TV networks embrace AI at the upfront to improve audience targeting and commerce

Sandwiched between performances by Little Big Town and Kelly Clarkson, generative AI shared the spotlight at Radio City Music Hall as NBCU execs showcased its upcoming slate of shows and various advertising updates. In recent months, NBCU has been beta-testing a platform with AI-generated audiences derived from identity-base signals, content and other context. While the topic was only briefly mentioned on a screen flashing a chart with various segments, NBCU chief data officer John Lee later spoke with Digiday about the process of coming up with more than 300 unique audience segments identified by AI.

Using large language models, NBCU trained its AI platform on every movie and show owned and created by NBCU along with hundreds of links to web-based content. After having the AI write summary descriptions for every piece, it trained the LLM on data from academic research about behavioral science, human emotion and motivation theories. The team then had the AI assimilate a system and mathematically map the content descriptions with the academic materials to score similarities between human emotions and the episodes of each show. 

“Why it represents a big breakthrough is it’s just really hard to manage all of the data, or even collecting all the data associated with all of that content,” Lee said. “…We’re talking about it at the episode level and in some cases as we go forward even storylines within an episode. It will be the sort of thing we are focused on to make this targeting even more finely grained.”

In an alpha phase last year, NBCU took a handful of clients — across travel, entertainment, auto and quick service restaurants — to test the AI-generated audience segments against existing established machine learning models for performance-based campaigns. According to Lee, the AI-generated audiences led to between 22% and 46% more sales than the existing machine learning models.

NBCU is just one of the companies that touted new AI features in the 2024 Upfronts. Earlier this week at Disney’s upfront, Rita Ferro, Disney’s global head of advertising, said Disney’s Magic Words advertising product “ties mood to messaging.” NBCU and Disney first mentioned their new features earlier this year.

“The technology analyzes scenes across Disney’s vast library, and then serves advertising around that specific emotion or cultural touchpoint,” Ferro said. “It’s just one example – there’s so much more innovation to come. All viewer-first and all allows for new products and action while they stream using the remote or phone.”

AI-enabled tools for everything from contextual targeting to shoppable ads also came up during upfronts hosted by YouTube, Amazon and Warner Bros Discovery. Using generative AI for audience targeting also came a few weeks ago in pitches from digital media companies during the 2024 Newfronts. Google has a new generative AI feature for finding adjacent audiences advertisers might not have thought to show ads to. Others that announced new AI audience tools at the Newfronts include Canela, Samsung and Samba TV.

There are a lot of other steps that go into creating generative AI tools for audience targeting — from cleaning and labeling data to finding ways to scale insights not just from language models but also visual AI models that analyze the content. There are also safeguards to put in place for privacy protections and structures to make sure the AI models don’t hallucinate and provide inaccurate results or create the wrong associations for audience segments. Companies also need to check for biases in both the data and the large language models. For example, a TV series from the 1990s in a network’s content portfolio might have different biases that are no longer accurate or appropriate. 

“The sequence by which something happens, the order in which something happens is super critical when you want to do this in a fine-grained context,” said Vyas Sekar, chief data officer at Conviva. “That’s the hard part. Machine Learning models can inspect the scene and tell you there are yellow sneakers from Nike on a screen, but you need to see that in the context of whether you’re watching a sports game or whether you’re watching some video and what you did before.”

There are still plenty of risks. In a separate report from Forrester published this month, 60% of U.S. marketing and advertising decision-leaders surveyed said they were concerned using generative AI in their marketing could create false or inaccurate information. However, the research firm also noted that some marketers using generative AI tools for ads are seeing higher click-through rates and lower acquisition costs.

Beyond AI-generated CTV audience segments, CTV ad creative could also become powered by AI video models like OpenAI’s Sora and Google’s newly debuted Veo that debuted this week at Google I/O. The production capabilities aren’t yet ready for prime time on addressable TV, said Forrester marketing analyst Jay Pattisall. However, he added that could change fast in the industry’s current pace of innovation. According to Forrester, marketing investments in generative AI could total $23.4 billion by the end of 2024 and reach $28 billion by 2025.

“Once genAI video technologies come out of beta they will easily power bespoke, brand specific models to execute in CTV,” Pattisall said. “The concept of Brand AI Models is to execute campaigns leveraging audience data and inventory data from historical campaigns to produce content for the right audience, right channel, conveying the right message.”