Introduction: Background and Context

In the past year, ChatGPT and other generative artificial intelligence (AI) large language models (LLMs) have become ubiquitous. This explosive growth has led companies across the media and advertising ecosystem to look for applications of this remarkable technology. While the full story of generative AI is yet to be written, it is worth taking a close look at where we are - and what is next - in applying AI to media measurement.

The ability to measure and analyze audience size and engagement is paramount for advertisers, distributors, and content creators alike. Yet traditional methods of media measurement often fall short in capturing the nuances of today's fragmented media landscape.

On November 12, 2024, I joined the panel ‘AI: Transforming Measurement & Attribution’ at the inaugural CIMM West Summit in LA along with Erin Foxworthy of Snowflake, Gary Mittman of Kerv, and Evan Cohen of CIMM as moderator.

I shared Comscore’s POV on how the digital age has ushered in an era of unprecedented data generation, and how AI and machine learning (ML) are becoming critical tools in navigating this sea of information. From social media platforms to streaming services, the sheer volume and diversity of media content – and associated consumption data – demands advanced analytics to extract meaningful insights. Artificial intelligence systems can really level this up.

Non-Generative AI Applications

There’s a distinction between “traditional” non-generative AI applications and those using generative AI. The two areas are sometimes conflated, but in my mind, they are quite different.

Non-generative AI applications include classification and regression. Given a photograph, is it a cat or a building? This is an example of classification. Given a set of programs and their ratings, what would be the rating of some other program? This would be a simple example of regression.

Non-generative AI applications focus on enhancing the efficiency and accuracy of media measurement processes without creating new content. Examples at Comscore include:

• Personification: Our measurement systems use household-level observations of media consumption (notably, Linear TV and Connected TV viewership) and demographics to model which persons within the household are consuming the content. Because we know that houses don’t watch TV, people do.
• Anomaly detection: Comscore ingests a large amount of data from external sources. These data streams are all subject to potential problems, but an AI classification system helps to identify anomalies in the data, and which require intervention.
• Audience segmentation: Media consumption behavior over time can be used to identify various “clusters” of behavior, which can then be used for audience reporting.

Generative AI Applications

Generative AI, which is what most of the LLM fuss is about, produces content based on, for example, natural language prompts.

It is safe to say we are in the early days of using generative AI in media measurement. Current LLMs are subject to truly absurd “hallucinations” as well as a general blandness and lack of specificity – a regression-to-the-mean effect that is almost inevitable given the scale of the training data.

So where is Comscore using it now?

We have started to use generative AI in the context of code generation, where a data analyst describes a task in natural language and asks the AI system to produce appropriate code, for example in Python or SQL. Invariably, this requires iteration, either by refining the prompt (“prompt engineering”) or by editing the code that came out of the system. Time will tell if this workflow is more efficient than having an experienced programmer write the code directly.

A related innovation has been the measurement of generative AI-based search, which we started late last year in our qSearch product suite. Although this isn’t an application of generative AI (the measurement is done using traditional methodologies), it is a measurement of generative AI and shows that increasingly online search is done using this technology.

Future applications of generative AI from Comscore might include survey instrument design, ad-hoc data analyses, “story finder” applications within measurement data sets, and so forth. More to come!

Conclusion: Where Are We Going?

The history of generative AI has yet to be written. Issues around content ownership, data governance, privacy, deepfakes, job displacement, and even existential threats to humanity are much discussed in our industry and beyond. My guess is things will get worse on these fronts before they improve. But in the meantime, it is imperative for media measurement to look at responsible and productive ways to use this innovative technology.

To continue the conversation on AI and measurement with me, get in touch via this link.