Free Google Generative AI Leader practice exam questions

What the 20 questions cover

Every question here passed the same checks as the full bank: the quoted sentence was found on the live official page, and an independent blind reviewer matched the answer. How we check questions.

The questions

Answers are hidden. Open “Show answer and source” under a question after you have picked yours.

1. Question 1 of 20 · S1: Fundamentals of gen AI · gal-s1-003

   During a vendor review, a procurement lead says, "This model was trained on images and audio, not text, so it can't be a foundation model." How should you respond?

   • A. Option A: The lead is right; image models are called diffusion models, and those are never foundation models.
   • B. Option B: The lead is right; a model only counts as a foundation model if it was trained mainly on text and code.
   • C. Option C: It depends only on parameter count; anything above a billion parameters is a foundation model.
   • D. Option D: Foundation models can be trained on images, audio, video or a mix; LLMs are one text-focused type.
   Show answer and source

   Answer: D. Think of LLMs as a subset of foundation models. Multimodal, vision and audio models can also be foundation models.

   • A (incorrect): Training data type does not exclude a model from being a foundation model.
   • B (incorrect): Text and code define LLMs. The foundation model category is broader.
   • C (incorrect): The definition rests on broad pre-training and adaptability, not a parameter threshold.
   • D (correct): Correct. Google Cloud explains that all LLMs are foundation models, but not all foundation models are LLMs.

   Official sourcecloud.google.com

   all LLMs are foundation models, but not all foundation models are LLMs

   cloud.google.com/discover/what-are-foundation-models (official page)

2. Question 2 of 20 · S1: Fundamentals of gen AI · gal-s1-035

   A data steward wants an ongoing measure of whether customer records are complete before they feed a model. What does Google Cloud recommend tracking?

   • A. Option A: The average response length of the chatbot
   • B. Option B: The number of GPUs used in each training run for the model
   • C. Option C: The number of model versions in the registry
   • D. Option D: Null-value percentages and field-fill rates
   Show answer and source

   Answer: D. Completeness is a core data-quality characteristic. Measure it continuously rather than checking once.

   • A (incorrect): Response length is an output metric, not an input data-quality metric.
   • B (incorrect): Accelerator count says nothing about data completeness.
   • C (incorrect): Version count is unrelated to whether the records are complete.
   • D (correct): Correct. Google Cloud says to assess completeness by tracking null-value percentages and required field-fill rates.

   Official sourcedocs.cloud.google.com

   Assess completeness by tracking null-value percentages and required field-fill rates.

   docs.cloud.google.com/architecture/framework/perspectives/ai-ml/reliability (official page)

3. Question 3 of 20 · S1: Fundamentals of gen AI · gal-s1-046

   A CIO asks which layer of the gen AI landscape Google's AI Hypercomputer belongs to. What is the best answer?

   • A. Option A: Infrastructure: an integrated supercomputing system optimized for AI workloads
   • B. Option B: Models: a family of pre-trained multimodal foundation models
   • C. Option C: Applications: a finished assistant that business employees use daily to search and draft content
   • D. Option D: Agents: software that pursues goals and completes tasks for users
   Show answer and source

   Answer: A. The landscape stacks infrastructure, models, platforms, agents and applications. Hardware and supercomputing belong to the bottom layer.

   • A (correct): Correct. AI Hypercomputer is the compute foundation that models and platforms run on.
   • B (incorrect): Models run on the infrastructure but form a separate layer.
   • C (incorrect): Employees do not use AI Hypercomputer directly as an application.
   • D (incorrect): AI Hypercomputer runs workloads; it is not an agent.

   Official sourcedocs.cloud.google.com

   AI Hypercomputer is an integrated supercomputing system that's optimized to support your artificial intelligence (AI) and machine learning (ML) workloads.

   docs.cloud.google.com/ai-hypercomputer/docs/overview (official page)

4. Question 4 of 20 · S1: Fundamentals of gen AI · gal-s1-052

   A bank wants a customer-facing voice assistant that holds natural spoken back-and-forth conversations in real time and can hand tasks to other agents during the call. Which kind of Gemini model is optimized for this?

   • A. Option A: A Gemini Flash Cyber model
   • B. Option B: A Gemini Live model
   • C. Option C: A Gemini Flash Image model
   • D. Option D: A Gemini Pro model
   Show answer and source

   Answer: B. Gemini is a family with variants for different strengths: Pro for hard reasoning, Flash and Flash-Lite for speed and cost, and specialized Live, Image and Cyber variants.

   • A (incorrect): Cyber variants are optimized for cybersecurity work, not customer voice conversations.
   • B (correct): Correct. Google describes Gemini Live models as optimized for live interactions, with improved voice quality, reliability and real-time agent orchestration.
   • C (incorrect): Flash Image models generate and edit images, not live voice conversations.
   • D (incorrect): Pro models target complex reasoning and coding; the Live variants are the ones optimized for real-time voice.

   Official sourcedocs.cloud.google.com

   Optimized for live interactions, with improved voice quality, reliability, and real-time agent orchestration.

   docs.cloud.google.com/gemini-enterprise-agent-platform/models/google-models (official page)

5. Question 5 of 20 · S1: Fundamentals of gen AI · gal-s1-053

   A hardware company wants a lightweight model it can run on its own devices and servers and adapt freely. Which Google model family fits?

   • A. Option A: Veo, a hosted video generation model family
   • B. Option B: Gemini 3 Pro Image, a hosted image generation model
   • C. Option C: Gemma, a set of lightweight, generative open models
   • D. Option D: Gemini Enterprise, a managed AI assistant for employees
   Show answer and source

   Answer: C. Gemma brings Gemini-derived research into open models that organizations can host and customize themselves.

   • A (incorrect): Veo is a hosted video model, not a lightweight model for your own hardware.
   • B (incorrect): Gemini 3 Pro Image is a hosted image generation model, not an open model you run yourself.
   • C (correct): Correct. Gemma models are lightweight open models you can run in your applications, on your hardware or on mobile devices.
   • D (incorrect): Gemini Enterprise is an application, not a model to embed on devices.

   Official sourcedocs.cloud.google.com

   Gemma is a set of lightweight, generative artificial intelligence (AI) open models.

   docs.cloud.google.com/gemini-enterprise-agent-platform/models/open-models/use-gemma (official page)

6. Question 6 of 20 · S1: Fundamentals of gen AI · gal-s1-055

   A field-service app must run a model offline on low-resource devices and accept text, image and audio input. Which Gemma variant matches?

   • A. Option A: CodeGemma, designed for code completion and generation on laptops
   • B. Option B: Gemma 3n, designed for efficient use on low-resource devices
   • C. Option C: ShieldGemma 2, designed for checking image safety on devices
   • D. Option D: TxGemma, designed for therapeutic prediction tasks offline
   Show answer and source

   Answer: B. The Gemma family includes specialized variants. Choose by deployment target and task.

   • A (incorrect): CodeGemma targets coding tasks, not multimodal field use.
   • B (correct): Correct. Google describes Gemma 3n as designed for efficient execution on low-resource devices, with multimodal input.
   • C (incorrect): ShieldGemma checks safety against policies; it is not a general on-device assistant.
   • D (incorrect): TxGemma targets therapeutic data, not field-service apps.

   Official sourcedocs.cloud.google.com

   An open model designed for efficient execution on low-resource devices

   docs.cloud.google.com/gemini-enterprise-agent-platform/models/google-models (official page)

7. Question 7 of 20 · S2: Google Cloud's gen AI offerings · gal-s2-001

   A general counsel asks whether employees' prompts and outputs in Gemini Enterprise will be used to improve Google's models. What does Google Cloud state?

   • A. Option A: Prompts, outputs and model tuning are never used to train Google models or other customers' models.
   • B. Option B: Only outputs are used to train Google models; prompts are deleted as soon as a response is returned.
   • C. Option C: Prompts are used to improve Google's models unless each employee opts out individually in their settings.
   • D. Option D: Data is used to train Google models, but only after it has been anonymized.
   Show answer and source

   Answer: A. Data control is a core enterprise-readiness claim: your business data stays yours. Agent Platform's service terms word it as no training or fine-tuning on your data without your prior permission or instruction.

   • A (correct): Correct. Google Cloud states that your data, including prompts, outputs and model tuning, is never used to train Google models or models for any other customer.
   • B (incorrect): Neither prompts nor outputs are used to train Google models.
   • C (incorrect): Google states this data isn't used to train its models at all, so there is no opt-out to manage.
   • D (incorrect): Google's commitment is not to train on the data at all, not to anonymize it first.

   Official sourcecloud.google.com

   Your data, including prompts, outputs, and model tuning, is never used to train Google models or models for any other customer

   cloud.google.com/gemini-enterprise (official page)

8. Question 8 of 20 · S2: Google Cloud's gen AI offerings · gal-s2-015

   Staff at a manufacturer waste hours hunting for answers across SharePoint, Confluence and Jira. Leadership wants one assistant that respects existing file permissions. Which product fits?

   • A. Option A: A fine-tuned open model trained on a one-time export of every file
   • B. Option B: Gemini Enterprise, with one permissions-aware search interface
   • C. Option C: Cloud Translation, to convert all documents into a single language before indexing them
   • D. Option D: Customer Experience Insights, to analyze contact center calls
   Show answer and source

   Answer: B. Gemini Enterprise connects to enterprise data sources and returns answers that respect each user's access rights.

   • A (incorrect): A one-time export goes stale and ignores per-user permissions.
   • B (correct): Correct. Gemini Enterprise gives employees one search interface with permissions-aware access to enterprise information.
   • C (incorrect): Translation does not unify search across systems.
   • D (incorrect): Insights analyzes contact center data, not employee knowledge search.

   Official sourcedocs.cloud.google.com

   This provides employees with a single, multimodal search interface offering permissions-aware access to enterprise information.

   docs.cloud.google.com/gemini/enterprise/docs (official page)

9. Question 9 of 20 · S2: Google Cloud's gen AI offerings · gal-s2-027

   A city government wants residents to search its public website of permits, policies and FAQs with natural questions like "Do I need a permit for a backyard deck?" and get relevant results. Which offering fits?

   • A. Option A: Natural Language API entity analysis
   • B. Option B: AI Commerce Search
   • C. Option C: Cloud Translation API
   • D. Option D: Agent Search for custom apps
   Show answer and source

   Answer: D. Agent Search (formerly Vertex AI Search) brings Google-quality, intent-aware search to an organization's own websites and data.

   • A (incorrect): Entity analysis tags text; it doesn't provide a search experience.
   • B (incorrect): AI Commerce Search is for commerce and retail product data, not a city's permit and policy pages.
   • C (incorrect): Translation converts languages; it doesn't improve search relevance.
   • D (correct): Correct. Agent Search for custom apps is a Google-quality search engine for website data and other content, and it goes beyond basic keyword matching.

   Official sourcedocs.cloud.google.com

   Agent Search for custom apps is a powerful, Google-quality search and content discovery engine

   docs.cloud.google.com/generative-ai-app-builder/docs/about-generic-search (official page)

10. Question 10 of 20 · S2: Google Cloud's gen AI offerings · gal-s2-042

    A support team's chatbot gives confident but wrong answers about internal policies the model never saw. Which managed Agent Platform service addresses this?

    • A. Option A: RAG Engine, to add private knowledge to the model's context
    • B. Option B: A model with a larger context window, used without adding any policy documents
    • C. Option C: Higher temperature for more varied answers
    • D. Option D: An image generation model, to illustrate each answer
    Show answer and source

    Answer: A. RAG Engine is Agent Platform's managed framework for retrieval-augmented generation over your own data.

    • A (correct): Correct. RAG Engine adds private knowledge to the context so the model can reduce hallucinations.
    • B (incorrect): More room in the window doesn't add missing knowledge.
    • C (incorrect): Randomness doesn't fix missing facts.
    • D (incorrect): Images don't fix factual errors.

    Official sourcedocs.cloud.google.com

    With RAG Engine, you can enrich the LLM context with additional private information, so that the model can reduce hallucinations

    docs.cloud.google.com/gemini-enterprise-agent-platform/build/rag-engine/rag-overview (official page)

11. Question 11 of 20 · S2: Google Cloud's gen AI offerings · gal-s2-053

    A COO asks how an AI agent can check live inventory in the company's ERP system instead of guessing. Which mechanism enables this?

    • A. Option A: Function calling, which describes external tools the model can use
    • B. Option B: Lowering the top-p value
    • C. Option C: A longer system prompt that describes typical stock levels for every warehouse and SKU
    • D. Option D: Fine-tuning on last year's inventory reports
    Show answer and source

    Answer: A. Tools let agents interact with the outside world; function calling is the standard way to describe them to the model.

    • A (correct): Correct. Function calling (tool use) gives the LLM definitions of external tools, such as an inventory lookup.
    • B (incorrect): Sampling settings don't connect to external systems.
    • C (incorrect): Describing typical levels doesn't retrieve live data.
    • D (incorrect): Tuning bakes in stale data rather than live lookups.

    Official sourcedocs.cloud.google.com

    Function calling , also known as tool use , provides the LLM with definitions of external tools

    docs.cloud.google.com/gemini-enterprise-agent-platform/models/tools/function-calling (official page)

12. Question 12 of 20 · S2: Google Cloud's gen AI offerings · gal-s2-055

    An IT architect wants agents to connect to existing databases and enterprise systems through open standards rather than custom adapters. What does Google Cloud support?

    • A. Option A: Only proprietary connectors, which work with Google databases
    • B. Option B: Manual CSV uploads before every agent run
    • C. Option C: Native support for MCP and OpenAPI 3.0
    • D. Option D: Screen scraping of each application's user interface
    Show answer and source

    Answer: C. Open protocols make tools reusable across agents and reduce custom integration work.

    • A (incorrect): The support is for open standards.
    • B (incorrect): That isn't a live connection.
    • C (correct): Correct. Google Cloud highlights native MCP and OpenAPI 3.0 support for connecting agents to systems.
    • D (incorrect): Open protocols replace brittle scraping.

    Official sourcecloud.google.com

    Native support for Model Context Protocol (MCP) and OpenAPI 3.0 ensures your agents connect seamlessly to existing databases and enterprise systems.

    cloud.google.com/ai (official page)

13. Question 13 of 20 · S2: Google Cloud's gen AI offerings · gal-s2-056

    A team needs a place to keep thousands of product manuals as files that agents and data stores can draw on. Which Google Cloud service fits?

    • A. Option A: Cloud Run, which runs containers
    • B. Option B: Natural Language API, which analyzes the text in uploaded manuals
    • C. Option C: Agent Gallery, which lists agents
    • D. Option D: Cloud Storage, which stores data as objects in buckets
    Show answer and source

    Answer: D. Cloud Storage is a common source for documents used by data stores and RAG.

    • A (incorrect): Cloud Run runs code; it isn't a file store.
    • B (incorrect): That API analyzes text but doesn't store files.
    • C (incorrect): Agent Gallery isn't storage.
    • D (correct): Correct. Cloud Storage stores data as objects in buckets.

    Official sourcedocs.cloud.google.com

    Cloud Storage is a scalable and managed storage service offered by Google Cloud that lets you store data as objects in containers called buckets

    docs.cloud.google.com/storage/docs/introduction (official page)

14. Question 14 of 20 · S3: Techniques to improve gen AI model output · gal-s3-001

    A chatbot answers questions about a product line launched last month with details about the old line. Which foundation model limitation is most likely at work?

    • A. Option A: Its knowledge is limited to training data, which goes out of date
    • B. Option B: Hallucination, because the model invents product details at random
    • C. Option C: The model was built with unsupervised learning only
    • D. Option D: Bias amplification, because older products dominate the training data
    Show answer and source

    Answer: A. Knowledge cutoff is why grounding and RAG matter for current or private information.

    • A (correct): Correct. LLMs only know what they were trained on, which can quickly become outdated.
    • B (incorrect): The answers describe a real older line, which points to stale knowledge rather than invented facts.
    • C (incorrect): The training method isn't the issue here.
    • D (incorrect): Size isn't the cause; missing training data is.

    Official sourcecloud.google.com

    Usually, LLMs only know what they were trained on, which can quickly become outdated.

    cloud.google.com/discover/what-is-model-context-protocol (official page)

15. Question 15 of 20 · S3: Techniques to improve gen AI model output · gal-s3-017

    A team expects to write a perfect prompt on the first try. What does Google's guidance say about the prompt engineering process?

    • A. Option A: It matters only for image models
    • B. Option B: It is a one-time task done carefully before launch, then left alone
    • C. Option C: It is a test-driven, iterative process
    • D. Option D: It replaces the need for any evaluation
    Show answer and source

    Answer: C. Treat prompts like code: test, measure, refine.

    • A (incorrect): It applies to LLMs broadly.
    • B (incorrect): Prompts should be tested and refined.
    • C (correct): Correct. Google describes prompt engineering as test-driven and iterative.
    • D (incorrect): Iteration depends on evaluation.

    Official sourcedocs.cloud.google.com

    Prompt engineering is a test-driven and iterative process that can enhance model performance.

    docs.cloud.google.com/gemini-enterprise-agent-platform/models/prompts/prompt-design-strategies (official page)

16. Question 16 of 20 · S3: Techniques to improve gen AI model output · gal-s3-030

    A delivery app's assistant should answer "which stores near me are open late" with accurate location context. Which grounding option fits?

    • A. Option A: Grounding with Google Maps
    • B. Option B: Few-shot prompting with sample addresses
    • C. Option C: Grounding with an internal policy data store
    • D. Option D: Chain-of-thought prompting
    Show answer and source

    Answer: A. Pick grounding sources that match the information the answer needs.

    • A (correct): Correct. Maps grounding adds geospatial context for more accurate responses.
    • B (incorrect): Examples don't provide real store data.
    • C (incorrect): Policies don't include live location data.
    • D (incorrect): Reasoning steps don't add location facts.

    Official sourcedocs.cloud.google.com

    Use Google Maps data with your model to provide more accurate and context-aware responses to your prompts, including geospatial context.

    docs.cloud.google.com/gemini-enterprise-agent-platform/models/grounding/overview (official page)

17. Question 17 of 20 · S3: Techniques to improve gen AI model output · gal-s3-032

    How does retrieval-augmented generation change a model's output?

    • A. Option A: It makes answers longer and more detailed by default in every case
    • B. Option B: It permanently changes the model's weights
    • C. Option C: It replaces the model's training data permanently
    • D. Option D: Retrieved facts enrich the context for more precise answers
    Show answer and source

    Answer: D. RAG improves accuracy and freshness without retraining.

    • A (incorrect): RAG changes what the model knows for the request, not answer length.
    • B (incorrect): That is tuning, not RAG.
    • C (incorrect): Retrieved content is added per request; the model itself is unchanged.
    • D (correct): Correct. RAG incorporates retrieved information into the LLM's context to improve responses.

    Official sourcecloud.google.com

    This augmented context enables the LLM to generate more precise, informative, and engaging responses.

    cloud.google.com/use-cases/retrieval-augmented-generation (official page)

18. Question 18 of 20 · S4: Business strategies for a successful gen AI solution · gal-s4-003

    A CFO wants AI initiatives tied to company strategy. What does Google recommend when defining them?

    • A. Option A: Keep objectives loose for the first year so teams can explore freely
    • B. Option B: Start with the use case that showcases the most advanced model capabilities
    • C. Option C: Let each business unit set its own AI goals and success measures independently
    • D. Option D: Align projects with strategic goals like revenue, cost and satisfaction
    Show answer and source

    Answer: D. Business requirements come first; technology choices follow.

    • A (incorrect): Loose objectives make it impossible to measure success or ROI.
    • B (incorrect): Showcasing technology is not the same as advancing strategy.
    • C (incorrect): That weakens alignment.
    • D (correct): Correct. Google recommends aligning AI projects with strategic objectives.

    Official sourcedocs.cloud.google.com

    Ensure that AI and ML projects are aligned with strategic objectives like revenue growth, cost reduction, customer satisfaction, and efficiency.

    docs.cloud.google.com/architecture/framework/perspectives/ai-ml/cost-optimization (official page)

19. Question 19 of 20 · S4: Business strategies for a successful gen AI solution · gal-s4-012

    A CISO asks what Google's Secure AI Framework (SAIF) is for. Which answer is accurate?

    • A. Option A: It replaces IAM and network controls for AI projects with one AI-specific policy
    • B. Option B: It is a compliance certification that AI products must pass before launch
    • C. Option C: It applies only to on-premises systems
    • D. Option D: It addresses AI model risk, security and privacy so AI is secure by default
    Show answer and source

    Answer: D. SAIF gives organizations a structured way to manage AI-specific security risk.

    • A (incorrect): SAIF guides security; IAM remains the access tool.
    • B (incorrect): SAIF is a conceptual framework and guidance, not a certification.
    • C (incorrect): SAIF applies to AI systems broadly.
    • D (correct): Correct. SAIF addresses AI/ML model risk, security and privacy concerns.

    Official sourcecloud.google.com

    SAIF is designed to address top-of-mind concerns for security professionals, such as AI/ML model risk management, security, and privacy

    cloud.google.com/use-cases/secure-ai-framework (official page)

20. Question 20 of 20 · S4: Business strategies for a successful gen AI solution · gal-s4-021

    A sales leader asks why responsible AI matters commercially, not just ethically. What does Google Cloud point to?

    • A. Option A: Responsible AI only matters for government customers
    • B. Option B: Responsible AI slows products so much that it lowers sales
    • C. Option C: Customers rarely consider AI practices when buying
    • D. Option D: Lack of trust in AI is a growing barrier to adoption
    Show answer and source

    Answer: D. Trust is a business asset; responsible AI protects it.

    • A (incorrect): Trust affects enterprise buying broadly.
    • B (incorrect): Google frames it as good for business.
    • C (incorrect): Google states the opposite.
    • D (correct): Correct. Google notes buyers increasingly select products based on AI commitments and practices.

    Official sourcecloud.google.com

    Lack of trust in AI systems is a growing barrier to adoption in enterprise

    cloud.google.com/responsible-ai (official page)