Complete guide
Google Cloud Generative AI Leader
Tap any service to see more details and official docs.
Exam breakdown
40–60
Questions
~40
Min read
700
Passing score
Fundamentals of Generative AI
30% of the examWhat is generative AI?
Generative AI creates new content — text, images, audio, video or code — rather than only classifying or predicting over existing data. It relies on foundation models trained on enormous volumes of data.
| Concept | Description |
|---|---|
| Traditional (discriminative) AI | Classifies, predicts or detects patterns over existing data. E.g.: detecting spam, predicting churn. |
| Generative AI | Generates new content that didn't exist before, learning the statistical distribution of the training data. |
| Foundation model | A large model trained on massive, general-purpose data, adaptable to multiple tasks without retraining from scratch. |
| LLM (Large Language Model) | A foundation model specialized in language — generates and understands natural-language text. |
How LLMs work
| Term | Meaning |
|---|---|
| Token | The smallest unit of text the model processes — can be a word, part of a word or a symbol. |
| Transformer | The neural network architecture behind modern LLMs, based on the "attention" mechanism to relate tokens to each other. |
| Context window | The maximum number of tokens (input + output) the model can process in a single request. |
| Temperature | Controls the randomness of the output: low values = more deterministic responses; high values = more creative/varied. |
| Top-k / Top-p | Limit the set of candidate tokens at each generation step, tuning the balance between coherence and diversity. |
Generative AI modalities
Text
Natural-language generation and understanding: summaries, writing, Q&A, translation.
Image
Image generation from text (text-to-image) and AI-assisted editing.
Audio
Speech synthesis and transcription, music and sound-effect generation.
Multimodal
The same model processes and combines text, image, audio and video in a single request.
Risks and limitations
| Risk | Description |
|---|---|
| Hallucinations | The model generates information that looks correct but is factually false. |
| Bias | The model can reproduce biases present in its training data. |
| Outdated data | The model's knowledge has a cutoff date — it doesn't know about later events without grounding. |
| Misuse | Generation of harmful content, misinformation or impersonation when there are no safety filters. |
The exam distinguishes between "traditional" (discriminative: classifies/predicts) AI and generative AI (creates new content). Every LLM is a foundation model, but not every foundation model is an LLM (some are image, audio, or multimodal models).
GenAI options on Google Cloud
35% of the examAgent Platform (formerly Vertex AI): the central platform
Click any service to see more details
Complementary tools
| Service | Function |
|---|---|
| Imagen | A text-to-image generation and editing model, available on Agent Platform. |
| Gemini Code Assist | An AI coding assistant integrated into IDEs and Cloud Workstations. |
| Gemini for Google Workspace | Generative AI built into Gmail, Docs, Sheets and Meet for productivity. |
| Agent Search (formerly Vertex AI Search) | A managed semantic search and RAG engine over your own enterprise data. |
| NotebookLM / Gemini Enterprise | Tools aimed at business users for exploring and synthesizing information with generative AI. |
From prototype to production
| Stage | Tool |
|---|---|
| Explore and test prompts | Vertex AI Studio — a no-code interface for experimenting with models |
| Customize the model | Model tuning (supervised tuning, RLHF) on top of a base model |
| Connect your own data | Grounding with enterprise data, Agent Search, or RAG with Embeddings + Vector Search |
| Deploy to production | Managed Agent Platform endpoints with autoscaling and monitoring |
| Orchestrate complex flows | Agent Platform pipelines and agents for multi-step flows |
D2 carries the most weight (35%). The exam expects you to map a business scenario to the right Google Cloud tool: does it need a model (Gemini/Model Garden), an agent (Agent Platform), or just document extraction (Document AI)?
Techniques for improving GenAI model output
20% of the examPrompt engineering
| Technique | Description |
|---|---|
| Zero-shot | The model is asked to solve a task with no prior examples in the prompt. |
| Few-shot | Some input/output examples are included in the prompt to guide the response's format. |
| Chain-of-thought | The model is asked to reason step by step before giving the final answer — improves reasoning tasks. |
| System instructions | High-level instructions that set the model's role, tone or constraints for the whole conversation. |
Grounding and RAG
Grounding connects the model's responses to verifiable data sources, reducing hallucinations. RAG (Retrieval-Augmented Generation) is the most common technique for achieving it.
| RAG pipeline step | What it does |
|---|---|
| 1. Embeddings | Converts your own documents into semantic vectors |
| 2. Vector Search | Indexes those vectors for semantic similarity search |
| 3. Retrieval | Given a question, retrieves the most relevant data fragments |
| 4. Generation | The model generates the response using those fragments as context |
Fine-tuning
| Approach | When to use it |
|---|---|
| Prompt engineering | First option: low cost and effort, enough for most cases |
| Grounding / RAG | When the model needs up-to-date or organization-specific data |
| Supervised fine-tuning | When the model needs to consistently adopt a very specific format, tone or domain |
| RLHF (Reinforcement Learning from Human Feedback) | When you need to align the model's behavior with specific human preferences |
Increasing order of effort: prompt engineering → grounding/RAG → fine-tuning. The exam expects you to choose the lowest-effort technique that solves the problem, not to jump straight to fine-tuning.
Business strategies for successful GenAI solutions
15% of the examResponsible AI
| Principle | Application |
|---|---|
| Fairness | Evaluate the model to detect and mitigate bias against specific groups |
| Transparency | Document the capabilities, limitations and training data of the model (model cards) |
| Privacy | Protect sensitive data used as context or in training/tuning |
| Accountability | Maintain human oversight (human-in-the-loop) for high-impact decisions |
| Security | Harmful-content filters and guardrails before exposing the model to end users |
Evaluating the business case
| Criterion | Key question |
|---|---|
| Build vs. Buy | Is it better to use an already-built model/agent (Agent Platform, Model Garden) or develop your own? |
| TCO (Total Cost of Ownership) | Does the cost of inference, tuning and maintenance justify the value generated? |
| Change management | Is the organization and are its end users ready to adopt the AI-powered workflow? |
| Measuring impact | What business metrics (not just technical ones) define the success of the solution? |
Governance and operations (LLMOps)
LLMOps extends MLOps practices to foundation models: prompt versioning, continuous evaluation of response quality, cost monitoring, and detecting drift in model behavior.
The business domain (D4) evaluates judgment, not technical architecture: the exam rewards identifying when a GenAI project isn't ready due to lack of governance or change management, even when the technology is already available.
Tips for exam day
Generative AI vs. discriminative AI: discriminative AI classifies or predicts over existing data; generative AI creates new content (text, image, audio, code).
Gemini Enterprise Agent Platform (formerly Vertex AI) is the umbrella platform. Inside it live Gemini (models), Model Garden (multi-provider catalog), Agent Search (enterprise search and RAG) and agent building. The exam asks "which would you use" through scenarios.
Pro vs. Flash in Gemini: Pro for deep reasoning and long context, Flash for high volume and low latency/cost. It's the exam's most repeated tradeoff.
RAG = embeddings + vector search + grounding. It reduces hallucinations by connecting the model to verifiable data instead of relying only on its trained knowledge.
Prompt engineering (zero-shot, few-shot, chain-of-thought) is the lowest-effort technique. Fine-tuning takes more effort and cost, but is necessary when prompting alone isn't enough.
Responsible AI on the exam: fairness, transparency, privacy and accountability. Google Cloud offers safety filters and model evaluation built into Agent Platform.
Business strategy: evaluate build vs. buy, TCO and change management before the technology itself. The exam rewards reasoning about the business case, not just the technical architecture.
Want to go deeper?
This guide is a quick summary. Visit the Deep Dive section for an exhaustive walkthrough with real cases, cost analysis and architecture decisions.
Ready to practice?
Put what you learned to the test with original Gen AI Leader exam questions.
Start Gen AI Leader mock examUnlimited mock exams with the Pro plan
Was this guide helpful?
Not ready to start yet?
Leave your email and create your free account whenever you're ready — start with 15 Gen AI Leader questions, no credit card required.
Already have an account? Go straight to the mock exam →