Grounding and guiding Google Docs' GenAI feature development

Overview

In 2023, Google Workspace sharply accelerated its focus on Generative AI. The UXR team already knew how to research how features got delivered — the open question was how to support fundamentally new kinds of features: which GenAI investments were actually valuable, how to evaluate model performance as the product team built it, and how to evaluate the delivery experience.

Split the problem into three concrete questions: roadmap and strategy (what GenAI features would be most valuable, and how could GenAI best help people get work done?), model evaluation (how do we assess model performance as the team builds?), and usability (how do we evaluate how those features are actually delivered?).

Approach

Roadmap and strategy: consolidated and evangelized a shared UXR point of view on Docs' top strengths and gaps from a user standpoint, built on four years of prior research. Designed and led the Product Values survey — asking paid Docs users which of a set of value statements already feel true of Docs (% fulfillment) and which they most want Docs to improve on (% demand) for each candidate product value. Plotting fulfillment against demand produced a 2×2: high-fulfillment/low-demand areas to consolidate, high-fulfillment/high-demand areas to grow, low-fulfillment/high-demand areas to invest in, and low-fulfillment/low-demand areas to leave alone — a simple, scalable way to prioritize where to point GenAI effort.

Usability: scaled tactical research two ways — focusing only on the design patterns UXD was proposing that hadn't yet been tested (not every individual interaction), and defaulting to unmoderated testing and rapid research for well-scoped questions, freeing up capacity for the riskier, more ambiguous usability calls.

Model evaluation: this was a genuinely new need with no established practice to draw on. Established an evaluation working group to define and iterate on process, and led the development of a centralized set of resources — stimuli/prompts to test models against, a measurement questionnaire for rating output, a tester pool, and standardized top-line and detailed-analysis reporting — so any team could run a consistent model evaluation session.

Impact

Directly supported the successful launch of multiple GenAI features: Help Me Create (Dec 2024), Corpus Grounding (Oct 2025), and Cover Images (Oct 2024).

Firmly grounded the 2025+ Docs GenAI strategy in user needs, attitudes, and behaviors via the Product Values survey.

Established an iterative model evaluation process now used across all of Docs — the same foundation later extended into the conversation-level evaluation methodology built at Meta.