We are Vincent and Jochen from Sitefire (https://sitefire.ai), with backgrounds in reinforcement learning/optimization at Stanford and software engineering at Technical University Munich. Sitefire (YC W26) helps marketing teams get recommended by AI search engines such as ChatGPT, Google AI Overviews, Claude, etc. Our platform monitors prompts, tracks which pages get cited, and suggests actions like creating YouTube videos or writing the right blog posts.
After talking to hundreds of marketing teams we learned that everyone despises “AI slop” yet still wants to use AI for content creation. We asked: what structural patterns define AI slop? Can we train a model to spot it without looking at the words?
Inspired by the University of Maryland and DeepMind paper StoryScope (Russell et al., 2026), we adapted their pipeline to commercial web pages. Using the Wayback Machine we gathered 2,250 blog posts from 268 B2B sites that were written before ChatGPT existed. For each post five large models (GPT‑5.4, Claude Sonnet 4.6, Gemini 3 Flash, DeepSeek V3.2, Kimi K2.5) generated their own version.
Instead of analyzing text, we asked an AI to answer 214 structural questions about every article – e.g., does it push its product repeatedly, does it back claims with sources, does it quote a named expert? These answers formed a feature vector that we used to train a binary classifier. On unseen blogs the classifier achieved 98% accuracy, misclassifying only 19 out of 1,740 posts.
The reason is that all five models produce a very similar shape. When we map each model’s feature values they cluster tightly, while human articles spread out and occupy a larger region. Among the 1% most unique posts, 149 are human and only 4 are AI. A key AI slop trait is the “tidy, self‑announcing” pattern: the title already promises the conclusion, the introduction repeats it, and the ending says it again. 77% of AI posts end by restating the main point versus 12% for human posts.
We also trained a second classifier to identify which of the five models (or a human) authored a post. It picks the correct author 79% of the time, far above random guessing (1 in 6 ≈ 17%). Most mistakes are swaps between AI models, not between AI and human.
Structural features are robust to re‑wording. We let each model rewrite its own posts until, on average, 73% of the original 13‑word sequences vanished, yet the AI‑slop detector performed just as well.
We are now embedding this model into Sitefire so that the posts our agents write vary more like human writing. Unexplored areas include the many human‑izer tools, manual rewrites, restructuring, or prompting models to explicitly avoid these habits. Another limitation is that our human corpus mainly spans 2020‑2022, while the AI‑generated posts were created in August 2026, so temporal effects cannot be ruled out.
The full study with figures is on arXiv (https://arxiv.org/abs/2609.15369) and the code is on GitHub (https://github.com/pulse-energy-eu/slopshape). We also built a public checker (https://sitefire.ai/slop-checker) and a small game to test human intuition (https://sitefire.ai/spot-the-slop).
Review