General Exam Tips
- 1.This is a business/leadership exam, not a technical one — there is no coding, no configuration, and no hands-on labs. Read questions for business framing, not implementation detail
- 2.You get 90 minutes for 50-60 questions — roughly 90 seconds each. This is generous; do not rush, and re-read multiple-select questions to see how many answers are expected
- 3.There is no penalty for wrong answers. Never leave a question blank — always select your best guess before moving on
- 4.Product-recognition questions dominate Domain 2 (35%). If you can't recall a product's exact name, match on FUNCTION first (customer-facing vs internal-productivity vs enterprise-search) and the name usually follows
- 5.When a scenario names a business problem (hallucinations, outdated answers, inconsistent style, needs review before publishing), translate it to a technique: hallucinations/outdated → grounding/RAG; style/behavior → fine-tuning; wording/format → prompt engineering; high-stakes → HITL
- 6.Watch for near-identical product names used as distractors: Gemini app vs Gemini Advanced vs Gemini Enterprise vs Gemini for Workspace; Google AI Studio vs Agent Studio; Conversational Agents vs Agent Assist. The exam deliberately pairs these
- 7.Domains 1 and 2 together are 65% of the exam (fundamentals + product catalog) — prioritize vocabulary and product recognition over deep technique nuance if time is short
- 8.Diffusion models generate images/video; LLMs generate text. If a question's modality is video, the Google answer is almost always Veo; if it's images, Imagen; if it's open/self-hosted, Gemma; if it's general multimodal at scale, Gemini
- 9.Responsible AI (ethics: transparency, privacy, fairness, accountability, explainability) and SAIF (security: protecting AI from attacks/misuse) are frequently confused as the same topic — they are complementary but answer different question types
Quick Navigation
Fundamentals of Generative AI
Must-Know Facts
- AI is the broadest term; ML is a subset that learns from data; generative AI is a subset of ML that creates new content. Foundation models are large pre-trained models adaptable to many tasks; LLMs are text-focused foundation models
- Prompt engineering (craft the input, no model change) is distinct from prompt tuning (a lightweight training technique that learns soft prompts and does change model parameters)
- The three ML approaches: supervised (labeled data), unsupervised (unlabeled data, finds clusters/patterns), reinforcement learning (reward/penalty feedback from an environment, no labels)
- The ML lifecycle: data ingestion → data preparation → model training → model deployment → model management, each with corresponding Google Cloud tooling
- Choosing a foundation model is a BUSINESS decision weighing modality, context window, security, availability/reliability, cost, performance, and customization needs — not just raw accuracy
- The create/summarize/discover/automate framework categorizes gen AI business use cases
- Data concepts: structured (schema, tables) vs unstructured (documents, images, free text — where gen AI adds the most value) is one axis; labeled vs unlabeled is a SEPARATE, independent axis
- Five gen AI landscape layers, bottom to top: infrastructure → models → platforms → agents → applications
- Google's foundation model family and their modality: Gemini (multimodal flagship), Gemma (open/lightweight text models), Imagen (image generation), Veo (video generation)
Common Traps
Confusing Pairs
Scenario Tips
A question describes clustering unlabeled customer support tickets into groups with no predefined categories...
Unsupervised learning — it finds structure/patterns in unlabeled data without predefined output categories
Supervised learning requires labeled examples upfront, which the scenario explicitly lacks. Reinforcement learning requires a reward signal and sequential actions, not present here
A question asks which gen AI landscape layer contains custom-designed TPUs, GPUs, and data centers...
The infrastructure layer — the bottom layer providing the compute/storage/networking foundation for everything above it
The models layer refers to the foundation models themselves (Gemini, Gemma), not the physical hardware that trains and serves them
A question describes condensing thousands of documents into a short executive brief...
Summarize — condensing large volumes of content into concise insights, per the create/summarize/discover/automate framework
Create implies generating brand-new original content, not condensing existing content. Discover implies finding/searching for information, not summarizing it
A question asks which model to use for generating a short promotional video from a product description...
Veo — Google's dedicated text-to-video model
Imagen only generates still images. Gemini and Gemma are general multimodal/text models, not purpose-built video generators
Last-Minute Facts
Google Cloud's Generative AI Offerings
Must-Know Facts
- Google's four differentiators: AI-first approach, enterprise-ready platform (responsible/secure/private/reliable/scalable), integrated AI ecosystem across Google products, and an open approach (first-party + open + third-party models)
- AI Hypercomputer combines custom TPUs, GPUs, data centers, and software — the infrastructure behind Google's scale and efficiency
- Gemini has FOUR distinct surfaces: the consumer Gemini app, Gemini Advanced (with Gems), Gemini for Google Workspace (productivity apps), and Gemini Enterprise (enterprise search + custom agents)
- Gems are user-created customizable personas within the Gemini app for RECURRING personal tasks — not full custom agents and not an enterprise offering
- Gemini Enterprise provides multimodal enterprise search, NotebookLM-style research, and custom agent building grounded in a company's first-party data
- Customer Engagement Suite components: Conversational Agents (talk directly to customers), Agent Assist (helps human agents in real time), Conversational Insights (analytics), and CCaaS (contact-center infrastructure)
- Agent Platform developer building blocks: Model Garden (model catalog), Agent Search (enterprise/external search), AutoML (low-code model training), plus RAG offerings and agent tooling
- Agent tooling categories: extensions/functions (call external APIs), data stores (grounding data), plugins (extend capability) — how an agent takes action beyond text generation
- Pre-built AI APIs map to specific tasks: Document AI (parse documents), Cloud Vision (analyze images), Video Intelligence (analyze video), Natural Language (sentiment/entities in text), Speech-to-Text/Text-to-Speech (audio conversion), Translation (language conversion)
- Google AI Studio = fast prototyping with Gemini in browser. Agent Studio = enterprise environment for building/deploying PRODUCTION agents with governance
Common Traps
Confusing Pairs
Scenario Tips
An enterprise wants employees to search internal documents in natural language and build custom agents grounded in that data...
Gemini Enterprise — purpose-built for enterprise multimodal search and custom agents over first-party data
Gemini for Workspace only helps with productivity tasks inside office apps; it is not an enterprise search or agent-building platform
A retailer wants a 24/7 chatbot that answers customer questions directly on their website...
Conversational Agents (Customer Engagement Suite) — a customer-facing virtual agent
Agent Assist only supports human agents; it does not talk to customers on its own
A developer wants to quickly test prompt variations against Gemini before committing to a full build...
Google AI Studio — fast, browser-based prototyping
Agent Studio is for building and deploying production-grade enterprise agents, which is heavier weight than quick experimentation
A company needs to extract invoice number, date, and total from scanned PDF invoices...
Document AI — purpose-built for structured extraction from documents
Cloud Vision analyzes images generally but is not optimized for document field extraction; Natural Language analyzes text sentiment/entities, not document layout
Last-Minute Facts
Techniques to Improve Generative AI Model Output
Must-Know Facts
- Foundation model limitations: data dependency, knowledge cutoff, bias/fairness issues, hallucinations, and edge cases — name the limitation before picking the mitigation
- Google-recommended mitigations map 1:1 to limitations: grounding/RAG for hallucinations and knowledge cutoff, prompt engineering for wording/format, fine-tuning for specialized style/behavior, HITL for high-stakes decisions
- Grounding connects model output to trusted data; RAG is a specific grounding METHOD that retrieves relevant documents at query time before the model answers
- Three grounding data sources: first-party enterprise data (a company's own systems), third-party data (licensed/partner), and world data (broad public knowledge via Grounding with Google Search)
- Google Cloud grounding offerings: prebuilt RAG with Agent Search, RAG APIs, and Grounding with Google Search
- Prompt technique count of examples: zero-shot = 0 examples, one-shot = exactly 1, few-shot = 2+. Role prompting assigns a persona. Prompt chaining links sequential prompts
- Chain-of-thought = step-by-step REASONING only. ReAct = reasoning PLUS taking actions with external tools, interleaved
- Sampling parameters: temperature (low = deterministic/factual, high = creative/varied), top-p (nucleus sampling cutoff), token count (output length only), safety settings (content filtering)
- Continuous monitoring/evaluation practices: automatic model upgrades, KPI tracking, security patches, versioning, drift monitoring, and Feature Store for consistent feature serving
Common Traps
Confusing Pairs
Scenario Tips
A support bot gives outdated product info and invents products that don't exist...
Ground the model with RAG over the current, authoritative product catalog — addresses both knowledge cutoff and hallucination
Raising temperature increases randomness and would worsen hallucinations; fine-tuning on style does not fix factual accuracy
A prompt includes three labeled example tickets before asking the model to classify a new one...
Few-shot prompting — 3 examples is 2+, which is few-shot by definition
One-shot would require exactly one example; the scenario explicitly has three
A financial firm requires a person to review every AI-drafted investment recommendation before it reaches a client...
Human-in-the-loop (HITL) — mandatory human review before high-stakes output is used
Prompt chaining links prompts together but does not describe human review; it is a technique, not a governance control
A team wants highly consistent, repeatable answers for a factual FAQ bot...
Set a LOW temperature to reduce randomness and increase determinism
High temperature increases creativity/variance, which is the opposite of what a consistent FAQ bot needs
Last-Minute Facts
Business Strategies for a Successful Generative AI Solution
Must-Know Facts
- Four steps to implement a transformational gen AI solution: identify the need → choose the right solution (business requirements + technical constraints) → integrate into the organization → measure impact
- Measuring impact means defining KPIs and ROI tied to BUSINESS outcomes (cost savings, resolution time, customer satisfaction, adoption), not just technical metrics like model accuracy
- Secure AI must be applied across the full ML lifecycle — data, models, and infrastructure — from ingestion through deployment and ongoing management, not bolted on at the end
- The Secure AI Framework (SAIF) is Google's framework for securing AI systems against attacks/misuse across the lifecycle — a SECURITY framework, distinct from Responsible AI's ethics focus
- Google Cloud security tools: secure-by-design infrastructure, IAM (least-privilege access control), Security Command Center (threat/misconfiguration detection), and workload monitoring
- Five Responsible AI principles: transparency (open about how/when AI is used), privacy (protect sensitive data), fairness (avoid biased outcomes), accountability (own AI decisions), explainability (understand why a model decided what it did)
- Privacy practices: anonymization (irreversible de-identification) and pseudonymization (reversible tokenization) protect sensitive data used in gen AI systems
- Data quality, bias, and fairness directly affect AI outcomes — poor or unrepresentative data produces biased results, which undermines accountability and trust
Common Traps
Confusing Pairs
Scenario Tips
A healthcare organization must protect patient identities but still needs to re-link records internally when authorized...
Pseudonymization — replaces identifiers with reversible tokens, allowing authorized re-linking
Full anonymization removes identifiers irreversibly, which would prevent the required internal re-linking
An organization wants a structured framework specifically to protect its AI systems from threats and misuse...
The Secure AI Framework (SAIF) — Google's security framework for AI across the lifecycle
Responsible AI principles address fairness and ethics, not primarily technical security defense against attacks
Regulators ask a company to explain why its gen AI hiring tool rejected specific candidates...
Explainability — the ability to articulate why a specific decision was made
Transparency is about general openness regarding AI use, not justifying a specific individual decision
Leadership wants to demonstrate the value of a new gen AI customer-support assistant to executives...
Define KPIs tied to business outcomes (resolution time, cost savings, satisfaction) and track them over time
Reporting only training accuracy or prompt volume are technical/activity metrics that do not demonstrate business ROI