PGA Tour Tees Up A Revolution In Fan Experience With Generative AI

Deriving value from GenAI means customizing content for fans.

The PGA Tour is redefining the golf fan experience with the help of cutting-edge technology, leveraging cloud infrastructure, artificial intelligence, and machine learning to bring the game of golf closer to its audience.

The PGA Tour organizes a series of professional golf tournaments played across the U.S. through the year. Scott Gutterman is senior VP digital and broadcast technologies at PGA Tour which has been working with AWS since 2012 when it moved digital operations to the cloud. The partnership has developed into a formal partnership with AWS now the official cloud, AI, and ML partner of the Tour. AWS is also a crucial part of PGA Tours studio production facility.

Gutterman explains how the Tour delivers “Every Shot Live,” where fans can follow their favorite players from start to finish.

“Our mission is to always show as much golf as we can, to show the best players in the world. Every Shot Live is a digital experience where we deliver every single shot from every player, from the very first drive to the very last. It's quite a big undertaking, because you have 100 to 150 acres, 18 fields of play with 18 holes, and on every hole there's typically three groups of three. We have approximately 120 cameras and we follow every group [there are 24 in total] and switch the cameras as we go from place to place.”

The cloud infrastructure enables people based in Atlanta and in London to do the camera switching, the footage goes back to a set of encoders in the production facility and from there distributed to ESPN plus and to different rights holders around the world.

“It’s all delivered over AWS media connect, so it's all delivered over IP, not over satellite or anything. It's a pretty big production and it’s been successful so far.”

Storytelling With GenAI

“We've been working a lot with Anthropic’s Claude models, particularly Sonnet 3.5 and Haiku figuring how to use that Shot Link data and use that to generate stories for our fans.

“One example in the past year during the President's Cup, we used Anthropic’s Claude models, to generate round recaps after every match. We took the live data and generated a story out of that live data that our fans could read on PGATour.com, or on our apps. Those stories could be generated within minutes after a match completed.

“They were reviewed by our staff for accuracy and then they were published. They turned out to be the most read stories that we had on the site during the President's cup.”

Another generative AI driven project will launch with the upcoming Players Championship in March. Gutterman provides an example courtesy of Rory Mcllroy.

“If Rory Mcllroy hits a 385-yard drive, what we have on our website maps is data points that say shot one was 385 yards, shot two is 125 yards left to the hole. But that’s it. It tells you exactly what happened and what the distance is, but it doesn't give you context about the importance of this shot for Rory.

“What we've developed is called Narrative Commentary. Instead of the statistics it will say Rory McIlroy just hit the shot off the 18-tee box, 385 yards. That's the longest drive of the day on this hole. He has 125 yards left to go into the hole. He has a 50% chance of hitting the green based on his past performance.

“We'll create these shot narratives for evey shot that will be published in our TOURCAST experience, which is a digital twin of everything that happens on the course. It's like watching a video game version of everything on the course. Fans will be able to see that commentary when it goes out, and they'll be able to understand how important each next shot is. “

Talking Points is the logical extension of Narrative Commentary.

“We produce four streams for ESPN every week, and that has live commentary on it. We also have the streams that we generate for our linear partners at NBC, CBS and Golf Channel. Historically what we’ve done is once a shot lands, staff will write queries against the database to figure what each shot means and feed that data to commentators. What we're building now is a talking points tool, which will generate three to five talking points for every shot hit.”

Back to Rory Mcllroy.

“Commentators will have three talking points straight away so for example, Rory’s shot is the closest to the pin for the day. He has a 10% chance of making this 10-foot putt. And what the commentators will be able to do is look at that and then weave that those talking points into their commentary. We're really looking forward to being able to release that tool hopefully this year.”

Prompting Context

What have been Gutterman’s learnings from the development of these tools?

“What we've learned is how important prompt engineering is when you're using these models," he says. “The models themselves can tell you about the basics of golf. They can tell you what the definition of a birdie is, but they can't interpret the golf action at a moment in time. When we were building the narrative commentary tool it would fixate on certain stats that weren't important at that moment. So, for example, what you don't want is on the very first hole of a golf tournament is for that to be talked about as the longest drive of the day, because that's the only drive of the day. So, you have to say, do not report on the longest drive of the day until half the field is on the course.

“We have thousands of stats in golf so what we’ve built is what I like to call context services. We're going to give the point in time data, but we're also going to give that data some context so that the model understands how to interpret that data. Context would be, this is the first round don't report on the longest drive. You can compare this to his longest drive in the same hole last year. Those are the types of contexts that we need to give the prompt engineering at the front end before it goes into the model."

Gutterman also has to sanity check model output.

"Once the narrative commentary is generated, we go through set of validation services. The first step is to check that the data that we put in the same as the data that we get out. The other step that we take is to check that the narrative make sense. There's a method called cosine similarity that says yes, this narrative about a drive is within the bandwidth of how you talk about drives on a golf course. That gets developed by creating thousands of what we call gold set prompts.”

When considering how to derive greater value from generative AI and develop these tools over the 12 - 24 months, Gutterman emphasizes the centrality of golf fans to the process.

“We want every fan to be able to follow any player that they want,” he says. “If we give fans what they want everything follows that. Sponsors want to know that what they're sponsoring is being watched so if I can interest a fan in a player, then I know that that player is being followed, and so does the sponsor.”

“That's really what is going on in media and sports - customization of content. Technology has given all of us in media sports the ability to customize content as a fan wants to see it, wherever they want to consume it, and for whatever players they may want to watch.”

This article originally appeared on our sister site Computing.