This paper presents a method for scaling up large language model (LLM) reviews for content moderation in Google Ads. The goal is to detect policy violations in ads before they are eligible to be served. The approach involves selecting candidate ads using heuristics, clustering, and duplicate removal, then using LLMs to review only the representative ads from each cluster. The LLM decisions are then propagated back to the clusters to label all similar ads. This method reduces the number of reviews by more than 3 orders of magnitude while achieving a 2x recall compared to a baseline non-LLM model. The success of this approach is strongly dependent on the representations used in clustering and label propagation, with cross-modal similarity representations yielding better results than uni-modal ones. The method combines funneling, LLM labeling, label propagation, and a feedback loop. Funneling reduces the volume of content processed by the LLM using heuristic-based selection, deduping, filtering, and cluster-based sampling. The LLM is then used to label the selected candidates, and label propagation is used to propagate the labels to similar images. A feedback loop from the final labeled images helps select similar candidate images for subsequent rounds of funneling. The approach was evaluated on the "Non-Family Safe" ad content policy, which restricts sexually suggestive content, sexual merchandise, and nudity. The pipeline labeled roughly twice as many images as a multi-modal non-LLM model, while also surpassing its precision on this policy. The method helped remove more than 15% of the policy-violating impressions among image ads for this policy. The technique is being expanded to more ad policies and modalities, such as videos, text, and landing pages. The quality of all pipeline stages is being improved by exploring better heuristics, tuning better LLM prompts, and propagating similarity through higher-quality embeddings.This paper presents a method for scaling up large language model (LLM) reviews for content moderation in Google Ads. The goal is to detect policy violations in ads before they are eligible to be served. The approach involves selecting candidate ads using heuristics, clustering, and duplicate removal, then using LLMs to review only the representative ads from each cluster. The LLM decisions are then propagated back to the clusters to label all similar ads. This method reduces the number of reviews by more than 3 orders of magnitude while achieving a 2x recall compared to a baseline non-LLM model. The success of this approach is strongly dependent on the representations used in clustering and label propagation, with cross-modal similarity representations yielding better results than uni-modal ones. The method combines funneling, LLM labeling, label propagation, and a feedback loop. Funneling reduces the volume of content processed by the LLM using heuristic-based selection, deduping, filtering, and cluster-based sampling. The LLM is then used to label the selected candidates, and label propagation is used to propagate the labels to similar images. A feedback loop from the final labeled images helps select similar candidate images for subsequent rounds of funneling. The approach was evaluated on the "Non-Family Safe" ad content policy, which restricts sexually suggestive content, sexual merchandise, and nudity. The pipeline labeled roughly twice as many images as a multi-modal non-LLM model, while also surpassing its precision on this policy. The method helped remove more than 15% of the policy-violating impressions among image ads for this policy. The technique is being expanded to more ad policies and modalities, such as videos, text, and landing pages. The quality of all pipeline stages is being improved by exploring better heuristics, tuning better LLM prompts, and propagating similarity through higher-quality embeddings.