10 min readLLMscreative testing
ChatGPT will pick your winning ad. It just picked both.
GPT-4 flips its verdict on identical ads 35% of the time. Sycophancy, non-determinism, and RLHF calibration collapse make LLMs unfit to pick creative winners.
The Splitroom blog
Creative testing, measured.
No listicles, no “10 tips.” Every number traces to a named source, and every piece answers its question in the first paragraph.
9 min readsynthetic audiencescreative testing
Synthetic panels agree with humans 85% of the time. They predict behavior 12% of the time.
85% aggregate preference agreement with humans (Zheng, NeurIPS 2023). 12% next-action accuracy on real shoppers (Lu, ACL 2026). Peer-reviewed evidence on what to trust.
9 min readMeta adscreative testing
There is no formula for a winning ad. Here's what 578,750 of them told researchers anyway.
Motion 2026: 5% of 578,750 Meta ads win. Meta: Reels ads 34.5% cheaper than image. Nielsen: 47% ad recall lift in the first 3 seconds. Here's the pattern.
9 min readMeta adscreative testing
95% of Meta creatives fail. Guess who pays to find out.
$1,383 in the US. $80 in India. $1,216 globally. That's the tuition per creative to reach Meta's Learning Phase verdict, and 95% never clear it.