Generative AI can now produce dozens of versions of an advertisement in minutes. The harder question is deciding which versions are worth showing to customers.
That is becoming a bigger challenge as advertisers create more content with AI. An Interactive Advertising Bureau report found that generative AI was used to create or modify 22% of digital video ads in 2024. Buyers expected that share to reach 39% in 2026.
Kiran Kumar Manku has spent more than 11 years working on the engineering systems behind digital advertising. His work has focused on the technology that decides which ads run, how they are delivered and how companies test new formats.
Manku is also the author of “Distributed Systems Anti-Patterns: How Production Systems Actually Fail,” a book about the difference between how large systems are designed and how they behave under real traffic.
He said advertising teams often have plenty of ideas but not enough engineering time to test them.
When one test takes months
Testing a new type of advertisement once required much more than uploading a different image or changing a headline. A new format could require custom code, new validation rules and changes to the system used to deliver it.
“We would sit in planning and someone would say, ‘What if we tried this?’ and everyone knew it was a good idea,” Manku said. “Then someone would say it’s a quarter of work. So it went on a list. Most of that list never got built.”
That process limited how many experiments a team could run. Instead of testing several ideas and comparing the results, teams sometimes had to select a few large projects and hope they worked.
Manku said the problem was that advertising systems treated each format as an entirely new product, even when many of the underlying parts were the same.
His approach was to build a framework that allowed teams to describe a new format through configuration instead of writing a separate system for it.
“Nobody gets excited about a configuration schema, but that schema was the whole thing,” Manku said. “If it was too narrow, teams would go back to custom builds. If it was too broad, it stopped being safe to change.”
According to Manku, the framework reduced some development timelines from months to weeks and was eventually used by several advertising teams.
AI makes the testing problem bigger
Generative AI has made it easier to create new images, videos and versions of ad copy. It has not automatically made those ads effective.
“People think the hard part of AI creative is the creative,” Manku said. “It isn’t. I can get 100 versions of a banner before lunch. The question is which three to keep and whether I can find that out this week.”
Testing systems must be able to handle that added volume. Otherwise, AI may simply create a larger backlog of ads waiting to be reviewed.
A useful system also needs to compare results consistently. If separate teams use different testing methods, it can be difficult to tell whether one version actually performed better or whether the test itself produced a different result.
Manku describes this as a knowledge problem. A company needs to remember what it tested, how the test was conducted and what happened afterward.
“A test you can’t run again is an anecdote,” he said. “The point was never to go faster for its own sake. It was so the second team didn’t have to relearn what the first one already paid for.”
Manku’s broader research includes work presented at KDD 2026, a data science and artificial intelligence conference held in Jeju, South Korea.
Keeping people involved
As AI becomes more common in advertising, Manku said companies still need people to review results and understand why a system made a recommendation.
That includes checking whether an ad follows brand standards, whether a test was fair and whether the results can be explained. A system that simply selects a winner without showing how it reached that decision may save time, but it provides little information for the next campaign.
“Everyone is going to have more ad variations,” Manku said. “What’s left is knowing what happened and why one thing beat another.”
The ability to generate ads may no longer be the biggest technical barrier. The next challenge is building testing systems that can keep up, preserve what teams learn and give people enough information to make the final decision.