fast changing local environment.
This transformation will pose a huge challenge for us. But maybe not in a way we think. It’s easy to believe that the fast changes in superficial media interfaces that we´ve seen through the past years is the actual challenge for us. But as we´lle see will in the following sections, the actual challenge lies on a much more profound level. A level that´s all about human context.
To understand what the real challenge in the future of news consumption will be, we first need to take a look of the history of news distribution.
News have been distributed through communities for as long as humans have been able to speak. But the first time that the business model of charging a fee for mass distributing of news information really took of was when the printed newspaper started to gain traction.
This was during the eighteenth century, and to be a consumer of the newspaper came with several prerequisites. For example, first of all you needed to be within the distribution area to be able to get a hold of a paper. More than that you also needed light (which wasn’t always easy to produce before the lightbulb) to be able to see the letters on a page. But most importantly you also needed to be able to read, to get any value at all out of the medium.
In the beginning of the twentieth century news started to be broadcast to radio devices. This was of course a huge jump in regards of user experience, since you only needed to be within hearing distance of the device to be able to consume the content. Presenting the news through sound also made the content accessible to a broader range of the population.
For example: small children, not yet able to read, could still take away some information. However, there was still an extremely important condition that need to be met for the user to be able to access the content. The radio needed to be tuned in on the right channel at precisely the right time.
In 1947 the first serialized news broadcast for television started to air. Of course, the television to a large extent is an extension of the radio format, with one huge addition. When consuming a news broadcast on television the user no longer needs to visualize what is being described, in most cases the screen will guide the user towards a greater understanding of what is being reported.
To continue this historical walkthrough we should jump much closer to present time. In the nineteen eighties computers started to connect to each other using the internet. Initially, the information sent was primarily text shown on monochrome monitors. After a while images also started to appear, and by the time the internet computer started to get a real foothold in the common household the medium was well capable of showing color images.
What followed was the introduction of sound. First as stored data, but soon enough streaming audio was a common way of consuming digital content.
And today, any mobile data plan needs to take into account that a lot of the media being consumed over the web is in the shape of streaming video.
So, is it just a strange coincidence that the first iteration of news media as text, audio and video correlates to how the internet medium came to evolve? Of course not. Any computer scientist will tell you that what we actually see in the evolution of digital media is a representation of Moore’s law. That law tells us that over any given time period we can expect hardware, or more specifically computing power, to increase at a linear rate.
This means that the reason that we had text before we had streaming videos on our phones is that the text is a much simpler type of information carrier. But, Moore’s law also meant that already during the eighties most computer scientist already knew that we would be consuming digital video within a few years.
A part of the definition of a revolution is that it is sudden and somewhat unexpected. This means that the introduction of digital media is to be seen as an distribution optimization, rather than a media revolution.
The media didn’t revolutionize the user experience, but the changes in user experiences is certainly revolutionizing the media industry. And it has everything to do with new contexts.
If this was twenty years ago, and you were sitting in a room listening to presentation or keynote, you really didn’t have that many choices if you found yourself getting bored. The most obvious thing to do would be to physically change the context by leaving the room.
Today, all of us are are only a few small interactions away from putting themselves in a completely different context. If you decide to change your context today, you will almost certainly do it by putting your focus into a non physical context. You are, at any time, seconds away from catching Pokemons with someone sitting in a building next door.
This type of context switching is the true revolution that Mittmedia, and the media industry at large, needs to handle. Context switching happens by routine.
If you decide to catch Pokémons or check Instagram right now it is not the first time you are doing it. The reason that you turn to these products is that you believe that they can reward you, right here and right now. It’s a habit.
To survive, Mittmedia needs the user to resort to the habit of using the Mittmedia news product provided.
And so, the key takeaway is just that. Regaining the position as a relevant information partner in a new contextual reality is the actual challenge for us. The key to do that is to strengthen our presence in the daily consumption routines and habits of our customers. First, we need to start understanding theses routines and habits.
- Exactly who are the customers?
- How do they behave during a normal day or week?
- What kind of information is relevant to them?
- When is it relevant?
And of course, we can´t start off by guessing. We need to map them, and the first step in that is to start collecting, compiling and visualizing big amounts of data on the routine patterns of our customers.
This is also what we´ve been doing for the past couple of years.
Above is one of many data visualizations of the information habits of Mittmedia customers.
Interaction data collected from the Mittmedia products by the Mittmedia data platform Soldr.
Interaction data collected from the Mittmedia products by the Mittmedia data platform Soldr.
These patterns show us how users interact with products and content during a given period of time.
How they consume Mittmedia information and when they do it.
But as you can see, this doesn’t give us any deeper understanding of who the users actually are or what type of information they find relevant.
These are just blind patterns.
If we want to get the full picture of customers and routines relative to our content, we need to understand the individuals, the actual people, behind the data. We need to dress the patterns up with stories of customers and their preferences.
First half of this year, the Mittmedia user experience team did a huge effort in collecting and compiling stories of actual local users.
Hundreds of questionnaires and interviews were made.
Hundreds of stories of actual local people were gathered.
The stories were compared, written down and visualized in a number of ways.
After this excellent user experience work, we had some unique knowledge on our local audience.
But to be able to combine these stories with the anonymous data patterns you saw in the previous slide, we needed to format and systemize them.
In Mittmedia, we´re doing just that by fitting all of the stories into the Mittmedia user model.
This is a simplified image of that user model.
In this model, every single one of the stories of actual local people must be fitted in.
So we started doing that, mapping and systemizing the user stories we had collected from ux work.
And when we were done with that mapping, we had the ability to combine actual people with the data patterns of their behaviour.
And so, where did that get us?
The true power of the Mittmedia ux mapping becomes obvious when we are able to quantify the parameters into a model that we can feed into a system. By translating user behaviours into data models we can easily cluster the user base to fit into the different habit patterns we are looking for.
To verify that we have actually identified different types of user routines we can compare clusters. In the above case, we’ve picked two clusters that to a large extent are each others opposites. Users in cluster nine are very active in the mornings, and users in cluster thirteen are mostly using Mittmedia’s products in the evenings.
Above, we draw a graph showing both clusters on a 24 hour time frame we can see a visual representation of the cluster definitions from the previous slide.
However, even if we clearly can see the differences between the clusters on this quite recent data we need to widen our perspective to verify that what we are actually seeing is a representation of routine behaviours.
Above, we go back 100 days and plot the same users, and can clearly see that the pattern seems to be consistent over time.
Also, for the sake of making sure the we don’t have a spotty dataset we can also look at the data from day to day. The pattern is obvious.
And when we´re sure of the consistency of this, we actually now have all the combined knowledge that we need.
- Actual users and their profiles.
- The routines of their everyday consumption behaviour.
- The different types of content they find relevant and interesting.
This is the full picture of the Mittmedia customers daily consumption routines and habits. And, as you can see, we have all the information we need to start creating personalized products. We´re now able to learn machines how to distribute information and content to regain local relevancy.
And in about a minute or so, we´ll show you how Mittmedia is doing just that.
But first, let´s take a step back and question why personalized products are so important for Mittmedia. Why have we chosen the strategy of autonomous personalization in the first place?
Well, almost 18 months ago, we initiated the personalization project of which this blog post tells a small piece (to know more on the background of the project, please read this blog post).
The start of the project consisted of three months of research done by the Mittmedia data team. The results from the research are pretty extensive.
But from a ”need-to-personalize-perspective”, the project came up with one basic conclusion in the early stages.
If Mittmedia wants to be able to monetize in the digital subscription business, we need to create products where supply of content meets individual customer demand far better than today.
The data showed us what is illustrated in the below graph, and this is really all we need to know.
Above is the process in which one of our customers comes to purchase a non- personalized Mittmedia product. The interaction between the customer and the product can be described as follows:
To the left, before the purchase, the soon-to-be customer consumes information that she or he finds relevant and interesting.
In doing so, the customer becomes more and more positive to the idea of paying for the product.
Finally, the customer completes a purchase. After the purchase, the customer will expect to find more of the same kind of information that was the driving force behind the purchase. As you can see to the immediate right of the purchase line, the activity to find relevant information intensifies.
But since the Mittmedia premium product up until now has been non-personalized, the customer won’t find enough of the type of information.
The content supply doesn’t meet the customer demand.
Therefore, the customer’s activity goes down, as you can see to the far right.
This is actually the phase when the customer will start to churn out from our product.
Simply put: In a non-personalized product, content supply won’t meet customer demand enough to create sustainability in monetization.
Another way of looking at this are the next two images.
Before personalization, when a customer enters a Mittmedia product at any given time, the customer will look through a list of articles and find and consume one article of relevance. What we need to do is to create products that will give the customer opportunity to…
…enter a Mittmedia product and find and consume two articles of relevance.
This is not rocket science. This slide simply shows a product doing a better job for its customer.
So, in short, this is why Mittmedia sees the need to personalize products to be able to monetize. Our personalization project has since long gone from research to actual production. Mittmedia is now, as we speak, live with personalized products built on machine learning. And now we´ll give a short run through on how we’ve been working towards that.
On a conceptual level, this is how Mittmedia is approaching personalization.
By using data from loyal users we are able to create different user profiles that the system can learn from.
These profiles are translated to cluster data, and each user in the system is assigned to at least one cluster. When we meet new users, or when a user behaviour changes, the algorithm re-assign a cluster belonging.
Article recommendations are calculated using the cluster data as a base, and for example a user belonging to a sport cluster will see more sport content in Mittmedia’s products.
Another way that we approach personalization is by applying unsupervised machine learning to learn about the different geographical clusters that our users belongs to. The obvious logic here is that we recommend articles depending on how well the geographical metadata of the content matches with a the users within a geo cluster.
Above: From a technical perspective, the setup looks like any common applied machine learning architecture. The data used for training, modelling and verifying the output of the system is the same type of data as we talked about before when we looked at the UX data mapping.
During the development period Mittmedia have conducted several live tests on sites, mobiles applications and in newsletters. As with all research projects the results have been mixed, depending on what the evaluation of the test is focused on. Both we have in several occasions noted that consumption patterns changes if a user is exposed to a personalized content feed. The image above shows the result of one of these live tests.
A better way to take away conclusions from a test is to compare changes between the control group (not personalized) and test group (personalized).
To again put it simple, this is what we now in our digital products.
The machinery Magnus we´ve gone through is raising degree of relative consumption, creating a better supply/demand product, pushing down churn levels, and creating a sustainable monetization of our business.
As of now, Mittmedia is live with autonomous personalization of all of our news sites and email newsletters. To find out how this works form an interface perspective, please read this excellent blog post form Mittmedias head of news sites, Katarina Ellemark.
We´ll soon go live with our mobile news apps.
Machines are now running the show of distributing personalized content to our 150 000 customers, every hour, every day, every week.
Every night the machine learning algorithms are recalibrating the clustering of customers based on behaviour for the past 24 hours.
The machines are doing a far better job in distributing content than humans.
So now, we can put all of the human Mittmedia journalists at work in producing unique content instead of manually distributing it.
But of course, this system isn´t just built for todays information interfaces.
If Mittmedia wants to stay relevant in the future the systems, logic and principles behind personalization must be able to serve tomorrows interfaces in VR, AR, voice or any other interface or form.
Magnus Engström, Head of data strategies at Mittmedia
Thomas Sundgren, Head of platforms and strategic parternships at Mittmedia
