Case Study: USA TODAY – How AI Is Revolutionizing News Publishing
With social media algorithms constantly changing, it’s time for news publishers to regain control of their audiences. Using AI and machine learning, publishers can program contextually relevant and personalized content down to the viewer level on their owned-and-operated sites for both breaking news and day-to-day operations. IRIS.TV and Gannett demonstrate how AI is helping to provide prescriptive insights into audience engagement and fueling future content strategies and video distribution on owned-and-operated sites.
Field Garthwaite, Co-Founder and CEO, IRIS.TV
Kara Chiles, Senior Director of Product Management, Gannett/USA TODAY NETWORK
Field Garthwaite, CEO & Co-Founder, IRIS.TV
My name’s Field Garthwaite, I’m the co-founder and CEO of IRIS.TV. We’re going to talk about some use cases, how we’re using machine learning in some live production environments, and how USA Today and Gannett used this around the Olympics, and some kind of great applications of AI in use cases around video programming. I’m here with Kara Chiles, who is the Senior Product Manager at USA Today and Gannett, and she has an amazing background in media and content, ranging from being a reporter to a product strategist, and a number of great companies, including WebMD and AOL, Whole Foods, and then USA Today for many years.
So, a little bit of background. We’re going to really focus on how publishers can regain control of audiences. This is a topic that’s particularly interesting for many publishers today, in terms of how they can utilize social media and marketing channels to grow their audiences. So we’re going to focus on video, and how AI machine learning can be used to really drive contextually relevant programming to audiences, and even leverage breaking and day-to-day content to people whom it’s relevant to when you have a high load of content being distributed and published. And IRIS.TV’s background, we’re a personalization and programming platform; we’ll show a quick video on us. And the work that we did with Gannett was very focused on creating a better user experience, and then deriving data and insight that would be actionable to a number of different teams across the organization, so this includes editorial, product, revenue teams, and management.
And so, with that, Kara’s going to be talking about how they utilize our platform around the Olympics specifically, and it’s a great resource to have her here because it’s a really exciting use case. I learned more about it talking to her, and so this will be a lot of fun. And just handing it off to Kara, I’m going to brag a little bit about them. They just won three Pulitzers in the last year. We know that’s a topic everyone is thinking about when it comes to fake news or false information, and really, this is the companies that are investing in media and journalism and factual information, and so it’s great to see them seeing great success.
Kara Chiles, Senior Director of Product Management, USA TODAY Networks
So, one of the things we’ll talk about, though, as well, and it’s particularly germane for this group, is a lot of people, when they hear “video,” they’re thinking about what they’re watching, and not just the platform that it’s sitting on. This group obviously is a little bit different, and part of where we’re taking the product strategy for USA Today Network, and the 108 local market sites that we support, as well as the USA Today flagship, is really around thinking about product lifecycle, as well as that product experience to meet a lot of different users at a lot of different communities and levels in a viewing experience context.
For us, part of what that involved was rolling out an enterprise-wide proprietary player that would allow us to distribute video across all of these different markets, some to very small, like a Fort Collins, Colorado, to a very large, like the Arizona Central in Phoenix. We’re equally looking at solutions right now, and we’re all in the media space feeling the challenge of the changing browser environment, and so really looking at ways that we can break loose from an “autoplay with sound on” default experience that we know is not a customer favorite, to other ways that we can use that experience to both capture and retain users, as well as satisfy advertisers. And then we’re also looking at ways that we’re going to continue to grow, just as everyone else is experiencing, mobile web as a primary part of our viewership and our video experience across the board.
In terms of that, what that means for us in terms of programming and personalization, what we’re doing right now is we’re moving from a very desktop-intensive, plug it in, promote it across the network footprint to one that is really moving in the direction of test and learn. While this is something that a lot of product management organizations embrace, it’s still, I would say, in its very early stages at USA Today Network, because part of what we’re looking at is really using this extended market audience that we have to do a lot of different types of testing, and it’s teaching us more about our video network, and how best to adapt that product experience based on those markets and those appetites.
As part of that, as well, we’re really looking at, how do we actually take this and create new revenue streams? How do we actually use franchises, both from a content perspective and a day-parting or week-parting perspective, to create different ways to approach content and get it in front of audiences in the right time and place? And then, as I mentioned, using those audiences, one of the things that we see is the ability to recognize that we went from a waterfall model, where what we delivered and developed for USA Today was really sort of trickled down to the rest of the markets, to recognizing that local markets and audiences have overlapping, but also distinctly different needs and interests. So, as we develop our video product capabilities, how are we thinking about some of those unique aspects?
So, again, that is really a great use case of this. I’m just going to quickly cover the high level with regards to how we are implementing machine learning in production environments. The first is creating a common data model, so this is something that, for example, Facebook, Google, Amazon, Netflix, they’re very good at doing this, but traditional media companies have been a little bit slow to adopt. The second component, which the video touched on, is how our APIs implement into video players to actually capture audience behavior. And lastly, how you essentially develop insight and get it to teams in an actionable way.
So, the first component here we’re going to talk about is metadata ingestion and taxonomy creation, and then we’ll move through those other steps. The major takeaway here in terms of creating a common data model is, every business is unique, so if you’re a broadcaster, a news publisher, you may have some common ground with other news publishers in best practice you can adopt, but you’re also going to have topics, like sections of your paper or areas that you have specific coverage. So, there’s a number of examples that Kara will talk about, but one is the Olympics, right, or other kinds of special series, and having a taxonomy and common data model around that makes the data model machine learning manipulable in the future, so you can actually structure business rules around it.
The second piece, the API integration, this is fairly straightforward, like most APIs, but IRIS.TV kind of sits on top of any player; in this use case, this is a Brightcove back end. And IRIS.TV is also able to implement API installation, so if you have your own video player, as Gannett does now, then it’s also easy to integrate. And finally, in terms of how this personalizes video, a little bit of background. Setting up instruction, creating, using NLP to create that contextual kind of relevant data on an asset level, sets up the first kind of machine learning system, as well as what device, time of day, other contextual information, and finally cohort analysis. So those three kinds of analysis sit under a business rules engine that allow an editorial team or product team like Gannett and USA Today to essentially control the machine learning. And Kara’s going to tell you more about how they used it in specific instances.
That’s right. So, with this in mind, I think one of the things that we were trying to set up here was really to be able to take advantage of what we saw as a major content, product, and advertising opportunity. So, with an event like the Olympics, one of the things we see is that it’s an international story, obviously. It has a national-level story arc, as well as one that could be individual to specific markets where, say, athletes may be from, or where there is a training center. And with all of those stories, we also know that sports is a unique highly engaged category, and it also, this event, in particular, draws in a lot of people who may not identify as sports junkies. So we really saw a lot of opportunities here to reach people who are highly engaged and very loyal, very casual, may not even be regular users of the site, but would definitely come in for the experience, and a lot of original content that would be getting created, and then advertisers really trying to reach that heightened volume around this global and centralizing event.
Out of this, then, there was a lot of coordinating complexity, and this is where, when you’re pulling together an event of this scale, one of the things we automatically look for on the product organization side is, how can we find efficiencies? Where do we find ways to optimize all aspects of this? When you think about a 14-hour time difference between Pyeongchang and DC, one of the things we had to factor is, who was going to turn on feeds at 5:00 in the morning? Or, in particular, when you get an email from one of our videographers from the ski slope at 5:30 in the morning about where is the video that they just thought was going to be featured, which happens. One of the things you find out is the complexity of layers here really is where you start to align on, how do we then get complexity out of the system? One of the ways we do that is actually looking at templates that will help us recirculate the content, and also how we can then make sure that the content flowing through the templates is being directed through intelligent means.
So, with that in mind, when we think about that, there’s also the added piece of editorial standards. Whether it’s our own network original content, if it’s third-party content, if it’s information that’s coming from our users, how are we maximizing all of those experiences and driving both user value, so they’re always seeing something fresh and new from constantly changing and evolving events, to also providing value for people who are trying to reach those viewers as advertisers? And so that’s where I think part of what we came out with was, by partnering with IRIS.TV on how we were going to surface video, both at a USA Today level as well as in all of our local markets, we were leveraging how we were going to surface that video and recirculate it to users, so they felt like they were always seeing something new in the experience.
Yeah, I think the most important takeaway in this case for a media company is the leap that occurs when you move from, say, just doing manual playlists, right, or a simple kind of algorithm that will always show the users the same thing, to something that’s learning. So, having that data and taking that away as an organization, whether it’s for, you know, what content should we produce more of, which is a use case that we work on all the time with your teams, and also actually being able to learn from all the users. So we see this a lot, where there are videos that are the best-performing assets on a client’s website, especially in news, where then that video actually has very few initial clicks. So it’s something that’s difficult to surface in the real estate of a web page, but if you’re able to learn from that, then you can take advantage of it, program it to, say, a first-time user who’s coming in, and actually help them turn into a new and loyal audience.
In particular, the screenshot you’re seeing on this particular element, this was a brand-new template that we were also creating as part of that effort. This was a video template; our previous version of this within our desktop experience was basically a single hero video with an ad unit on the page, a very simplistic, circa 2011 type of page template. In designing something that put more video above the fold, that was being served through a dynamic module, we were really acknowledging that if you have reached this page as a user, you’re all ready to consume more Olympic video, or would be more likely to consume it if it was presented to you in a way that was easily accessible. And so, not only just providing continuous play, but providing more of that visual prompt in this space was intended to leverage what we were also then implementing through the player with IRIS.TV.
Beyond that, one of the other things we were able to do using the lock-off capabilities to really super-serve some of our advertiser clients. So, beyond the day-to-day of trying to make programmatic run through the system, this was actually a unique case at this level where we actually had high-level direct-sold campaigns and were really able to provide a lot of value from that side of the operation as well. So, when we had our Farmers Insurance campaign, one of the unique aspects of this was the spot that was put on the front of these pre-rolls was six seconds, as opposed to a normal 15-second spot and the eternity of waiting to skip ad from a user experience. Knowing it was so mildly disruptive, and knowing that it was actually going to be flighted through all of this highly promoted content, we actually ended up with a very happy scenario of actually needing to throttle that campaign and make sure we didn’t deliver too early in the game’s cycle. So we really saw a lot of value in being able to create prominence, create capture and a qualified audience, and then be able to be very mindful about making it persistent across the course of the event that we were providing coverage on.
And then when we looked at the results out of this, there were a lot of things that made the Winter Olympics Games a win for us in terms of our blended approach, but when we saw 50% video lift, that was clear that video was one of the key aspects of this event that shone out of what we had created. And then the increase in user in-stream retention versus regular content, I think also speaks to this being a unique event, where we know that people, if you can bring them in a little bit, but provide more opportunities to consume, they will come with you and return with you. And I think one of the things that we also saw that was great about this was, our expectations were modest in comparison to, say, the Rio Summer Games, where you see a lot of people having more appetite, and especially in Rio, where your time zone challenges are less pronounced. We actually saw this Winter Games outpace Rio coverage on video several times throughout the cycle because of this type of presentation and concerted effort.
So when we think about that, one of the ways that we looked at how do we reproduce this in the future, part of it is thinking about, do we continue to iterate, and do we have the right product templates and experiences that allow video to elevate, to really make it feel like it’s the appropriate experience for a user, whether they’re consuming it on their phone, on their smart TV, on their desktop? How do we make sure that editorial has the ability and has the resources either to create the most compelling original content, or is partnering with the right outlets, and that they are feeding through our distribution channels in a way that will make sure that we are surfacing those elements that get the benefit of being the most completed videos, the ones that have the most viewability, and the ones with the longest lifespan in terms of continuing coverage?
Our revenue teams also, in the ability to prioritize content that they know will work for a specific audience. I think one of the key things here, too, is not to continue to be incessant in front of the user, but to be very smart about really cadencing out how we put that revenue in front of the user in a way that allows them to consume as much as they want without feeling like they are being tackled by the ad units in place.
And so, out of that, I think one of the things that we have been talking about with Field and the team with IRIS.TV is, when we think about reproducing this, one of the next events that’s going to be coming up for us, and in fact for a lot of us here in the room, at least as individuals, is midterm elections. So, the initial primaries in some states kick off today, and I think one of the things that we look at as a unique opportunity, again, for us, is this is an event that plays almost on the flip side of the way that we approach the Olympics. By that, I mean this is obviously a very nationally relevant story, but it has deep roots at the local level.
So how do we actually show that immediacy, how do we show the strength of our network, where we have real embedded reporters who know these communities intimately, who know the politicians and the issues on the ground? And then how do we elevate those to universal issues and concerns no matter where someone lives at a national level? And one of the ways that we can do that, and one of the ways that we’re exploring, is how do we actually do that through video, then? Where we see those common interests, where we see the types of videos that people are using as we start to capture that metadata, how do we surface that from a local to national and then back again way, so that we’re constantly keeping that moving picture of this story as it unfolds over the next several months?
Yeah, it’s just a great use case of how to deploy this technology in a really actionable way that really drives business results. The USA Today Network has over a hundred different properties. Our goal is to surface content, say, from Des Moines that’s trending, and then actually bring it back to a national audience, so that’s just another small example of things that we’re trying to achieve.
And I think we’re at our time, so I’ll just close up and say that really, there is a methodology for building large businesses online. There’s been a lot of work done by YouTube, even, with Content ID, and now Facebook has a new platform for detecting what content is there. So if you’re an intellectual property holder, if you produce content and it’s great content like Gannett’s, you can actually manage the windowing it, the distribution of it, and then utilize best practices, which is what companies like Facebook and YouTube do, to then drive great results in terms of a good user experience, and then leveraging the data to inform how you proceed and build the business.