Scribd classified 400 million-plus documents, spanning 12 billion-plus pages, using Gemini Enterprise batch inference over several months.

Google Cloud’s customer account reports that Scribd classified more than 400 million user-uploaded documents containing more than 12 billion pages of text and images. The completed trust-and-safety backfill covered Scribd and Slideshare and took a matter of months, with Google Cloud scaling batch throughput to meet the timeline.

Native PDF processing allowed more than 99% of the corpus to be handled as-is. Scribd therefore did not need to build an optical character recognition (OCR), rendering, or screenshotting pipeline for that portion of the corpus.

Gemini Enterprise batch prediction was priced at a 50% discount to interactive inference, which the account identifies as important to the economics of corpus-scale classification. The article does not disclose the total project cost, exact processing duration, or classification-accuracy measurements.

Scribd’s team said previously evaluated moderation tools and open models had not delivered the required quality at its scale, while Gemini consolidated the work into one model, prompt, and pipeline. Scribd, Inc. also owns Everand and Fable, but the reported 400-million-plus-document backfill specifically covered Scribd and Slideshare; the company continues to describe its content-review approach as a mix of human and automated methods.