An icon of an eye to tell to indicate you can view the content by clicking
Signal
Original article date: Sep 10, 2026

Generative AI Is Cutting Ad Production Costs — But the Harder Problem Is Picking Which Ad to Run

September 10, 2026
5 min read

Generative AI has made it dramatically cheaper to produce advertising creative. What it has not solved is the evaluation problem: once you can generate hundreds of design permutations in minutes, how do you decide, in milliseconds, which one to serve to a given user?

That question is the subject of a piece by Tom White, citing Peter Paton of Quantcast, that explores where generative AI is creating value in digital advertising — and where the real complexity lies.

The production side of the equation is largely resolved. GenAI tools can produce vast numbers of creative variants at a fraction of traditional production costs, a meaningful shift for an industry running billions of ad auctions each day.

The harder problem, as Paton frames it, is selection. Evaluating which creative performs best across billions of daily impressions requires real-time optimization systems — specifically, multi-armed bandit algorithms that balance exploration (testing new creatives) with exploitation (serving known performers). The scarcity of high-quality conversion signals makes this harder: not every impression results in a click or sale, so systems must draw reliable inferences from incomplete feedback.

The global digital ad market has surpassed $1 trillion, and the economics of even marginal improvements in creative selection are significant at that scale.

Key Takeaways

  • Creative production is no longer the bottleneck: GenAI reduces cost and time for ad design, but volume of options shifts the burden to selection systems.
  • Real-time optimization is the new differentiator: Multi-armed bandit methods are now central to competitive ad performance, especially as creative volume increases.
  • Signal scarcity constrains optimization quality: Limited conversion data means algorithms must make probabilistic decisions on thin feedback — a structural challenge GenAI does not address.

Note: This article is sponsored/partner content on Benzinga.

Read the full article on Benzinga