CertPrepNow
Google CloudGenerative AI Leader4 domains

Generative AI Leader Exam Notes

Last-minute traps, must-know facts, and scenario tips for the Google Cloud Certified – Generative AI Leader exam.

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  • 5.When a scenario names a business problem (hallucinations, outdated answers, inconsistent style, needs review before publishing), translate it to a technique: hallucinations/outdated → grounding/RAG; style/behavior → fine-tuning; wording/format → prompt engineering; high-stakes → HITL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
Domain 130% of exam

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

TrapTreating AI, ML, and generative AI as interchangeable synonyms
RealityThey form a strict hierarchy: AI ⊃ ML ⊃ Generative AI. Generative AI is always a subset of machine learning, and machine learning is always a subset of artificial intelligence. A question testing 'which is broadest' expects AI
TrapAssuming structured data cannot be unlabeled, or unstructured data cannot be labeled
RealityStructured/unstructured and labeled/unlabeled are two INDEPENDENT axes. A spreadsheet (structured) can lack target labels; a set of images (unstructured) can be fully labeled with categories. Do not conflate the two axes
TrapConfusing prompt tuning with prompt engineering because the names sound alike
RealityPrompt engineering only changes the input text — no training occurs. Prompt tuning is a lightweight TRAINING technique that learns 'soft prompts,' meaning it does adapt model parameters. The similar names are a deliberate exam trap
TrapPicking an LLM or Gemini for an image- or video-generation scenario
RealityDiffusion models (Imagen for images, Veo for video) generate visual media by denoising; LLMs and general Gemini text tasks generate text. Match the model family to the OUTPUT modality described in the scenario
TrapChoosing a foundation model based only on which one seems most powerful or most 'advanced'
RealityThe exam frames model selection as a leadership trade-off across modality, context window, security, cost, availability, performance, and customization needs — the 'best' model depends on the business constraint stated in the scenario

Confusing Pairs

Prompt EngineeringPrompt Tuning

Prompt engineering = craft better input text, zero training, instant, no data needed. Prompt tuning = a lightweight training method that learns soft prompt parameters — technically a form of model adaptation, unlike prompt engineering

Structured vs Unstructured DataLabeled vs Unlabeled Data

Structured/unstructured describes the FORMAT (schema vs free-form). Labeled/unlabeled describes whether examples have tagged outcomes. These are orthogonal — any combination of the two axes is possible

ImagenVeo

Imagen = text-to-IMAGE (still images). Veo = text-to-VIDEO (moving images with time dimension). Both are diffusion-based generative models but differ strictly by output modality

GeminiGemma

Gemini = proprietary, managed, flagship multimodal model family with the highest capability. Gemma = open, lightweight, self-hostable models derived from Gemini research. Openness/self-hosting → Gemma; maximum capability/managed convenience → Gemini

Scenario Tips

If the question asks about:

A question describes clustering unlabeled customer support tickets into groups with no predefined categories...

Answer:

Unsupervised learning — it finds structure/patterns in unlabeled data without predefined output categories

Distractor to avoid:

Supervised learning requires labeled examples upfront, which the scenario explicitly lacks. Reinforcement learning requires a reward signal and sequential actions, not present here

If the question asks about:

A question asks which gen AI landscape layer contains custom-designed TPUs, GPUs, and data centers...

Answer:

The infrastructure layer — the bottom layer providing the compute/storage/networking foundation for everything above it

Distractor to avoid:

The models layer refers to the foundation models themselves (Gemini, Gemma), not the physical hardware that trains and serves them

If the question asks about:

A question describes condensing thousands of documents into a short executive brief...

Answer:

Summarize — condensing large volumes of content into concise insights, per the create/summarize/discover/automate framework

Distractor to avoid:

Create implies generating brand-new original content, not condensing existing content. Discover implies finding/searching for information, not summarizing it

If the question asks about:

A question asks which model to use for generating a short promotional video from a product description...

Answer:

Veo — Google's dedicated text-to-video model

Distractor to avoid:

Imagen only generates still images. Gemini and Gemma are general multimodal/text models, not purpose-built video generators

Last-Minute Facts

1AI ⊃ ML ⊃ Generative AI — strict nesting, always in that order
2Gemini = multimodal flagship (proprietary). Gemma = open/lightweight. Imagen = image. Veo = video
3Prompt engineering = no training, changes input only. Prompt tuning = lightweight training, learns soft prompts
4Five landscape layers bottom-to-top: infrastructure → models → platforms → agents → applications
5Structured/unstructured and labeled/unlabeled are independent axes — never conflate them
6create / summarize / discover / automate is Google's gen AI use case framework
Domain 235% of exam

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

TrapChoosing Gemini for Workspace when a scenario asks for enterprise search or custom agent building over company data
RealityGemini for Workspace only embeds gen AI inside office apps (Gmail, Docs, Sheets, Slides, Meet) for productivity. Enterprise search and custom agents grounded in first-party data are Gemini Enterprise's job — this exact swap is the single most common Domain 2 trap
TrapPicking Conversational Agents when the scenario describes helping a HUMAN employee, or Agent Assist when the scenario describes a customer talking directly to AI
RealityConversational Agents = customer-facing (the AI talks to the customer). Agent Assist = employee-facing (the AI helps a human who talks to the customer). The direction of the conversation determines the answer
TrapUsing Google AI Studio and Agent Studio interchangeably because both have 'Studio' in the name
RealityGoogle AI Studio is for QUICK PROTOTYPING and experimentation with Gemini. Agent Studio is the ENTERPRISE, PRODUCTION-GRADE environment on the Agent Platform with governance controls for deploying agents at scale
TrapThinking Model Garden is a single model or a specific Google model
RealityModel Garden is a CATALOG containing first-party (Gemini, Gemma, Imagen, Veo), open-source, and third-party models. It represents Google's open, choice-driven approach — it is a discovery/deployment hub, not a model itself
TrapMatching the wrong pre-built API to a document-parsing task (e.g., choosing Cloud Vision for extracting invoice fields)
RealityDocument AI is purpose-built for extracting structured fields from documents like invoices and forms. Cloud Vision handles general image analysis (labels, objects) but is not optimized for document field extraction

Confusing Pairs

Gemini EnterpriseGemini for Google Workspace

Enterprise = search, research, and custom agents over a company's OWN data. Workspace = gen AI features embedded INSIDE office productivity apps (Gmail, Docs, Sheets, Slides, Meet). The exam tests this pairing directly and often

Conversational AgentsAgent Assist

Conversational Agents talk directly TO customers (chatbots/voicebots). Agent Assist helps a HUMAN agent serve the customer (real-time suggestions, knowledge surfacing). Both live in the Customer Engagement Suite but serve opposite audiences

Google AI StudioAgent Studio

Google AI Studio = fast browser-based prompt/app prototyping, individual experimentation. Agent Studio = enterprise agent building, testing, and PRODUCTION deployment with governance on the Agent Platform

Model GardenAgent Search

Model Garden is a CATALOG of models (first-party/open/third-party) for discovery and deployment. Agent Search is an enterprise SEARCH capability that retrieves grounded answers from company (and optionally external) data — different purposes entirely

Document AICloud VisionNatural Language API

Document AI parses/extracts structured data from documents (invoices, forms). Cloud Vision analyzes general images (labels, objects, faces). Natural Language analyzes TEXT for sentiment, entities, and syntax. Match the input type and extraction goal to the right API

Scenario Tips

If the question asks about:

An enterprise wants employees to search internal documents in natural language and build custom agents grounded in that data...

Answer:

Gemini Enterprise — purpose-built for enterprise multimodal search and custom agents over first-party data

Distractor to avoid:

Gemini for Workspace only helps with productivity tasks inside office apps; it is not an enterprise search or agent-building platform

If the question asks about:

A retailer wants a 24/7 chatbot that answers customer questions directly on their website...

Answer:

Conversational Agents (Customer Engagement Suite) — a customer-facing virtual agent

Distractor to avoid:

Agent Assist only supports human agents; it does not talk to customers on its own

If the question asks about:

A developer wants to quickly test prompt variations against Gemini before committing to a full build...

Answer:

Google AI Studio — fast, browser-based prototyping

Distractor to avoid:

Agent Studio is for building and deploying production-grade enterprise agents, which is heavier weight than quick experimentation

If the question asks about:

A company needs to extract invoice number, date, and total from scanned PDF invoices...

Answer:

Document AI — purpose-built for structured extraction from documents

Distractor to avoid:

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

1Four Gemini surfaces: app, Advanced/Gems, Workspace (productivity), Enterprise (search + custom agents)
2Workspace = productivity in office apps. Enterprise = search + custom agents over company data — do not swap
3Conversational Agents = talks to customers. Agent Assist = helps human agents
4Google AI Studio = prototyping. Agent Studio = enterprise production deployment
5Model Garden = catalog of first-party + open + third-party models, not a single model
6Document AI = documents/invoices. Cloud Vision = images. Video Intelligence = video. Natural Language = text sentiment/entities
Domain 320% of exam

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

TrapReaching for fine-tuning to fix hallucinations or outdated answers
RealityFine-tuning changes STYLE/BEHAVIOR — it does not reliably add new facts or guarantee accuracy. Hallucinations and outdated knowledge are fixed by grounding/RAG, which connects the model to current, trusted data at query time
TrapMiscounting prompt examples and calling a 2-example prompt 'one-shot'
RealityZero-shot = 0 examples. One-shot = exactly 1. Few-shot = 2 or more. Read the scenario carefully and count the literal number of examples given in the prompt
TrapConfusing chain-of-thought with ReAct because both involve 'reasoning'
RealityChain-of-thought only makes the model think step by step INTERNALLY — no external tool use. ReAct interleaves reasoning WITH actions (calling tools, searching, calculating) — it is reasoning plus acting
TrapBelieving higher temperature is always 'better' output quality
RealityHigher temperature increases randomness/creativity but REDUCES consistency and factual reliability. For a factual FAQ bot or deterministic task, LOW temperature is correct; high temperature suits creative brainstorming
TrapAssuming grounding and fine-tuning solve the same problem and picking either interchangeably
RealityGrounding changes what FACTS a model can access (query-time retrieval, no retraining). Fine-tuning changes how the model BEHAVES or its style (via retraining on niche data). They solve different problems and are not substitutes

Confusing Pairs

Grounding / RAGFine-Tuning

Grounding/RAG gives the model access to trusted, current facts at query time — fixes hallucinations and knowledge cutoff. Fine-tuning retrains the model to change behavior/style — does not keep the model current with new facts

Chain-of-ThoughtReAct

Chain-of-thought = internal step-by-step reasoning only, no tools. ReAct = reasoning interleaved with ACTING (using tools/APIs/search) to gather information or take real steps mid-response

Zero-ShotOne-ShotFew-Shot

Distinguished purely by the NUMBER of examples in the prompt: 0 = zero-shot, exactly 1 = one-shot, 2+ = few-shot. All three provide an instruction; only the example count differs

TemperatureTop-P

Temperature scales the overall randomness of token probabilities (low = deterministic, high = creative). Top-p (nucleus sampling) restricts choices to the smallest set of tokens whose cumulative probability exceeds P, trimming unlikely options. Both control diversity but via different mechanisms

Grounding with Google SearchPrebuilt RAG with Agent Search

Grounding with Google Search taps WORLD data (the public web) for current facts. Prebuilt RAG with Agent Search retrieves from an organization's FIRST-PARTY enterprise data. Match the data source (public vs internal) to the offering

Scenario Tips

If the question asks about:

A support bot gives outdated product info and invents products that don't exist...

Answer:

Ground the model with RAG over the current, authoritative product catalog — addresses both knowledge cutoff and hallucination

Distractor to avoid:

Raising temperature increases randomness and would worsen hallucinations; fine-tuning on style does not fix factual accuracy

If the question asks about:

A prompt includes three labeled example tickets before asking the model to classify a new one...

Answer:

Few-shot prompting — 3 examples is 2+, which is few-shot by definition

Distractor to avoid:

One-shot would require exactly one example; the scenario explicitly has three

If the question asks about:

A financial firm requires a person to review every AI-drafted investment recommendation before it reaches a client...

Answer:

Human-in-the-loop (HITL) — mandatory human review before high-stakes output is used

Distractor to avoid:

Prompt chaining links prompts together but does not describe human review; it is a technique, not a governance control

If the question asks about:

A team wants highly consistent, repeatable answers for a factual FAQ bot...

Answer:

Set a LOW temperature to reduce randomness and increase determinism

Distractor to avoid:

High temperature increases creativity/variance, which is the opposite of what a consistent FAQ bot needs

Last-Minute Facts

1Hallucinations/outdated facts → grounding/RAG. Style/behavior → fine-tuning. Wording/format → prompt engineering. High-stakes → HITL
2Zero-shot = 0 examples. One-shot = 1. Few-shot = 2+
3Chain-of-thought = reason only. ReAct = reason + act with tools
4Low temperature = deterministic/factual. High temperature = creative/varied
5Three grounding sources: first-party enterprise, third-party, world data
6Google grounding offerings: prebuilt RAG with Agent Search, RAG APIs, Grounding with Google Search
Domain 415% of exam

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

TrapTreating SAIF and Responsible AI as the same initiative because both relate to 'safe AI'
RealitySAIF is a SECURITY framework protecting AI systems from attacks, misuse, and technical threats. Responsible AI is an ETHICS framework covering fairness, transparency, privacy, accountability, and explainability. They are complementary but answer different question types — security threat vs ethical/fairness concern
TrapMeasuring gen AI success using only technical metrics like model accuracy or number of prompts sent
RealityThe exam expects BUSINESS KPIs and ROI — cost savings, resolution time, customer satisfaction, adoption rate — tied to organizational objectives. Technical accuracy alone does not demonstrate business value to leadership
TrapAssuming anonymization and pseudonymization are interchangeable privacy techniques
RealityAnonymization irreversibly strips identifying data — it can NEVER be re-linked to the individual. Pseudonymization replaces identifiers with reversible tokens — authorized parties CAN re-link the data later. Choose pseudonymization when future re-linking is required (e.g., internal record matching); choose anonymization when it is not
TrapConfusing transparency with explainability because both relate to 'openness'
RealityTransparency = being open about HOW and WHEN AI is used in a process generally. Explainability = the ability to articulate WHY a specific model decision was made. A hiring-tool scenario asking 'why was this candidate rejected' tests explainability, not transparency
TrapChoosing a gen AI solution based purely on which is most innovative or technically impressive
RealityThe exam frames solution choice as matching BUSINESS REQUIREMENTS and TECHNICAL CONSTRAINTS — cost, timeline, compliance, existing systems — not novelty. A leader picks the solution that fits the organization's actual needs and limits

Confusing Pairs

Secure AI Framework (SAIF)Responsible AI Principles

SAIF = securing AI against attacks/misuse (a security discipline). Responsible AI = ensuring AI is fair, transparent, accountable, and explainable (an ethics discipline). Security threat scenarios → SAIF. Fairness/ethics scenarios → Responsible AI

AnonymizationPseudonymization

Anonymization = irreversible removal of identifiers, cannot re-link. Pseudonymization = reversible tokenization, CAN re-link with authorization. If the scenario needs future re-linking for legitimate purposes, pseudonymization is correct

TransparencyExplainability

Transparency = openness about the fact that AI is being used and how it generally works. Explainability = the specific ability to justify WHY a model produced a particular output/decision. Explainability is decision-level; transparency is process-level

Technical MetricsBusiness KPIs

Technical metrics (model accuracy, latency, training loss) measure engineering quality. Business KPIs (ROI, cost savings, resolution time, customer satisfaction, adoption) measure organizational value. Leadership-facing impact measurement always favors business KPIs

Scenario Tips

If the question asks about:

A healthcare organization must protect patient identities but still needs to re-link records internally when authorized...

Answer:

Pseudonymization — replaces identifiers with reversible tokens, allowing authorized re-linking

Distractor to avoid:

Full anonymization removes identifiers irreversibly, which would prevent the required internal re-linking

If the question asks about:

An organization wants a structured framework specifically to protect its AI systems from threats and misuse...

Answer:

The Secure AI Framework (SAIF) — Google's security framework for AI across the lifecycle

Distractor to avoid:

Responsible AI principles address fairness and ethics, not primarily technical security defense against attacks

If the question asks about:

Regulators ask a company to explain why its gen AI hiring tool rejected specific candidates...

Answer:

Explainability — the ability to articulate why a specific decision was made

Distractor to avoid:

Transparency is about general openness regarding AI use, not justifying a specific individual decision

If the question asks about:

Leadership wants to demonstrate the value of a new gen AI customer-support assistant to executives...

Answer:

Define KPIs tied to business outcomes (resolution time, cost savings, satisfaction) and track them over time

Distractor to avoid:

Reporting only training accuracy or prompt volume are technical/activity metrics that do not demonstrate business ROI

Last-Minute Facts

1SAIF = security (attacks/misuse). Responsible AI = ethics (fairness/transparency/privacy/accountability/explainability)
2Anonymization = irreversible. Pseudonymization = reversible tokens
3Transparency = general openness about AI use. Explainability = justifying a specific decision
4Measure impact with business KPIs/ROI, not just technical accuracy
5Four implementation steps: identify need → choose solution → integrate → measure impact
6Security tools: IAM (access control), Security Command Center (threat detection), secure-by-design infrastructure

Feeling confident?

Put your knowledge to the test with a timed Generative AI Leader mock exam.