Contextual targeting is getting a generative AI glow-up.
In the old days of contextual targeting, a media planner came up with a target audience and had basic segments (“fashionistas”) to choose from.
Now, a media planner can copy and paste a creative brief into a prompt field and watch a long list of potential URLs to target pop up.
If that experience feels reminiscent of ChatGPT, that’s because new contextual targeting tools by startups like Cognitiv and RTB House use generative pre-trained transformers (GPTs) to build contextual segments. Because the GPT approach builds up a big pool of contextual information and uses natural language processing technology to more precisely find more relevant ad placements for advertisers, these companies believe GPTs can address some of the problems that plague contextual targeting.
For instance, contextual offerings can be overly broad. “Often, you just click a box: I want articles about baseball,” said Aaron Andalman, Cognitiv’s co-founder and chief science officer.
Contextual can also be too narrow and lack nuance if it takes a keyword-based approach, said Jeremy Fain, Cognitiv’s co-founder and CEO. Typically, keyword blockers block all content that contains specific words. A blanket prohibition might block all mentions of “cheat,” regardless of context.
Cognitiv’s sentiment model makes it possible for a brand like Nike to block the word “cheat” when it’s used in a negative way, but keep “cheat” when it’s used in terms like “cheat sheet.” Or distinguish using the word “shoot” in the context of basketball vs. a gun.
Another company experimenting with generative AI to analyze websites and build contextual audience segments is RTB House, a European ad tech provider that focuses on retargeting and real-time bidding. To develop its audience-related products, RTB House launched PrimeAudience, a DMP and ad network spin-off, last June.
If a marketer wants to target people planning to visit Paris next July, PrimeAudience’s algorithm will find content about Paris in the summer (referring to the area’s weather and seasonal events, for example) and content about restaurants and hotels in the city, according to Mateusz Rumiński, VP of product at PrimeAudience.
How contextual targeting built with generative AI works
Marketers using either company’s tool start with a prompt. Say a marketer is looking for prospective buyers of Yeti cups.
Marketers using Cognitiv’s tool can copy and paste a creative brief or several paragraphs from an article about Yeti cups to use as their prompt. They can then modify their prompts and otherwise refine the deal until the results meet their criteria, Andalman said. For example, they can give a thumbs up or down to individual URLs. When they give a site a thumbs down, the tool removes not only that site but all similar sites.
Cognitiv’s contextual GPT can also layer custom models for advertisers on top of the embeddings to filter by criteria such as sentiment and inclusivity, Andalman said. Clients can specify their desired sentiment band, whether positive, neutral or negative. And the tool’s inclusivity feature rates pages on a scale of zero to five based on the Linguistic Society of America’s definition of inclusive language, which weeds out content that’s biased against particular groups.
In PrimeAudience’s audience creation tool, the marketer starts by describing who they’re looking for, such as people who are interested in buying Yeti cups. PrimeAudience finds articles that fit the description. Its algorithm classifies the articles as “Yes” or “Probably,” depending on its level of confidence that the article fits the prompt, according to Rumiński. In addition to the “Yes” or “Probably” classification, the tool includes an explanation of what each article is about and the likelihood that someone who reads that article is part of the marketer’s target audience.
The end result for marketers using both Cognitiv’s and PrimeAudience’s tools is a list of matching URLs that might encapsulate their target audience, such as people who are shopping for thermos mugs or who have an interest in outdoor activities.
Next, the marketer can target the specific URLs of these articles in a DSP. Both PrimeAudience and Cognitiv create private marketplace deals (PMPs) to pipe through the DSP to the SSPs. Cognitiv is integrated with Magnite, PubMatic and Xandr, while PrimeAudience’s supply-side partners include SSPs OpenX and Microsoft (formerly Xandr Curate) as well as publisher network Raptive.
PrimeAudience can also extend the audience by targeting users who visit other pages in the same domain, other domains in the same publisher group or – through technologies like PAAPI – other sites they visit across the open web. But it will only do that type of behavioral targeting based on the context of a page if a publisher allows it. For example, for a user who reads an article ranking the top 10 washing machines on CNN’s website, PrimeAudience might show washing machine ads on other news websites, Rumiński said.
Many publishers shy away from cross-site identification of individual users because of the risk of someone stealing the data and attaching it to a third-party identifier based on personally identifiable information, like an email address, Rumiński said. To ease publisher misgivings, PrimeAudience can use technologies like SharedID from Prebid, which assigns users pseudonymous IDs that only show a publisher domain.
“There are two elements of privacy here – the privacy of the user and the privacy of the publisher data,” he said. “We will not use the publisher data in any context they don’t want it to be used, and we don’t want to use that data if the user doesn’t allow us to do that.”
As for Cognitiv’s take on privacy, its cookieless and ID-less tech means “there are no privacy concerns,” the company said.
On the backend
Before they can use generative AI tools for contextual targeting, companies first have to build a content library. Cognitiv uses a web scraper to scrape the content of every URL it sees in its DSP, Andalman said. The company scrapes millions of pages a day and feeds that content into its database, which updates every 15 minutes.
Similarly, PrimeAudience builds a content library by looking at the URLs in its bid requests, which it views as a form of consent. “We wouldn’t get a bid request if we didn’t have consent from the user,” Rumiński said. PrimeAudience also excludes audiences, such as medical segments, that don’t meet the requirements of the EU’s Digital Services Act (DSA).
The companies then feed their content libraries into their AI models. Both companies use a mix of publicly available models – Cognitiv uses OpenAI and open-source large language models (LLMs), while PrimeAudience uses open-source LLMs – and their own in-house GPTs and deep learning models.
Behind the scenes, when a marketer enters a prompt, both companies’ AI models are communicating with their content libraries to build contextual segments. PrimeAudience’s algorithm analyzes all the content on a website, including the URL, title, metadata, headline, article text (including tone, phrasing and keywords) and images. Based on this holistic understanding of the page, PrimeAudience can assign a URL to a specific user group or contextual segment.
Cognitiv’s tool returns a numerical representation, called an embedding, of the prompt. The embedding’s associated numbers denote where the prompt text might be located in relation to other content in the library along multiple dimensions – hundreds or thousands, depending on the LLM.
Cognitiv’s GPT compares the prompt’s embedding to the embeddings of all the scraped websites it has on file to find the closest matches. A cosine function measures the distance between embeddings.
The closer together the embeddings are, the more similar they are – and the higher the relevancy score. If the prompt was a few paragraphs from a “New York Times article” about Yeti cups, a “USA Today” article on the same topic would have a high relevancy score.
Calculating relevancy involves a “complex calculus,” Andalman said. If a marketer enters a paragraph that starts with a sentence about Nike, the most relevant stories the tool generates will probably be about Nike. But “if the rest of the paragraph is talking about shoe comfort or something else, those will also get weighed.”
Both Cognitiv and PrimeAudience built their contextual targeting solutions in anticipation of the end of third-party cookies. Rumiński declined to name clients but said a few brands are running tests and expects adoption to accelerate in the coming months. Cognitiv has seen interest from CPG brands, Andalman said. The Olympics also present an opportunity if consumers are searching for current articles about an athlete’s win or backstory.
Even for advertisers who understand what kind of content an audience reads, targeting that audience via contextual signals “can be tricky to get right with keywords,” he said. GPT-powered targeting tools “can make it a lot easier, quicker and more transparent that you’re getting the content you want.”
Correction: A previous version of this article described the PrimeAudience contextual targeting audience tool as a product of RTB House. Though part of RTB House Group, PrimeAudience is a company spun off from RTB House.
