RecSys 2026 · Industry Track · Minneapolis

The Text on the Creative

An Under-Exploited Ranking Modality for Short-Form Video Ads

Shubham Goel, Andrew Treadway, Yundi Jiang, Yiding Wen, Jianing Fu, Alan Li, Angli Liu, Antoine Simoulin, Himanshu Thakur, Haibo Zhang, Selahattin Akkas, Calvin Ma, Zhen Zeng, Yunyu He, Qin Huang, Benjamin Au, Guy Lebanon, Sagar Chordia

Meta, Instagram Ads Ranking. Underlined names are presenting the poster.

Tap the poster for the full-resolution PDF.

The short version

Short-form video ads put their pitch right on screen: prices, offers, calls to action. The Instagram ads ranker wasn’t using any of that text. We got it in cheaply with three pieces.

1. Frame selection. We reuse the CTR-trained scorer that already picks ad thumbnails to rank frames, and drop near-duplicates. At 6 frames it recovers more text than uniform sampling at 12.
2. Ads-OCR. An in-house OCR model trained on ad creatives in 41 languages. It gets 93% of a VLM’s word F1 at about 80× the throughput.
3. Compact features. An RQ-VAE turns the text embedding into three 1-byte codes, and TF-IDF keeps the exact words. Both go into the ranker as sparse ID features, about 500× smaller than the dense embedding.

Offline, the bundle gives a 0.055–0.06% NE gain across surfaces and ≈0.12% on Reels-class video. Online, a multi-week A/B on 80% of traffic gave 0.06–0.08% (p < 0.01), and the features are now in production.

Questions or want to compare notes? shubhamgoel27@gmail.com