After introducing a six-second “Bumper” ad format back in 2016, YouTube is unveiling a new tool that uses machine learning to automatically pull out a six-second version from a longer ad.
It might seem a little ridiculous to try to compress (say) a 90-second video into a six-second message. In fact, Debbie Weinstein, Google’s vice president of YouTube and video global solutions, acknowledged that there was some skepticism when Bumper ads were first announced, with advertisers wondering, “Can we actually tell our story in six seconds?”
However, Weinstein argued, “We learned over time that creatives love constraints. They’ve historically been constrained to 30 seconds, and then 15 seconds, and constrained by whatever dimensions of a particular media format.”
For some advertisers, she said, a Bumper may simply be a short teaser for a longer ad. For others, the format could provide a way to break down a 30-second ad into a sequence of six-second clips.
And with Bumper Machine — which YouTube is currently alpha testing, which will then lead into beta testing and eventually general availability — advertisers will have a tool to create a Bumper by scanning a longer ad for “key elements,” like a voice-over or a tight focus on human beings or logos or products. The result always ends with “the final call to action in the last two-to-three seconds of the video,” Weinstein said.
For example, as an early test, GrubHub took a 13-second ad and used Bumper Machine to create the six-second version below.
Weinstein suggested that Bumper Machine could be used by “many different advertisers of all shapes and sizes” — some of them might be smaller advertisers who want to create Bumpers with as little time and effort as possible, while larger brands and agencies may treat them as more of a “jumping off point,” which can be refined or serve as inspiration.
Either way, Weinstein isn’t expecting advertisers to just start posting machine-created Bumpers willy-nilly. The idea is to always have “some level of human review.”
“You’ll get three to four executions, the best guesses that the machine is going to make,” she said. “A human is going to go through and decide which of the three or four is best, or decide all of them are great, or do some light editing on top of that.”