New AI Lets Marketers Test Ads on a Fake Internet First

A team of researchers published a paper describing OranSim, software that simulates a social media marketing campaign from start to finish. It models the whole chain: the creative (the ad itself), the creator (who posts it), the targeting (who sees it), and the budget (how much you spend). Fake consumers get exposed based on how well the content matches them and how the platform hands out attention. They then respond — and those responses ripple outward through 60 different population segments.

The clever part is making comparisons fair. Different candidate campaigns run on the exact same simulated crowd, using the same random numbers, so when you change one thing you can actually see what that change caused. Think of it like replaying the same day twice, with only one detail different. In one test, doubling the budget roughly doubled how many people were reached — but it also lowered how well the content matched those people and how likely they were to engage. Total simulated responses over 14 days rose to 1.96 times the baseline: nearly double, but not quite.

To keep the fake world realistic, the team trained a standard machine-learning predictor on 39,000 real posts from RedNote, a Chinese app similar to Instagram. It estimates how much engagement a post will get, scoring between 0.56 and 0.62 on a 0-to-1 accuracy scale, where 1 is perfect. A separate set of 12,154 posts tested whether it held up on future posts, unfamiliar creators, and less common topic areas. The code is publicly available for others to check and build on.

The honest catch: this is a simulation of a synthetic campaign, not a real ad buy. It can't capture everything — sudden algorithm changes, viral cultural moments, or what competitors do at the same time. And it learned from one platform's data, so its lessons may not transfer neatly to Instagram or TikTok. Still, it points toward a future where your first draft ad budget gets tested in software, not in your wallet.