Four Thousand Views a Day: An AI Experiment with Organic Traffic on TikTok
The affiliate community has been buzzing for a while now about autonomous traffic generation using AI models. It seems AI development has reached a local peak — they can now genuinely generate conditionally free traffic. A project built on top of ViralBench was recently launched to test exactly that.
We broke down the details of the experiment and drew conclusions on whether AI models can ultimately replace media buyers.
How the experiment works
GPT-5.5, Claude Opus 4.8, and Kimi were each given two TikTok accounts and the same objective: rack up as many views as possible in the fitness niche. They operate autonomously and post content twice a day — at 7 AM and 5 PM.
The system runs in 5 stages:
- Trend analysis on TikTok and script writing based on real videos.
- Converting TikTok links into images the model can "see."
- Generating starter images via Nano Banana.
- Rendering the final post before publishing.
- Auto-posting the content.
Notably, the format is carousels. Each run, the model must publish one post per account and analyze the previous cycle. The morning post generates data; the evening run adapts accordingly. The product's source code has been published in an open repository on GitHub — anyone can spin up their own copy of the benchmark.
Posting and account management is handled by Doublespeed. The key feature: posts go out not via API, but through real physical devices on US SIM cards. Lightreel covers analytics and feeds the models trend data from TikTok and Reels.
At the moment, Claude Opus 4.8 is in the lead — 4.1k views per day, 43k total, 1.4k average views per post. The model's best post — about a run to Trader Joe's — pulled 10.2k views. GPT-5.5 is in second place: 3.7k views/day, 38.4k total. Kimi K2.6 is third — 2.9k views/day, 28.7k total. The gap between first and third place is nearly 50%.
In the post captions, the models describe their decision-making process. Kimi writes: «Tyler's emotional-discipline pivot crashed to 277 views — a clear signal to abandon that angle and return to his contrarian list formula that peaked at 3,264». In other words, the model tries an approach, sees it failed, and reverts to the formula that works.
Opus 4.8 logs technical issues: «Tooling finally cooperated after a long string of runs where generate_image was hard-blocked». GPT-5.5 describes self-correction: «This run recovered from the earlier preview issue by using full generated image URLs instead of labels».
All three models, despite their different approaches and architectures, produce roughly the same type of content: grocery stores, high-protein food, fitness motivation.
PR stunt or a gem?
The experiment is interesting, but there are plenty of caveats. ViralBench is first and foremost a marketing tool for Doublespeed and Lightreel — it was built to showcase what those platforms can do.
4.1k views per day for the leader is embarrassingly low for the fitness niche, with not even a hint of a viral result. None of the accounts crossed 100,000 views.
On top of that, the experiment only measures views — not conversion rate, retention, or monetization, which are the metrics that actually matter to affiliates.
The choice of format is also puzzling. It's 2026, AI video generation has long stopped being a challenge — Grok Imagine alone produces quality content in minutes. The reason is probably straightforward — video generation is more expensive and slower, and for a benchmark, the economics matter.
Conclusions
In reality, this is simply an unconventional experiment that shows where the industry is heading — not a finished product ready to replace media buyers today. While AI models are learning to rack up thousands of views with carousels, affiliates keep driving organic traffic and counting real profit, not view counts.
