He broke down 724 ads for 9 cents, then let 30 fake buyers judge them for 22 cents
The 40 second test cost less than a postage stamp, and points to cheap dry runs before brands pay for real ones.
The AI Daily Brief (Formerly The AI Breakdown): Artificial Intelligence News and Analysis · How People Are Actually Using JevMatthew Berman started with a pile that no sane marketing team would analyze by hand: 724 live ads from 37 brands.
He fed every one to Jev, an artificial intelligence model built by the company Typesafe. Jev does not write like a chatbot. It makes fast, snap judgments, the kind of work where you read something, make a small call, and take a predictable next step.
For each ad, Berman asked the same questions: what is the hook, the opening line or image meant to grab attention, what is the format, what is the offer, and what is the call to action. The idea was to compare those answers with actual ad performance, to see which patterns lined up with success in a way that would have taken huge manual effort to pull together before. The full breakdown took 40 seconds and cost nine cents in computing power.
Doing that kind of review by hand would have taken far too much time and effort for the payoff, but software can finish it quickly and for almost nothing by running the same few questions across the whole set.
Two days after that first test, Berman tried a stranger experiment. He took 723 of the ads and had Jev scroll past each one as 30 different buyer personalities, including a gym owner, a dental office manager, a toddler mom, and an AI curious engineer.
For every pairing of buyer and ad, Jev had to make one call: whether that shopper would pause or scroll past. That meant 21,690 stop or scroll decisions, delivered for 22 cents.
Those were not 30 real buyers. They were buyer archetypes imagined by an AI, so the results are better seen as a hypothesis generator than as data. It is hard not to think marketers will use this kind of dry run to try out ad ideas before spending money on live experiments.
Other users are already doing the same kind of archive work, running Jev over email histories, past social media posts, folders of documents, and databases of research projects by applying the same few questions to every item in the archive.
Questions people ask
How can analyzing hundreds of ads cost only cents?
Jev is designed only for small judgments, like which category something fits or yes versus no, and it asks the same questions about thousands of items at once. Because each judgment uses very few computing tokens, thousands of calls can add up to pennies rather than dollars.
Can this replace real ad testing?
Only partly. It can tell you how an ad breaks down by hook, format, offer, and call to action, and it can suggest which angles might appeal to which buyers. It cannot tell you what actually sold, so advertisers still need to compare its readouts with real performance and real customer tests.