Google Cloud blueprint for recommender builds

A recommendation system is software that guesses what a person should watch, read, or buy next from past clicks, views, or ratings. Google Cloud has now published a reference build for running that kind of system on its managed services. (docs.cloud.google.com) The stack it points to centers on BigQuery for storing and querying large datasets, Vertex AI for training and managing models, Dataflow for moving and transforming data streams, and Cloud Run for serving application logic. Google Cloud promoted the blueprint in April 2026 as a pattern for large content catalogs and portfolio-scale media businesses. (docs.cloud.google.com) (cloud.google.com) In plain terms, BigQuery acts like the warehouse, Dataflow acts like the conveyor belt, Vertex AI acts like the training lab, and Cloud Run acts like the front counter that answers live requests. Google’s own architecture guidance describes this kind of reference design as prescriptive documentation for building production systems on its platform. (cloud.google.com) (docs.cloud.google.com) Google already sells a more packaged recommendation product through Vertex AI Search, formerly Recommendations AI. That service says media companies can optimize for goals such as engagement, ad inventory, or trial-subscription conversion without manually provisioning infrastructure. (cloud.google.com) The new blueprint matters because it sits between a turnkey product and a fully custom build. Companies that want their own data pipelines, model choices, and serving layer can keep those controls while still using managed Google Cloud components underneath. (cloud.google.com) (docs.cloud.google.com) Google’s current media-recommendation tutorial shows the shape of that workflow. It imports catalog data and user events from BigQuery into Vertex AI Search, trains a movie recommendation model on the MovieLens dataset, and says initial setup takes about 1.5 hours while model training can take about 24 hours. (docs.cloud.google.com) For teams that need fresher signals than nightly batch jobs, Dataflow is the piece that keeps behavior data moving. Google says Dataflow supports real-time extraction, transformation, and loading into BigQuery, scales to 4,000 workers per job, and is used for patterns including personalized recommendations. (cloud.google.com) BigQuery and Vertex AI have been moving closer together for several years. Google said in February 2024 that BigQuery Machine Learning usage had grown 250% year over year and added tighter Vertex AI integration so teams could blend structured data, unstructured data, and models in one pipeline. (cloud.google.com) Google has also been pushing recommendation tools directly at media and commerce customers. Its recommendations page cites Newsweek with a 10% increase in total revenue per visit and says STARZ uses Google’s recommendation technology to help viewers find more relevant content. (cloud.google.com) The pitch in this blueprint is not that recommendation software is new. It is that Google wants broadcasters and large platforms to assemble it from BigQuery, Vertex AI, Dataflow, and Cloud Run instead of stitching together more infrastructure on their own. (docs.cloud.google.com)