Abandoning intuition: using Generative AI for advertising creative

Of the various enterprise use cases served by Generative AI, none is perhaps more immediately actionable than advertising creative production. The value proposition is obvious, given some qualities of digital advertising:

  • The expense of advertising creative production scales more or less linearly with advertising spend;
  • A large proportion, if not the majority, of advertising creatives are not useful or viable in the sense that they do not lead to profitable advertising outcomes when utilized in campaigns (and relatedly: the dollar value of wasted effort also scales more or less linearly with ad spend);
  • Creative asset utilization follows something like a Pareto rule, where perhaps 20% of the creatives produced account for 80% of all advertising spend (the skew can potentially even be more extreme than this).

Advertising creative is a critically important component of successful, scaled advertising, but the cost of production and testing can be immense. I describe a workable creative production process in detail in Mobile ad creative: how to produce and deploy advertising creative at scale, and I outline a creative testing framework in this QuantMar thread. Ultimately, the purpose of a creative production and testing process is to establish a production cadence such that new, viable creatives are available before active creatives reach the performance inflection point visualized in the below diagram.

Successfully achieving this cadence results in creatives being launched on a schedule that maintains some average level of satisfactory performance as older, degrading ad creatives are replaced with vetted alternatives. This replacement schedule is visualized below.

The ability of a Generative AI tool to contribute to an advertising creative production process is conceptually clear: tools like Stable Diffusion and DALL-E conjure unique, bespoke images from both text and image inputs. These tools, as described, fulfill the same purpose as entire advertising creative production teams: to render written prompts or concepts into pieces of ad creative. But in understanding how Generative AI functionality can be embedded into an advertising creative production workflow, it’s important to first consider the discrete steps of the production process. I conceptualize a creative production process as being comprised of three steps: Ideation, Production, and Analysis.

It’s important to recognize that this process is iterative and, ideally, accretive. But the notion of accretive optimization for a creative process can be deviously confounding. My sense is that teams tend to focus less on improvements to the process when gauging the efficacy of their creative production workflow than they do on the subjective qualities of their ad creatives, the demonstrable values of which are difficult if not impossible to ascertain. Examples of these qualities are: aesthetic styles, product placements, narrative styles, themes, the chronology of content (for video), etc.

Put a different way: creative teams tend to overestimate their ability to understand why specific ad creatives are performant. This results in teams becoming anchored in feedback loops that might not result in the best possible outcome for their advertising campaigns: the teams decompose performant advertising creatives on the basis of the qualities they can intuitively discern, and they produce more creative that acknowledges those qualities.

The problem with this approach: I believe that creative teams are not very adept at recognizing the aspects of an ad that impel some reaction (for example, a click), either because they cannot fathom all possible qualities that led to that reaction or they simply approach the exercise with inherent and tacit bias. I often see creative teams declare with forceful certitude that some creative performed well for a specific reason. But that is mostly hubris. More often than not, when I see a creative that outperformed its peers in a test, I am dumbfounded as to why. To my mind, the appropriate course of action when a specific ad creative results in exceptional performance is to reinforce the production process from which it was generated — not to attempt to understand why that creative performed well.

To that end, I believe that the value that Generative AI tools like Stable Diffusion bring to bear in the creative production process is realized less by replacing the mechanical human efforts related to asset creation — such as illustration — and more in obviating the risks that human biases present in determining which specific pieces of creative outperform others. Yes, tools like Stable Diffusion are helpful for materializing creative assets into existence with just a text or image input, and that capability will indeed result in production cost savings for the creative teams that adopt them. But more crucially: these tools can create value by untethering the ideation process from the somewhat arbitrary qualities of advertising creative that teams believe to be responsible for success.

It’s therefore important to pinpoint where in the production process Generative AI tools can be deployed. Obviously, these tools will be used in production: illustration, footage procurement, asset re-sizing, etc. can all be managed by these tools. This will result in substantial cost savings: a team might need just one or two artists or designers to work with the output of a Generative AI model compared to the army of artists and designers needed to produce dozens or even hundreds of pieces of advertising creative on a weekly basis. There exists an obvious, undeniable use case for creative asset production with Generative AI tools.

But I believe that the ideation step in the process is served more consequentially with Generative AI. Competitor analysis is a perfect example: rather than presuming why a competitor’s ads have been successful, a team might use the textual inversion process explained in this video to feed Stable Diffusion with competitor assets in order to produce ad creative variants that can be deployed with minimal editing. The same approach can be used with owned assets: feed them to the Stable Diffusion tool and allow the tool to create variants. Again, the production capabilities of Stable Diffusion reduce cost, but the ideation capabilities deliver value: not only is an artist not needed to produce visuals, but neither is intuition or gut feeling around “why” specific pieces of creative performed well historically or for competitors.

The middle and right columns in the diagram above capture cost, but the leftmost column represents value. Utilizing a Generative AI tool in both aspects of the process — especially in a way that can be automated — is where the substantive opportunity exists in applying this technology to advertising creative production.

A cynical interpretation of this approach might assume that genuine creativity plays no role in the production of advertising creatives; that the production process is simply an exercise in procedurally parsing existing creatives for persuasive power and generating variants from those components. But the reality is that the introduction of Generative AI to the process imbues real creativity with a profound premium. The fast-follow process and the tools that enhance it, like Generative AI, are good at quickly enabling convergence around the effective elements of performant ad creative. But wholly new, blank-canvas creatives are best produced by humans, and the ability for a creative team to generate those creatives ex nihilo will confer a competitive advantage.

Fast-following will always necessarily remain one production cycle behind genuinely new ideas, and the performance delta between genuinely novel, unique creative and fast-follow convergence likely widens with the broad-based adoption of Generative AI. The configuration I describe here simply assigns different parties and mechanisms with the tasks they can best complete: to humans, creative ingenuity; to machines, deciphering patterns from a complex mosaic of information.