A recurring theme in Nielsen’s monthly viewing figures is the rising prominence of YouTube on TV screens, with the Google-owned video service now accounting for around 13 percent of total TV viewing time in the US. And with the sheer volume of content on YouTube increasing by the day, advertisers are faced with an ever-growing list of channels and videos where their ads might be running – whether they want them to or not.
To help manage those ad placements, Pixability last month launched ‘pixie’, an agentic AI solution for curating YouTube campaigns. The product builds on PixabilityONE, the brand suitability firm’s buying and activation platform for managing spend through Google Ads and DV360. But with 1,200 different signals used by the platform to categorise inventory, across more than 30 million YouTube channels, sifting through all that data can prove time-intensive for advertisers.
Designed to simplify this process for advertisers, pixie is built as an agentic layer on top of PixabilityONE, with AI agents operating each of Pixability’s machine learning (ML) inventory categorisation tools, and an orchestrator agent that automatically selects the best tools to use for each campaign.
“We essentially have a tool for every one of those ways that you can curate inventory on YouTube, and we’ve trained that tool with our data to say this is how to get the best out of YouTube,” Jackie Swansburg Paulino, Chief Product Officer at Pixability, tells VideoWeek. “And it would take months to go through all those channels. We roll up the data so it’s easy for you to digest. Do I have any Made for Kids content here? Am I running only in the language I’m trying to reach? What content categories am I running in? Going through that manually is not impossible, if you threw a whole team at it, but it would take a lot of time.”
Minimising risk
Advertisers using Pixability can curate inventory based on 600 content types, using the IAB taxonomy, such as sports or news content; the Pix TV taxonomy, which covers premium, TV-like content; and audience-based behaviours, powered by Comscore data. Customers can also search for specific channels or creators, enabling them to find content similar to their brand’s channel, or similar channels to those that are already performing well.
“We also analyse sentiment of the video, so if you want to run around competitor brands where there’s negative sentiment, or you want to avoid anyone who’s talking negatively about your brand, we can remove that as well,” adds Paulino. “And we find that there’s a correlation between suitability and performance, and recommend most people run only on channels that have over 5,000 subscribers. But maybe I want a higher subscriber threshold, and we pull out those insights as well. So you can say, I want to make sure I’m running on channels that will scale well or perform well, that typically have a high view rate, or are viewed on TV screens.”
Then there are brand suitability filters, allowing brands to avoid certain types of content they feel might be damaging to their reputation. Pixability has 15 ML models to categorise potentially unsuitable content, such as one built to find videos that mention terrorism.
The platform then categorises this content by risk profile. Using terrorism as an example, low-risk content could be fact-based news talking about a terrorist attack, with media watchdog Ad Fontes assessing news channels for fact-based reporting and to weed out bias; medium-risk might be breaking news coverage of an attack, or video game, TV or film content that features terrorism; and high-risk content are videos “glamorising” terrorism, such as live-streamed footage or content owned by terrorist organisations. Paulino notes that these videos, including hate speech content, are removed from Pixability’s platform.
“There isn’t much of it on YouTube,” she adds. “YouTube has done a good job of demonetising and removing content that most advertisers find unsuitable. But we can be very strict with our rules, since we only work with advertisers, and so we can remove that content.”
Drawing the line
There is however borderline content that falls short of hosting extreme views but may still prove politically sensitive, such as podcasts and commentators discussing cultural and political topics, with ‘manosphere’ content being a widely discussed example. Paulino says most of these edge cases are submitted for human review, since they can be “difficult to solve with machine learning.”
But the level of tolerance for the more nefarious types of content varies depending on the brand and their objectives, according to Paulino. “We definitely see varying levels between those brands who just want the best performance and don’t care where their ads run, to those that only want to run on content they would find on TV screens, premium inventory, and a very select number of channels.”
And for advertisers using Google’s tools alone, monitoring every channel where their ads are running proves almost impossible. “If you’re running without an inclusion list, you’re probably running on hundreds of thousands of channels, or millions of channels depending on your budget,” says Paulino. “So it’s hard to analyse exactly which videos they’re running on. You might be running behavioural content, and you say, I want to reach people who are likely to buy a car next week. And that may be running in an environment that the brand just doesn’t know about.”
This is where the platforms’ different objectives come into play. While Pixability works solely with advertisers, YouTube is invested in its creators, making Google less likely to impose restrictions on monetisable content. This dynamic is evident in Google’s hands-off approach to labelling Made for Kids (MFK) content, leaving it to the content creators to ensure videos are properly labelled. Last month saw Disney agree a $10 million settlement with the Federal Trade Commission for failing to label certain YouTube videos as MFK, including content around Toy Story and Frozen.
“Google has to keep the creators happy,” comments Paulino. “And there’s a lot of borderline content. A 30-year-old man talking about Minecraft probably shouldn’t be labelled as MFK content. But from an advertiser point of view, again, if you’re trying to sell cars and those users are not logged in, the likelihood is they’re 13 or under. So we can label those content as likely MFK, where Google probably shouldn’t. The line is different.”
This assurance from Pixability helps advertisers invest safely in YouTube, which Paulino says garners a positive relationship between the two companies. “I think we help sell YouTube, and we’ve worked with them for over a decade,” she remarks. “We’re also in the YouTube Measurement Program, which verifies that our data does what says it does.”
And looking ahead, Paulino sees agentic technology playing an increasing role in YouTube campaigns, without eradicating human involvement. “There’s always going to be context that humans have, that agents don’t necessarily,” she says. “And from a product leader standpoint, building an agent that does 90 percent of the process, instead of trying to get to 100 percent, is easier to get adoption. People still want to be part of the process and feel like they’re collaborating with an agent, instead of feeling that it’s doing what they used to do.”
Follow VideoWeek on LinkedIn.
