The Text on the Creative: An Under-Exploited Ranking Modality for Short-Form Video Ads

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Video advertisements contain rich textual signal overlaid on the creative itself — promotional copy, prices, calls-to-action, brand and product mentions — that current production ranking systems largely under-exploit. We argue this is a high-yield, low-cost ranking modality and demonstrate a production deployment that brings it into Instagram’s ads ranking model at an Internet scale. Making this practical requires three components, each motivated by the modality’s deployment economics: (1) a specialized OCR backbone delivering near-VLM word-level quality at ∼ 80 × higher production throughput, which makes per-frame perception economical at an Internet scale; (2) a smart frame-extraction stage tailored to short-form video, combining a CTR-trained quality scorer with windowless-SSIM de-duplication, that recovers more OCR-relevant frames than uniform sampling at half the per-video frame budget; and (3) a feature stack — TF–IDF tokens and discrete RQ-VAE semantic IDs over a domain-tuned sentence-embedding backbone — that integrates into the ranker’s existing sparse embedding tables without a dedicated dense tower. We validate the stack with a four-lens ablation (LLM-as-judge with 200, 000-ad coverage, deterministic lexical-recall and reconstruction anchors, and a 5-rater human-vs-LLM agreement check with Spearman ρ = 0.74). The launched bundle delivers 0.055–0.06 offline relative NE gain aggregated across all surfaces served by the ranker — concentrated at ≈ 0.12% on short-form video (Reels-class) surfaces where promotional overlay text is densest — and a 0.06–0.08% NE gain in a multi-week 80%-traffic online A/B; per-feature attribution shows the 3-byte discrete codes (looked up through the ranker’s existing embedding tables) carry the majority of the lift at a 500 × smaller per-ad storage footprint than the dense backbone they were trained from, with the TF–IDF tokens contributing complementary high-recall lexical signal that the codes blur (brand names, numeric promotions, tail vocabulary).