When it comes to attracting and acquiring subscribers in today’s media landscape, good data is fundamental.

During INMA’s Subscriptions Town Hall, Jonathan Harris, senior director/product at Zuora, explained that getting that good data can also be tricky.

The difficulty with data

Recommendation systems in media are among the most challenging, Harris said, for very important reasons. The data is “highly unbalanced,” with large parts of the audience that come in and don’t return, or switch between devices, or go incognito. The percentage of audiences that actually buy something is small.

“There is a lot of unobservability in the data,” he said, “and that makes targeting or personalisation challenging.”

Doing manual testing and experimentation, however, is prohibitively time consuming and full of our own biases and errors because of the difficulty of working with the data. Additionally, he said, the data is never really personalised because A/B testing “doesn’t respect minority preferences and any kind of propensity or cohort score, compressing the audience into such large buckets that it’s largely meaningless.”

Unlike a Starbucks, where you can sit at a table all day using your laptop and not buy a coffee, Harris said media businesses don’t want to give away all of their content without some kind of compensation for it. The key, then, is the requirement to limit access. 

Unfortunately, the reality of media is that “a lot of the data that comes in at the top is very sparse in terms of personalisation signals,” which means recommendation systems are working with imbalanced data, he said.

“On the other hand, importantly, we’ve seen that the boundaries of decision-making are very narrow,” he added. “If you move a metre by one page, you can have a profound impact on a number of things, including the number of people who return to the site, conversion rates, or the amount of ad revenue.”

It should not be underestimated, he said, that the “kind of North Star of targeting and personalisation must be underpinned by a backdrop of the data, and that’s extremely challenging.”

Targeting the most engaged audiences with the most data and hyper-personalising on that level is a legitimate approach, Harris said, but it’s important to recognise that is also the smallest part of your audience — and that there are other portions of the audience with less data that drive equal amounts of revenue.

Zuora has a client, he said, that was targeting people on their second session, working from the assumption that “this is a more engaged audience.” But the factors leading people to return to a site quickly is a “range of behaviours that are extremely different,” and therefore “limiting the view of the data through simple metrics like sessions or page views hides a lot of the important granularity.”

Key strategic challenges of subscription models (and how AI can help)

Harris outlined the strategic challenges of subscription models (including dynamic pricing and packaging, timely deployment of retention tools, and coexistence with the KPIs of other teams). He then drilled down on some of the specific ways AI can help support a winning subscription model. 

While we often use that term in a generic sense, we don’t really talk about what it does — and “understanding the utility of AI for use cases is fundamental,” he said.

One key feature of AI is identifying recurring patterns and relationships within the data. AI can then use that pattern recognition to refine its understanding of both the product and the consumer’s needs to improve product offerings dynamically. Being able to adapt in real-time is crucial to delivering value “both to the customer and to your bottom line.”

Continuous growth and adaptation of the system’s knowledge base, something Zuora calls a “Knowledge Graph,” can lead to things like packaging of “content adjacencies” based on a user’s unique interests. The system can tailor offerings based on the utility of the content to each individual user.

“Data quality is fundamental” to implementing a strategy like this, Harris emphasised. “There is no substitute for a well-developed Knowledge Graph that understands relationships between all of the components that make up a media business.”

One of the main challenges people face, he continued, is that trying to stitch that data together through multiple systems is almost impossible. The “consistency and cleanliness of the data is paramount,” so Zuora’s approach is to “own that data pipeline.” This allows for flexibility, extensibility, adaptability, and dynamism, while reducing the effort required to make decisions and predictions.

“We think AI will execute whatever we give it,” he said, “but actually inputs have a profound impact on output. And if you create noisy data, you create bad outcomes.”

Use case: building a structured pricing framework using AI learning

Clear subscription pricing boundaries create a structured environment for reinforcement learning agents to effectively optimise users, resulting in more accurate recommendations, improved customer experiences, and better outcomes. 

But, he said, “when we talk about building a pricing framework, what we’re really talking about is a data framework.” To do that, you have to disaggregate the audience and get people to make decisions.

“When you get the right offer to the right person at the right time, without ambiguity or noise in the data, you start to understand relationships between the thing that you showed the user and the engagement that caused the user to take an action.”

This, he said, is the reinforcement any learning model requires.

What AI is doing on your behalf

Harris mentioned several things that AI is “doing on your behalf” in the case of building a structured pricing framework — including efficiently exploring options without ambiguity, accurately evaluating customer responses to specific offers to enhance the learning process, and modeling customer responses to understand how different people will react to different offers.

He concluded his presentation with the “AI checklist for an optimal subscription model:”

• Dynamic: It must be driven by real-time data. “We don’t live in a world anymore where you can build scores, deploy those scores, and hope that those scores will always be relevant.” 

• Self-learning: It must adapt based on feedback. “We don’t have the budget to deploy teams” to gather feedback and adapt to it “at the level of frequency or scale that we’re talking about” with AI.

• Autonomous: It must automatically and continuously take the best action on your behalf.

“This doesn’t mean replacing teams,” Harris said. “It actually means allowing teams to focus on value package bundling, disaggregation of content, and other strategic ways of acquiring customers — and letting systems like ours continuously take the best action on behalf of the customer within, importantly, guardrails and a transparent environment.”

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .