Read this the night before or on exam morning. It assumes you’ve studied the notes. It’s here to refresh, not to teach.
Exam day
- 50 to 60 multiple-choice and multiple-select questions in 90 minutes. Fee $99. Online-proctored or at a test centre. English, Japanese and Spanish. Valid for 3 years. No prerequisites. Google does not publish a passing score.
- About 1.5 minutes per question. Flag hard ones and return. Answer everything; there is no stated penalty for guessing.
- No coding and no hands-on tasks. It is a business-level exam.
- Section weights (official guide, all approximate): fundamentals of gen AI 30%, Google Cloud’s gen AI offerings 35%, techniques to improve model output 20%, business strategies for a successful gen AI solution 15%.
- 2026 names: Vertex AI is Gemini Enterprise Agent Platform (Agent Platform), Vertex AI Studio is Agent Studio, Vertex AI Search is Agent Search, Agentspace is Gemini Enterprise. Old names still appear in older material.
How to read a question
Read the last sentence first, then find the problem or constraint word.
| The question says | Think |
|---|---|
| Wrong, made-up or outdated facts | Grounding or RAG, not a bigger model |
| Needs current world information | Grounding with Google Search |
| Needs answers from company documents | RAG / Agent Search over enterprise data |
| Inconsistent format or style | Few-shot examples, then fine-tuning if needed |
| No AI team, fast results | Prebuilt product |
| High-stakes decision | Human in the loop |
| More creative or varied output | Higher temperature or top-p |
| Factual, repeatable output | Low temperature |
| Reasoning goes wrong | Chain-of-thought or prompt chaining |
| Roll out across the company | Pilot first, measure, then scale |
Elimination tricks:
- “Train a new model from scratch” is almost never the answer for a business leader.
- “Raise the temperature” never fixes hallucinations.
- Answers that skip measurement, governance or human oversight are usually wrong.
- Pick the lowest-effort option that meets the need (prompting before RAG before fine-tuning).
Terms to define in one line
| Term | One line |
|---|---|
| Foundation model | Large, broadly trained model adapted to many tasks |
| LLM | Foundation model for language, predicts next tokens |
| Multimodal | Handles text, images, audio, video together |
| Diffusion model | Builds images by removing noise step by step |
| Prompt tuning | Trains a small add-on, base model frozen |
| Fine-tuning | Further trains the model on your labeled examples |
| Grounding | Ties answers to verifiable sources |
| RAG | Retrieve relevant content, add to prompt, generate |
| Hallucination | Fluent but false output |
| Context window | Max tokens the model can consider at once |
ML basics
| Learns from | Example | |
|---|---|---|
| Supervised | Labeled data | Spam filter |
| Unsupervised | Unlabeled data | Customer segments |
| Reinforcement | Rewards | Game-playing agent |
Lifecycle: data ingestion, data preparation, model training, model deployment, model management.
Model choice factors: modality, context window, security, availability, cost, performance, customization.
Layers: infrastructure, models, platforms, agents, applications.
Data quality: completeness, consistency, relevance, availability, cost, format.
Google models
| Model | Job |
|---|---|
| Gemini | Multimodal flagship, long context |
| Gemma | Open-weights, run it yourself |
| Imagen | Text to image |
| Veo | Video generation |
Product matching
| Need | Pick |
|---|---|
| Personal assistant, custom helpers (Gems) | Gemini app |
| AI in Gmail, Docs, Sheets, Meet | Gemini for Workspace |
| Search and agents over all company data | Gemini Enterprise |
| Customer chat or voice bot | Conversational Agents |
| Real-time help for human agents | Agent Assist |
| Analyze call transcripts | Conversational Insights |
| Full cloud contact centre | Contact Center as a Service |
| Site or app search with answers | Agent Search |
| Compare, tune and deploy models | Agent Platform, Model Garden |
| Custom model, labeled data, low code | AutoML |
| Fields from invoices and forms | Document AI |
| Speech to text, text to speech | Speech-to-Text, Text-to-Speech APIs |
| Quick prototype with API key | Google AI Studio |
| Enterprise prototype and governance | Agent Studio |
| Google-designed AI chips | TPUs |
Agents and RAG
- Agent tools: extensions (agent calls an API directly), functions (model proposes, your code runs it), data stores (retrieve your data), plugins (add-on capabilities).
- RAG effort: prebuilt with Agent Search (least), RAG APIs (medium), build your own (most).
- Grounding data: first-party (your data), third-party, world (Google Search).
Prompting and sampling
| Technique | Use |
|---|---|
| Zero-shot | No examples |
| One-shot, few-shot | One or several examples |
| Role prompting | Assign a persona |
| Prompt chaining | Multi-step pipeline of prompts |
| Chain-of-thought | Step-by-step reasoning |
| ReAct | Reason, act with tools, observe, repeat |
| Setting | Effect |
|---|---|
| Temperature | Low is deterministic, high is creative |
| Top-p | Samples from tokens up to cumulative probability p |
| Max output tokens | Caps length |
| Safety settings | Block harmful content categories |
Secure and responsible AI
- SAIF (six elements): strong security foundations, extend detection and response, automate defenses, harmonize platform controls, adapt controls with fast feedback, contextualize risks in business processes.
- Tools: IAM (who can do what), Security Command Center (posture and threats), secure-by-design infrastructure, logging and monitoring.
- Threats: prompt injection, data poisoning, data leakage, model theft.
- Responsible AI: transparency, privacy, fairness, accountability, explainability.
- Anonymization cannot be reversed. Pseudonymization can be re-linked with a key.
- Measuring impact: baseline first, then ROI, productivity, quality, customer satisfaction, risk.