Generative AI Leader · CHEAT SHEET

Google Cloud Certified: Generative AI Leader

Key concepts grouped by official exam weight.

Download as PDF (free, no sign-up)

1. Fundamentals of gen AI

30–30%

Describe core generative AI (gen AI) concepts and use cases

AI is the broad field; machine learning learns patterns from data and NLP lets computers understand and generate language. Generative AI uses foundation models, pretrained on huge datasets and adaptable to many tasks: LLMs focus on text, multimodal models combine text, images, audio and video, and diffusion models create images by removing noise. Prompt engineering steers outputs through the prompt; prompt tuning learns a small prefix.
Supervised learning trains on labeled data to classify or predict values, such as the risk of a loan default. Unsupervised learning discovers patterns in unlabeled data, for example to segment customers. Reinforcement learning learns by trial and error from rewards and penalties, like a robot learning to grasp objects. Semi-supervised learning combines a little labeled data with a lot of unlabeled data.
The ML lifecycle runs from data ingestion (Pub/Sub for event streams) to data preparation (BigQuery, Dataflow), model training (code-free AutoML or custom training on Gemini Enterprise Agent Platform), deployment to endpoints for online or batch inference, and model management: Model Registry for versions, Model Monitoring for drift and Pipelines to automate the workflow.
Choosing a foundation model starts from the use case: modality (text, image, video or audio), the context window needed, quality, latency and cost, security and compliance, and whether it can be tuned with your own data. A bigger model isn't always better; Model Garden brings together Google, partner and open models so they can be compared.
Gen AI excels at creating content (text, images, video and code), summarizing documents and feedback, discovering information through conversational search and automating repetitive tasks. Forecasting or classification on structured data is usually a job for traditional AI, and the two combine well: traditional AI segments customers and gen AI creates personalized content for each segment.

Describe how various data types are used in gen AI and the business implications

AI depends on data quality: data must be complete, consistent across systems, valid, fresh and relevant to real deployment conditions, or bias and hallucinations follow. It must also be accessible: silos and scattered formats stall projects, and Google recommends a unified, governed platform rather than copying everything to one place, which is often costly and impractical.
Structured data follows a fixed schema of rows and columns and is queried with SQL, such as inventory or financial records. Unstructured data has no predefined model: emails, PDFs, images, audio, video or satellite imagery, and it makes up most of a company's data. JSON and XML are semi-structured. Google recommends BigQuery for structured data and Cloud Storage for unstructured data.
Labeled data pairs each example with its correct answer (a category, a sentiment or a value) and underpins supervised learning and supervised tuning; unlabeled data has no such answer and feeds unsupervised learning and the self-supervised pre-training of foundation models. Labeling is costly, so teams combine people and automation, with clear guidelines and quality control.

Identify the core layers of the gen AI landscape and the business implications

The infrastructure layer supplies the compute, storage and networking used to train and serve models. On Google Cloud it is grouped under AI Hypercomputer, which integrates accelerators (NVIDIA GPUs and Google-designed TPUs), open software and flexible consumption options, such as Spot VMs or reservations, to maximize goodput and keep costs under control.
The models layer is made up of foundation models: large models pretrained on massive data that can be adapted to many tasks. LLMs are one type of foundation model; others work with images, audio or video. Models differ in size, modality and cost, so choosing one is a per-use-case trade-off, and adapting a pretrained model is usually faster and cheaper than training from scratch.
The platforms layer connects models, data and agents and hides the complexity of infrastructure. On Google Cloud it is Gemini Enterprise Agent Platform, the evolution of Vertex AI, organized into four pillars: build (Agent Studio, Agent Development Kit, Model Garden), scale (Agent Runtime), govern (Agent Registry, Agent Identity, Agent Gateway) and optimize (evaluation, simulation and observability).
An AI agent pursues a goal autonomously: it uses a model as its reasoning engine, plans, remembers and acts on external systems through tools. Unlike an assistant, it does not wait for the user to decide every step, and unlike a bot, it does not follow fixed rules. Google notes that deterministic tasks such as summarizing or classifying do not need an agentic workflow.
The applications layer is the part people interact with. It includes AI built into everyday tools, such as Gemini in Gmail or Docs, and the Gemini Enterprise app, a single front door to search company data with permissions, chat and use agents. It delivers value quickly without in-house development, and Google states that company data is not used to train its models.

Identify the use cases and strengths of Google’s foundation models

Gemini is Google's family of most capable multimodal models, built by Google DeepMind: it understands text, images, audio, video and code in a single request and reasons across them. Its long context window lets it analyze large volumes of documents or video without splitting them, and it is the reasoning engine behind assistants and agents such as Gemini Enterprise. It is used as a managed service on Agent Platform.
Gemma is a family of lightweight open models built from the same research as Gemini. Because the weights are open, it can be tuned with any framework and run on your own hardware, mobile devices or hosted services. It has specialized variants such as MedGemma (medical), ShieldGemma 2 (safety) and EmbeddingGemma (search), and Google recommends Agent Platform to serve it without an MLOps team.
Imagen is Google's specialized text-to-image model. Its strengths are image quality, photorealism, artistic styles, typography and brand, logo and product designs, with low latency; it also supports mask-based editing and upscaling. For multi-turn conversational editing Google recommends Gemini instead, and content generated with Imagen carries the invisible SynthID watermark.
Veo is Google DeepMind's model for generating video from text or images, across a wide range of visual and cinematic styles. It can animate existing photos, create the footage between a first and a last frame, extend and edit clips and generate native audio with dialogue and sound effects. Companies use it to speed up ads and marketing content, and its videos carry the SynthID watermark.

2. Google Cloud’s gen AI offerings

35–35%

Describe Google Cloud's strengths in the field of gen AI

Google describes itself as an AI-first company with a full-stack approach: its own infrastructure with TPUs and GPUs, research from Google Research and Google DeepMind, the natively multimodal Gemini model family, and products that bring it to millions of users. That integration, guided by its AI Principles, lets research advances reach enterprise solutions faster.
Google Cloud presents its AI platform as enterprise-ready: customer prompts, responses and tuning data are not used to train models without permission, it offers IP indemnity, safety filters, Model Armor and Sensitive Data Protection, grounding reduces hallucinations, and everything runs on reliable, scalable global infrastructure.
Gemini is built into Google's whole ecosystem: Google Workspace (Gmail, Docs, Meet, Vids), Gemini in BigQuery and Looker for data, Gemini Code Assist and Gemini Cloud Assist for technical teams, Gemini in Security Command Center for security, and Gemini Enterprise as a single front door that connects to Workspace, Microsoft 365 or Salesforce data. AI reaches the tools each team already uses.
Google Cloud's open approach gives freedom of choice: more than 200 models from Google, open models such as Gemma and third-party ones such as Claude, swappable without rebuilding agent logic; open standards such as MCP to connect agents to systems and A2A so agents from different vendors can collaborate; and open technologies such as Kubernetes for multicloud portability. This reduces dependence on a single vendor.
AI Hypercomputer is Google Cloud's integrated supercomputing system: it combines optimized hardware (Google-designed TPUs, NVIDIA GPUs, networking and storage), open software such as PyTorch and JAX, and flexible consumption models such as Spot VMs or reservations to maximize goodput. TPUs are custom ASICs for large-scale matrix computation, running in efficient global data centers.

Describe how Google Cloud’s prebuilt gen AI offerings enable AI powered work

The Gemini app helps people draft, summarize and analyze, and includes Deep Research for reports built from hundreds of sources and Canvas for creating prototypes and text. Gems are saved versions of Gemini with their own instructions and files that save time on repetitive tasks; with Drive files they always use the latest version, and when shared, viewers can see the instructions and files.
Gemini Enterprise (formerly Agentspace) combines permissions-aware multimodal search over Workspace, SharePoint, Jira and other systems, an AI assistant and agents: prebuilt, built without code, or custom with ADK, all under central governance. Gemini Notebook Enterprise is for work over hand-picked sources, and its API creates, shares and loads notebooks and generates audio overviews.
Gemini in Workspace builds AI into the apps people already use: it summarizes threads and drafts in Gmail, creates drafts in Docs, analyzes data in Sheets, takes notes with decisions and next steps in Meet and helps create videos with Vids. Company data is not used to train the models, and Gemini only retrieves content the user already has access to.

Describe how Google Cloud’s gen AI offerings improve the customer experience

Agent Search (formerly Vertex AI Search) builds Google-quality search and recommendations over websites and your own data: it understands intent, gives AI answers with citations and works as ready-made RAG, including for media. Grounding with Google Search connects Gemini to the real-time web for information after the training cutoff; Agent Search is used when the answer lives in company data.
Customer Engagement Suite (now Gemini Enterprise for Customer Experience) covers the whole customer service journey: Conversational Agents offer voice and chat self-service with generative AI and deterministic flows; Agent Assist helps human reps live with replies, coaching and summaries; Conversational Insights analyzes topics, sentiment and quality; and Contact Center as a Service routes interactions across channels.

Describe how Google Cloud empowers developers to build with AI

Gemini Enterprise Agent Platform, the evolution of Vertex AI, brings together Model Garden, a library to discover, test, tune and deploy Google, partner and open models; Agent Search (formerly Vertex AI Search), semantic search with cited generative answers over your own data without deep ML skills; and AutoML, which builds code-free models from tabular or image data.
Google Cloud offers prebuilt RAG with Agent Search, an out-of-the-box semantic and keyword search engine, and RAG Engine, a managed service that connects private data to LLMs. For maximum control there are lower-level APIs: Layout Parser, embeddings, Vector Search, the ranking API, the grounded generation API and the check grounding API, which measures whether an answer is supported.
Agent Platform covers the full life cycle of a custom agent: build it with open-source ADK or other frameworks, with low-code Agent Studio or starting from Agent Garden; scale it on Agent Runtime with Sessions and Memory Bank; govern it with Agent Registry, Agent Identity and Agent Gateway; and optimize it with evaluation, simulation and observability.

Define the purpose and types of tooling for gen AI agents

Tools let an agent obtain up-to-date information and act on external systems. With functions, the model states what to call and with which parameters and the client application executes the call; extensions are executed by the platform itself; data stores give answers based on websites and documents with citations. Clear names and descriptions help the model choose well.
An agent can use pre-built APIs as tools: Speech-to-Text and Text-to-Speech for voice, Translation and Document Translation (keeps formatting), Document AI to extract fields from documents, Vision and Video Intelligence for images and video, and Natural Language for sentiment and entities. Cloud Run functions (formerly Cloud Functions) runs custom logic serverlessly and API Library helps discover APIs.
Google AI Studio is for quickly trying Gemini models and experimenting with prompts using a Google Account and an API key. Agent Studio, on Agent Platform, is the enterprise option: it compares prompts and models, designs low-code agents and offers IAM, SLAs, 24/7 support, compliance such as HIPAA, data residency, CMEK and third-party models in Model Garden.

3. Techniques to improve gen AI model output

20–20%

Describe how to proactively overcome foundation model limitations

Foundation models depend on the quality of their data and prompts, only know what was in their training data (the knowledge cutoff) and can hallucinate, producing plausible but false content. They can also amplify bias, serve underrepresented language varieties less well and fail with overconfidence on rare edge cases.
Grounding connects model output to verifiable sources (Google Search, Agent Search, RAG Engine), and RAG brings in current or private data without retraining. Google recommends starting with prompt design and moving to fine-tuning only with labeled data and tasks prompting cannot solve. Decisions that affect people need meaningful human oversight (human in the loop).
Production apps should pin a stable model version; the latest alias moves automatically to each new release. Migrations are validated with the Gen AI evaluation service, drift is watched with Model Monitoring, and Feature Store keeps feature data consistent and versioned. Success is measured with model, system and business KPIs, and incidents end with validated versions and patches.

Describe prompt engineering techniques and how they drive better results

Prompt engineering is the iterative process of designing prompts and assessing responses to steer a model toward the desired result without changing the model. Every prompt needs a task; system instructions, examples and context are optional. It brings accuracy, control and predictability, and still matters for complex tasks even though current models handle simple ones well.
Zero-shot gives only the instruction; one-shot and few-shot add one or more examples to fix format, style or tone, although too many examples cause overfitting. Role prompting sets a persona, ideally in system instructions. Prompt chaining splits a task into sequential steps where each output feeds the next, improving control and accuracy.
Chain-of-thought asks the model to reason through intermediate steps, improving accuracy on complex problems and showing how it reached the result; the reasoning must come before the answer and can be separated with an output format. ReAct alternates reasoning, acting with external tools and observing the result, ideal for agents that need live data.

Identify grounding techniques and their use cases

Grounding connects model output to verifiable sources: it reduces hallucinations, anchors responses to your data and provides links to sources. It can use first-party enterprise data (documents, databases, applications), specialized third-party data from external providers, and world data from the public web, for example through Google Search.
RAG first retrieves relevant facts from your sources and adds them as context to the query before the model generates. This reduces hallucinations and lets the model answer with private or up-to-date data without retraining. Answer quality depends on what is retrieved: inconsistent or outdated data gives equally poor answers, and reranking results improves relevance.
Agent Search (formerly Vertex AI Search) is the out-of-the-box RAG option: Google-quality search over enterprise data. The RAG APIs (layout parser, embeddings, Vector Search, ranking and check grounding) let teams build their own system with full control, and RAG Engine orchestrates the flow as a managed service. Grounding with Google Search adds current web knowledge with no index to manage.
Temperature controls randomness: low for predictable answers, high for more varied or creative ones. Top-p limits candidate tokens by cumulative probability. Maximum output tokens and stop sequences control length, and safety filters block harmful output using per-category thresholds without changing how the model behaves.

4. Business strategies for a successful gen AI solution

15–15%

Describe the Google Cloud-recommended steps to successfully implement a transformational gen AI solution

Gen AI creates new content: text (marketing copy, summaries, translation), images (Imagen), video (Veo), music (Lyria), code, conversational search, and personalized content or recommendations based on each user's profile. Traditional AI predicts or classifies from historical data (spam, forecasting, segmentation). The two can be combined, for example a churn model explored through a chatbot.
Google Cloud recommends starting from measurable business goals and documenting security, privacy and compliance constraints. Then assess whether the task needs accuracy or creativity, whether data is accessible and trustworthy, the risk tolerance and the speed required. Without enough training data, gen AI with prompt engineering is advised.
Choosing a solution starts with prioritizing high-value, high-feasibility use cases. Gemini Enterprise (formerly Agentspace) offers enterprise search and no-code agents; Gemini Enterprise Agent Platform (formerly Vertex AI) is for custom solutions, with Model Garden to compare models. When answers must follow changing data, grounding with RAG beats one-time fine-tuning.
Google Cloud suggests starting with a specific domain and persona through low-risk internal pilots, forming a dedicated team and setting goals and KPIs before launch. Scaling needs strong data foundations, change management, executive education and an AI Center of Excellence, plus a cost-benefit analysis. Guided, task-focused tools are adopted more readily than open-ended chat.
Google Cloud measures gen AI with KPIs for model quality, system quality, business operations (containment, handle time, CSAT), adoption (adoption rate, frequency, feedback) and business value. Set baselines before deployment and use A/B tests. Low adoption suggests an awareness problem; falling adoption suggests a performance problem.

Define secure AI and its importance in protecting AI systems from malicious attacks and misuse

AI security should be built in from design and carried through development, deployment and operations. Each stage has its own risks: data poisoning is mitigated early with sanitization and integrity controls, prompt injection by filtering inputs and outputs, and model reverse engineering or denial of service by rate limiting at the API.
SAIF is Google's conceptual framework for securing AI systems: it addresses model risk management, security and privacy so that AI is secure by default. It has six core elements, such as extending detection and response to AI and automating defenses, and SAIF.google offers a map of risks and controls plus a risk self-assessment.
Google's infrastructure is secure by design, with layers from custom hardware up to operations. IAM controls who can do what on each resource; Security Command Center detects vulnerabilities, misconfigurations and threats, including in AI workloads; Cloud Monitoring tracks performance and Cloud Audit Logs record who did what, where and when.

Describe the importance of responsible AI in business

Responsible AI matters to business: lack of trust slows adoption and many companies abandon projects over ethical concerns. Google recommends reviewing each use case in context from the design stage, with diverse reviewers, and being transparent about how decisions are made. Model cards document a model's performance and limitations.
Removing names is not enough: quasi-identifiers (ZIP code, age) combined can re-identify people. Pseudonymization replaces identifiers with tokens that keep joins possible and, with the key, allow re-identification; redaction and masking are not reversible. Sensitive Data Protection discovers and de-identifies sensitive data, including in prompts.
AI is only as good as its data: poor-quality or unrepresentative data produces biased outputs. Bias types include historical bias (data reflecting past inequities), sampling bias and automation bias. Detect them by evaluating performance per group; the most straightforward mitigation is to augment the data with representative examples.
Explainability builds trust, makes errors easier to debug and helps uncover bias. Local attributions explain a single prediction and global importance explains overall model behavior; neither proves causation. Accountability requires clear owners, audit trails and appropriate human oversight of AI decisions.

Practice this certification