You nailed it with "design patterns." That is exactly the internal mental model we use. We are trying to move the industry away from "copying a vibe" to "implementing a successful structural pattern."
To answer your questions on the depth of the RE:
Text vs. Visuals: We capture both. Our analysis pipeline runs audio transcription, but it also extracts visual text (OCR) from the frames. So if a video has no spoken words but relies on text overlays and scene changes, we still capture the full narrative structure. If you drag the creative into the chat, we feed all of this—spoken words, on-screen text, and visual descriptions—into the LLM context. This allows you to ask it to identify the specific beats (Hook, Problem, Solution, CTA) or analyze the pacing yourself.
The "Why": Currently, without access to private metrics, the Chat generates strong hypotheses based on established frameworks (e.g., "This likely works because of the scarcity trigger"). However, we plan to allow users to connect their own Meta accounts in the future. The goal is to let users pull real metrics (CTR, CPA, ROAS) and map them against the creative analysis. This will allow us to close the loop and correlate specific "design patterns" with actual performance data.
Avoiding Overfitting: This is exactly why we built the "User Ingestion" feature. If we only showed a curated "Best of" feed, everyone would copy the same 5 trends. By allowing users to paste any Facebook Ad Library link to ingest a specific brand, they define their own dataset. This allows them to study niche-specific patterns rather than just generic platform trends.