Amazon’s Neal Richter on AI’s impact on advertising

Amazon is really good at selling stuff online. Now, as its programmatic ad tools continue to grow, it’s eyeing the rest of the internet.

Marketing Brew spoke with Neal Richter, director of advertising science at Amazon DSP, the company’s ad-tech tool that merges its marketplace data with the rest of the internet. We chatted about which advertisers make sense for the DSP, industry automation, and what the heck AI has to do with programmatic advertising.

This interview has been lightly edited and condensed for clarity.

Help me understand how Amazon’s DSP fits into the marketplace. How much differentiation is there? Don’t all the pipes go to the same place?

When it comes to connectivity of supply-side systems, in general, what you’ve said is true. Most publishers have connections with more than one supply-side platform or ad server, so you can access your inventory through those systems. As a buyer, you can access inventory through a bunch of different mechanisms. In that sense, access to inventory isn’t that much of a differentiator for a buyer, for a bidder, anymore, a DSP.

What is a differentiator is kind of the models they use, and what those models predict. What do the models predict, and how do they perform for customers given the intention of the campaigns?

Amazon’s marketplace is sitting on a lot of data. Does that data inform your models?

Yeah. The signals that can be derived from shopping activity [have] pretty high utility in predicting purchase propensity and the types and combinations of products that shoppers are interested in.

Does Amazon’s DSP make sense for an advertiser that isn’t using Amazon’s marketplace?

We believe so, because when you’re leveraging Amazon’s DSP, you’re using models that are informed by the kinds of shopping activity of a very broad selection of people—humans—that buy things and get boxes delivered to their doorstep.

We believe that our models are very predictive in that sense and uncover a lot of basic purchase intent of consumers.

Amazon is using AI in everything from delivery vehicles to the warehouse. How is it being integrated within the DSP?

The most obvious use of AI and machine learning is modeling the probability of interest in a product, the probability of consideration of the product, which would be some degree of serious consideration, and the probability of purchase. I think those are the basic bones of what a good ad system should do.

What we also care about is helping advertisers build good media plans that help them reach relevant audiences that are going to consider and purchase their products. We also care about helping advertisers create great creative, and we’re working on those things as well.

We see AI in some ways as controllable automation to help advertisers be more efficient at creating good campaigns.

Language models aren’t new to ad tech. What’s different this time?

I think it’s different because, [today], large language models are able to extract and summarize content and context in a different way than before. Prior mechanisms of understanding context were more about keywords, or coarse-grained descriptors that happened to occur on the page, in the content they’re viewing versus really understanding the content.

Large language models are much better able to summarize the content that a user is consuming in a way that’s superior to just a bag of keywords or a bag of phrases. That approach was the n-gram approach. If you look at really early versions of statistical machine learning for natural language processing, n-grams were a very popular mechanism.

With large language models, we’re in a different space where we’re really talking about models that understand what the content is about.

Great, now I have to Google what an n-gram is. How would you explain it to your parents?

[laughs] They’re like two, three, four, five, six, [or] seven word combinations that tend to represent content, like sentence fragments.

Where do you see this technology going?

If you’re a publisher, you now have an opportunity to use these emerging AI models to build an audience strategy. You will have better tools yourself to think about your content and how your audience consumes your content—watch, read, etc.,—and think about how to summarize those patterns of consumption and build audiences…I think there’s going to be a new wave of audience-building that’s more about affinity of content than viewership.

On the advertiser side, there’s a lot to do. One is concentrating more on the strategy of the campaigns, instead of having to work on a daily basis on optimizing the campaigns. The latest wave of machine learning and AI algorithms are much better able to optimize the campaigns to goals, and that’s a lot of what we’re investing in…Building great media plans is going to be much more automated, and building great campaigns is going to be much more automated.