Generative AI, radical personalization, and personal value functions

Product personalization is a commanding theme within the digital economy. To my mind, two catalysts have emerged that not only amplify the opportunity inherent with product personalization but necessitate its use:

  1. The widespread availability of Generative AI tools, such as LLMs and text-to-image models. These tools render the large-scale content production needs of product personalization far more economically viable than they otherwise would be;
  2. The steady degradation of conversion signals for use in advertising campaign optimization through privacy restrictions. This limitation of inputs to targeting models reduces the capacity by which advertising platforms can personalize ads to known user behaviors, resulting in user acquisition campaigns delivering more heterogeneous cohorts of users to products.

In Why in-app personalization, not fingerprinting, is the future of post-ATT advertising, I argue that in-app personalization is a potential solution to the challenges faced by advertisers in the evolving digital privacy landscape. Because the new users acquired through digital advertising initiatives cannot be assumed to have been vetted and filtered by ad platforms to the same degree as was previously possible, advertisers must take care to segment users and curate their experiences within the product to ascertain which users are likely to be the most engaged. In other words, product developers must fulfill the function that was previously served by ad platforms: to gauge an individual’s affinity for the product.

One difference in approach — which represents the opportunity inherent in product personalization — is that the curation that was performed by ad platforms in facilitating user acquisition generally funneled users into a singular product experience. This was by necessity: given the nature of advertising conversion optimization, targeting ads to audiences by their presumed in-app preferences would have required the ad platform to make assumptions about user intent that may not have proven true, since an ad platform can’t influence the user experience of the product being advertised.

To understand the power of personalization, it’s important to consider what it aims to achieve: that each user’s preferences are satisfied by the product with perfect precision. This is an ambitious if unrealistic proposition. But the process for pursuing it is no different than what is already employed by product managers, albeit for large groups of users: optimizing the product for target KPIs. Some products do not segment the user base at all to do this; they merely institute the changes that result in the highest possible target KPI (as an example: Day 7 retention) for the entire user base. Other products might do this for more specific segments: providing different product experiences for different segments of users such that each segment expresses the highest possible value for eg. Day 7 retention. Taking this approach to its logical extreme, a product might institute a Bayesian Bandits mechanic that alters each user’s product experience to optimize for some KPI.

But what if a product could curate an individual’s experience not just from a pre-defined catalog of content, but generate content in real-time that best catered to that specific user’s expressed preferences? As a thought experiment: imagine three different video streaming services. Each streaming service features a “home page” feed that presents thumbnails of available content to the user.

  • Product A curates the home page universally, for all users, prioritizing content in a way that produces the highest possible “match rate” across the entirety of the user base. The content thumbnails for Product A are drawn from an existing library of video content;
  • Product B curates the home page for each user based on their observed preferences, prioritizing content in a way that produces the highest possible “match rate” for that specific user. The content thumbnails for Product B are drawn from an existing library of video content;
  • Product C curates the home page for each user based on their observed preferences, similar to Product B, but Product C generates the library of content from which the feed is drawn specifically for that user, in real-time, based on that user’s observed preferences.

A theoretical set of histograms of “content match rates” for 100 users within each of these products might look like the toy example visualization above. Because Product C not only curates its feed based on observed preferences but also produces the content catalog for each user on that same basis, its average match rate is the highest. In essence, Product C implements a personal value function: each user’s satisfaction with the product — possibly measured in retention or engagement — would be maximized as a result of content being both produced and prioritized to suit their tastes.

Product C’s distribution of match rates isn’t represented as a straight vertical line on the 100% value. This is because there are obvious frictions to implementing such a system:

  • The cold start dilemma prevents content from being curated to a user’s preferences before those preferences have been observed. A user’s preferences must be known to be accounted for: Product C would need to test various types of content for engagement with any given user before it could optimize content explicitly for that user;
  • Choices from a constrained set of options don’t necessarily represent absolute preferences. If given the option between a Political Thriller and a Romantic Comedy, a user might select the Comedy, but that doesn’t reveal much about that user’s favorite genre of movie — merely the selection they’d make when faced with a constrained choice. To discover the types of content that best match a user’s tastes, many different options would need to be exposed, across many different types of content;
  • It may not be realistic to generate enough data for every user to derive a personal value function. Deriving a personal value function could require a large number of content interactions. And the experimentation process required to derive that value function could conceivably result in such a poor and unpleasant experience that the user would have been better served by a broader and more aggregated value function.

The promise of Generative AI is that content production becomes commoditized to the point that users’ tastes can be accommodated in real-time. But as that extreme is approached, the availability of input data, rather than production capacity, may become the primary constraint in optimizing for the highest possible level of retention or engagement. What Generative AI offers is the ability to achieve radical personalization: tailoring content to the preferences of individuals.

By