The underlying deep learning capabilities of ChatGPT allow it to discern linguistic nuance, which has the potential to transform media planning and buying. Deep learning capabilities will improve how advertisers select inventory for bidding, both contextually and behaviourally.

ChatGPT is nothing short of a cultural phenomenon. It set a record as the fastest app ever to reach 100 million users. Its ability to understand information means that it can be used for a huge variety of applications, from writing an email to scanning code for errors.

For digital advertising, the true potential of ChatGPT is not the app itself, but in its underlying deep-learning capabilities. ChatGPT has demonstrated that deep learning systems can acquire a breadth and depth of knowledge that was previously considered solely the domain of humans. This base of knowledge allowed humans to have powerful intuitions about how to market and advertise effectively, that machines were unable to replicate. The newfound ability of machines to understand language, culture, and behavioral patterns, means this technology is now poised to revolutionize how advertisers approach media buying. More specifically, deep learning will transform how advertisers select inventory for bidding, both contextually and behaviorally.

Context in full color

Contextual targeting today is almost exclusively based on URLs, keywords, categories and sets of relatively simple rules. Many webpages are categorized at the URL level or one or two layers below that based on meta-data and subsections of a website’s architecture. A blog might have a “health and fitness” page, for example, that gets labeled in that specific bucket. Other pages are categorized based on the words on the page. For example, a page that has words like “recipe” and “sandwich” will be placed in a “cooking” category.

However, contextual targeting doesn’t pick up on true context. What’s obvious to a human is not so clear to a simple rule-based categorization system. That blog might be a joke blog, with a “health and fitness” page that’s filled with jokes about being lazy. A page that says “recipe for disaster” and “knuckle sandwich” should not be placed in a “cooking” bucket, although it’s quite possible to do so using common contextual targeting rules.

Deep learning far surpasses today’s contextual targeting. The AI behind ChatGPT could easily pick up on the nuance of the language to determine that the blog is a joke, not a real health site. It can also discern that a “knuckle sandwich” has nothing to do with cooking. But that’s just the beginning. GPT doesn’t need to rely on buckets and categories. It’s perfectly comfortable with nuance.

Take an example of an article written by a scientist who happens to be a baking enthusiast. The article explains that she had an epiphany one day while baking bread. While the article includes a mention of baking bread, its content is about scientific discovery. The AI would understand the merging of multiple topics, but that the real topic is science. Perhaps it would score it as a 3 out of 10 for “cooking” and a 9 out of 10 for “science and technology.” But, even that is simplistic because ChatGPT can do this for a thousand or more similar categories. It could also determine the education level of the ideal reader, the level of positive sentiment, and how inclusive the article is, for example.

With deep learning like GPT, contextual targeting flips from black-and-white to full spectrum color. Not only is the categorization much more accurate, it has much more depth. An advertiser will be able to dial their contextual targeting up or down to achieve more or less precision or scale across a host of different important attributes. Perhaps an advertiser wants at least a 7 out of 10 for “science and technology” with at least a 5 or more for “positive sentiment.” Advertisers could also negatively target things like scientific propaganda or politically driven skepticism. By refining this targeting, advertisers will be able to understand which attributes drive better outcomes and continue to improve their approach.

Understanding human behavior

Behavioral targeting will be similarly reimagined. Today, advertisers buy audience segments that are in buckets based on a set of rules. Those rules are like a mirror image of today’s contextual targeting. Rather than the buckets reflecting the content on the page, they reflect the people who visited those pages. Someone who visits a cooking site will be put in a bucket for cooking enthusiasts.

Again, AI will bring a new level of depth and detail to the process. GPT is a “large language model.” It understands very long complex sequences of words. There could just as easily be a “large online behavior model” which understands sequences of online behaviors. Perhaps someone visited a string of different sites that indicates that they don’t just like cooking, they’re likely to buy an air fryer sometime soon. Someone else who visited a slightly different group of sites may be less likely to buy an air fryer, and could score lower for an advertiser that sells them.

This kind of AI-based marketing is gaining traction on retail sites using first-party data and shopping behavior, but could be extrapolated across advertising channels to gain insights earlier in the purchase process.

Don’t be afraid to try

ChatGPT has been a fun tool to test out, but few media buyers would jump at the idea of using it to help with media planning based on their experience thus far. Many of us have asked it to write an email or blog post, or find information online, with mixed results. However, our expectations for ChatGPT are that it will be better than the best human writer or best researcher we’ve ever worked with. That’s a very high bar.

In the case of having GPT AI improve our ad targeting, we don’t need it to operate at a near-perfect genius level. As long as it improves upon the simple rules-based approach we have today, it’s a highly valuable addition – and it’s doing so at a scale that no human could hope to replicate.

Not only can we get value from AI today, we can provide it with feedback that will train it to get even better over time. There’s no harm in trying it out. If we do, we’re bound to learn and get better.