I co-authored a piece on the official Google Cloud blog with Hugo Selbie, on a problem that comes up constantly for generative AI and media companies: how do you analyse unstructured assets — images, video, audio — alongside the structured behavioural data that describes how people actually use them?
Read it here: Product analytics for generative AI model and media asset companies using BigQuery (Google Cloud blog, May 2024)
The problem
If you run an image generation product, your most valuable signal is split across two worlds. The generated assets sit in object storage. The user behaviour — prompts, retries, downloads, shares — sits in your warehouse. Answering “which kinds of output actually get kept?” means joining across both, and traditionally that meant building a pipeline to move and pre-process petabytes of media before you could ask the first question.
The approach
The article walks through using BigQuery as the single hub instead:
- Object tables over Cloud Storage via a BigLake remote connection, so unstructured assets are queryable in place rather than copied
- Native BQML functions to call generative models directly from SQL
- Remote UDFs backed by Cloud Functions, for the cases where you need to reach Vertex AI APIs directly
The result is that analysts can interrogate media assets and user behaviour in one query, at petabyte scale, without a separate preprocessing pipeline and without duplicating storage.
Why it matters
The use cases this unlocks are the ones product teams actually care about: understanding which model outputs perform, flagging inappropriate content, improving search over a media catalogue, and doing it without standing up bespoke infrastructure for each question.
The full walkthrough — including the five-step implementation and the SQL — is on the Google Cloud blog.
Annie Xu 
Leave a comment