How Google autocomplete works in Search (2018) (blog.google)

Google’s 2018 explainer on Search Autocomplete describes a prediction system that helps finish queries as you type across Google surfaces (Search, Chrome Omnibox, mobile apps). Built from aggregated, real search logs and signals like location and prior queries, autocomplete updates instantly with each additional character (e.g., “san f” → San Francisco predictions; “san fe” → San Fernando). It’s a significant time-saver—about 25% less typing and an estimated 200+ years of typing saved per day—and supports richer results (dates, weather, Knowledge Graph icons) with UI differences by device (up to 10 suggestions on desktop, ~5 on mobile). For the AI/ML community the post highlights key engineering and safety trade-offs: models must balance relevance and utility with content-moderation constraints. Predictions are explicitly “predictions,” not endorsements, and Google applies automated filters plus human-reviewed policies to remove sexually explicit, hateful, violent, dangerous, spammy or pirated queries. In 2018 Google expanded hate and violence criteria to catch broader prejudiced or glorifying predictions, added a reporting/feedback loop, and uses reported examples to retrain and propagate removals to similar queries. Technical implications include reliance on massive real-time telemetry, incremental prediction models that update per character, Knowledge Graph enrichment for structured items, and ongoing work to combine ML relevance with policy-driven filtering and user feedback.