Pass the Google Generative AI Leader — the First Time

1. A retail company's chief digital officer is briefing the board on the technology behind their new customer-service assistant. She explains that it runs on a very large model that was pre-trained on a broad, diverse dataset and can be adapted to many different downstream tasks such as summarization, classification, and chat. Which term best describes this type of model?

   • A. A relational database engine
   • B. A rules-based expert system
   • C. A traditional regression model trained for a single task
   • D. A foundation model

   Answer: D. A foundation model is a large model pre-trained on broad data that can be adapted (via prompting, fine-tuning, or grounding) to many downstream tasks. A rules-based expert system and a single-task regression model are narrow and not adaptable across many tasks, and a relational database engine stores and queries data rather than generating content.

   Exam domain: Fundamentals of Generative AI

2. When a leader is estimating the cost of using a large language model through Agent Platform, they are told that pricing is often based on the number of "tokens" processed in both the input and the output. What is a token in this context?

   • A. A security credential that authenticates the API call
   • B. A reward point a user earns for giving the model feedback
   • C. A unit of text, such as a word or part of a word
   • D. A single pixel in an image the model generates

   Answer: C. In LLMs, a token is a chunk of text (a whole word or a fragment of one) that the model reads or produces, and most usage-based pricing counts input and output tokens. The other options confuse the term with authentication credentials, loyalty points, or image pixels, none of which is how text token pricing works.

   Exam domain: Fundamentals of Generative AI

3. A media company wants to build a search experience where employees can type a plain-language question and retrieve conceptually similar documents even when the documents use different wording than the query. Their data team proposes converting both the documents and the queries into numerical vectors that capture semantic meaning, then matching them by similarity. What is being described?

   • A. Manual tagging of every document with metadata labels
   • B. Diffusion-based image generation
   • C. Keyword indexing with exact-match lookup
   • D. Embeddings used with vector search

   Answer: D. Embeddings represent text (or other data) as vectors that capture semantic meaning, so similar concepts land near each other and can be matched with vector search even when wording differs. Exact-match keyword indexing would miss synonyms and paraphrases, diffusion generates images, and manual tagging does not produce the semantic vector matching described.

   Exam domain: Fundamentals of Generative AI

4. A product leader wants a single Google model that can accept a photo of a damaged product, read an attached text complaint, and produce a written summary of the issue for the support team. Which characteristic of the model makes this possible?

   • A. It is deterministic, returning identical output for any input
   • B. It is a narrow model fine-tuned only for image classification
   • C. It is an embedding-only model that outputs vectors rather than text
   • D. It is multimodal, combining inputs such as images and text

   Answer: D. Multimodal models (such as Gemini) can take multiple input types like images and text together and produce a combined response. Determinism is unrelated to handling multiple input types, a narrow image classifier cannot read the text complaint and write a summary, and an embedding-only model outputs vectors rather than the written summary required.

   Exam domain: Fundamentals of Generative AI

5. During a pilot, a company's legal team notices that the generative AI assistant confidently cited a court case that does not actually exist. What is this failure mode called?

   • A. Overfitting to the validation set
   • B. Latency
   • C. Hallucination
   • D. Rate limiting

   Answer: C. A hallucination is when a generative model produces content that is plausible-sounding but factually false or fabricated, such as a nonexistent legal citation. Latency refers to response speed, rate limiting caps request volume, and overfitting describes a training problem, none of which describes inventing false facts.

   Exam domain: Fundamentals of Generative AI

6. A retail company wants to give employees a single, conversational AI assistant they can access directly through a web interface or mobile app to brainstorm ideas, draft text, and analyze uploaded images, without setting up any cloud project. Which Google offering is designed for this consumer- and employee-facing chat experience?

   • A. Cloud Run
   • B. Model Garden on Agent Platform
   • C. The Gemini app
   • D. Agent Platform

   Answer: C. The Gemini app is Google's ready-to-use, end-user chat interface (web and mobile) powered by Gemini models, requiring no cloud project or development. Model Garden and the broader Agent Platform are aimed at builders and developers, and Cloud Run is a compute service for deploying containers, not a chat assistant.

   Exam domain: Google Cloud's Generative AI Offerings

7. A marketing director wants AI assistance embedded directly inside Gmail, Google Docs, Google Sheets, and Google Slides so staff can draft emails, summarize threads, and generate slide content without leaving those apps. Which offering provides this?

   • A. Gemini Notebook
   • B. Imagen for image generation
   • C. Gemini for Google Workspace
   • D. Agent Search

   Answer: C. Gemini for Google Workspace integrates Gemini's generative capabilities directly into Workspace apps like Gmail, Docs, Sheets, and Slides. Gemini Notebook (formerly NotebookLM) is a separate research and note-grounding tool, Imagen is an image-generation model that Google has discontinued in favor of Gemini image models, and Agent Search builds search experiences over enterprise data.

   Exam domain: Google Cloud's Generative AI Offerings

8. A bank's AI team wants a single place within Agent Platform to discover, compare, and select from Google first-party models (like Gemini and Veo), partner models, and open-source models before building. Which component meets this need?

   • A. Model Garden
   • B. Agent Search
   • C. Gemini Code Assist
   • D. Vector Search

   Answer: A. Model Garden is the Agent Platform catalog where teams browse, evaluate, and deploy a wide range of foundation models, including Google first-party (Gemini, Veo), partner, and open models. Agent Search builds search apps, Gemini Code Assist helps developers write code, and Vector Search is the similarity-search service, not a model catalog.

   Exam domain: Google Cloud's Generative AI Offerings

9. A leader is concerned that a customer-facing chatbot built on a foundation model invents plausible-sounding but false answers about the company's own products. The team wants responses anchored to the company's current, approved product documentation. Which technique most directly addresses this?

   • A. Switching to a smaller, cheaper model
   • B. Increasing the model's temperature setting
   • C. Adding more example conversations to the system instructions only
   • D. Grounding the model with retrieval over the company's product documents

   Answer: D. Grounding (often via Retrieval-Augmented Generation) connects the model to authoritative external or enterprise data so answers are based on real, current sources rather than the model's parametric memory, directly reducing hallucination. Raising temperature increases randomness, a smaller model does not fix factual grounding, and a few examples alone cannot supply the up-to-date product facts.

   Exam domain: Google Cloud's Generative AI Offerings

10. A product leader wants a low-code environment on Google Cloud's Agent Platform where business and technical staff can quickly test prompts, tune parameters, and compare model responses before committing to a production build. Which component is intended for this rapid experimentation?

    • A. Vector Search, for similarity lookup
    • B. Agent Studio, the visual AI workspace
    • C. BigQuery, the analytics data warehouse
    • D. Cloud Storage, for object file storage

    Answer: B. Agent Studio (formerly Vertex AI Studio) provides an interactive environment for prompt design, testing, parameter tuning, and comparing model outputs, ideal for rapid experimentation. Vector Search handles similarity search, Cloud Storage stores files, and BigQuery is a data warehouse, none of which is a prompt-experimentation interface.

    Exam domain: Google Cloud's Generative AI Offerings

Want more? Finish the quiz above to unlock 10 more free questions, or open the full Google Cloud Generative AI Leader practice app.