WTF is vector-based ad targeting?

“I would love it as a vendor if [Meta] had vector-based buying. So instead of doing keywords, I actually want to buy this vector,” said Jon Morra, chief AI officer at brand suitability platform Zefr. He added, “I think that could be the future of agentic buying.”

WTF is vector-based ad targeting?

Vector-based ad targeting is taking a vector and—

Hold on. What’s a vector?

A vector is a numerical sequence, like a set of coordinates. You know how lat-long coordinates are a set of numbers separated by commas and representing a position in two-space? Vectors are like that, but instead of just two dimensions — latitude and longitude — they can contain thousands of dimensions.

What’s the point of these dimensions?

The more dimensions in the vector, the more precise and complete the record it represents. Case in point: (27.9881569, 86.9253667) is a lat-long coordinate that tells you the location of a place on a flat, two-dimensional map. Which is helpful. But it’d be even more helpful to have a third coordinate, or dimension, that represented the location’s elevation in case you didn’t recognize the lat-long location is Mount Everest.

What does this have to do with ad targeting?

Vectors, like coordinates, are ultimately representations of data. Lat-long coordinates represent physical locations; vectors can represent data including text, images, video. And similar to how coordinates represent proximity numerically — (28.0, 87.0) is a handful of miles from Everest — vectors can be used to establish relations between pieces of data that have been converted into vectors by effectively organizing the data into clusters. Which means a vector could be used as an ad targeting parameter, similar to how keywords and CRM databases have been used to set ad targets.

How does data get converted into vectors?

Through a process called vector embedding. Companies including Google and OpenAI have AI models that can take a piece of data, like an article text or a video, and convert it into vector coordinates, which can be stored in a vector database. 

And then what?

And then the vector coordinates can be queried like a traditional database.

But then you’d have to know the vectors to find something?

Not necessarily. The vector coordinates can be associated with a traditional database entry so that regular parameters can be used to look up a vector. Like how you can input a lat-long coordinate into Google Maps to find a location or just type its name.

Again, what does this have to do with ad targeting?

Right. So you know how in traditional ad targeting, an advertiser sets keywords or audience characteristics? “I want to target pet owners. Run my ads on web pages containing the word ‘cat.’ And use this database of registered pet owners to identify their characteristics — shopping habits, household type, etc. — and find people with shared characteristics.” 

Sure.

Well, that traditional ad targeting works. But it’s also kinda basic. Keyword-based targeting can be crude, which is why keyword blocklists get a bad rap for preventing ads from running on articles about mobile phones because a blocklist contained the word “mob.” And just because cat owners can over-index on buying seafood for their fur babies, so would pescatarians. These are oversimplifications, but the point is that vector embedding can enable way more sophisticated ad targeting.

How? 

For starters, vector embeddings can be used to create keyword clusters that represent their semantic meaning, not just the order of their characters. “Mob” would be given coordinates that plot it in the neighborhood of other organization types but also include dimensions that effectively cluster it around crime. Meanwhile, an audience profile of a known pet owner would put them in the neighborhood of other pet owners but also include them in clusters that may have no clear relation to pet owners, unless it turns out that a lot of pet owners’ vectors also appear in those other clusters. 

“The way it works is essentially you’re looking for similarities of the users. And so as long as you define it all the same – as long as you take the CRM and define the audiences, define a common identity, like the RampID – you are able to do that. So our view is that that becomes a way of targeting audiences in the future,” said Travis Clinger, chief connectivity and ecosystem officer at LiveRamp, in an interview earlier this year. 

LiveRamp developed the User Context Protocol, which has been donated to IAB Tech Lab, renamed Agentic Audiences and aims to provide a standard for AI agents to exchange vector embeddings as signals for advertising.

Disclaimer: The vectors would need to be created using the same embedding model in order for the relations to be established and maintained.

Huh?

Yeah, the multi-dimensionality of vector embedding starts to break the brain a bit. 

“Most things that you embed have many ideas. This is why you need many dimensions. It’s not only the royalty dimension, but it’s the pets dimension, it’s the weather dimension, it’s the day of the week – all these things encode different dimensions. So that’s why you need this high-dimensional space to represent it,” Morra said.

And this is why vector embeddings become better suited to AI agents, which are able to handle this multi-dimensional data. In fact, they already do. Technologies like semantic search and the transformers underpinning generative AI incorporate vector embeddings. 

Cool, but I’m a human.

Fair point. Let’s consider four food items: a beignet, a chimichanga, a churro and a pickle. 

All four foods are sold at Disneyland, so they would all share that single dimension and appear on the same flat plane. But a beignet, a chimichanga and a churro are fried foods, so they would move along a second axis to appear closer together on the plane. But then a beignet and churro are sweet treats, so they would share a coordinate along a third, sweet-to-savor axis.

The vectors allow for someone looking at the four food items in three-dimensional space to see that all four food items are in a given cluster, but the pickle is a bit farther apart and the beignet and churro are the closest together.

Again—

What does this have to do with ad targeting? Right. Well, because vector embeddings plot data in multi-dimensional space and establish relations among data points based on the proximities of their coordinates, an advertiser could specify one or more vectors as the seed target(s) and then set a radius to include the nearby vectors representing keywords, people or content as part of the campaign’s target.

Huh?

OK, so you know lookalike targeting? The idea of taking a known audience segment and targeting ads to people with shared characteristics? Vector-based ad targeting is like an advanced version of lookalike targeting. But it can also be used as an advanced version of holdout targeting – excluding certain parameters – by specifying conditions for filtering out dimensional clusters. And then, as data is updated and embeddings are recalculated, the vector coordinates can change to move closer to or farther from certain clusters. If these changes are tracked over time, trajectories can be established to forecast someone’s propensity to make a purchase, enabling a form of predictive targeting.

Like 4-D chess to traditional targeting’s checkers?

More like 4,000-D chess, but yeah.