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
- Standard exam: 50 to 60 multiple-choice and multiple-select questions in 90 minutes. Fee $99. Online or test-centre proctored. Valid for 3 years. Google does not publish a passing score.
- Renewal exam (active certification, within 180 days of expiry): 20 questions, 45 minutes, $60, same content.
- About 1.5 minutes per question. Flag hard ones and return. Answer everything; there is no stated penalty for guessing.
- No hands-on tasks. It’s a business-level exam about products and outcomes.
- Section weights (exam guide launched August 2026): digital transformation about 18%, data transformation about 18%, AI about 18%, modernizing infrastructure and applications about 18%, trust and security about 18%, scaling with Google Cloud operations about 10%.
- 2026 naming: Vertex AI is now Gemini Enterprise Agent Platform (Agent Platform), Vertex AI Studio is Agent Studio, and Dataproc appears as Managed Service for Apache Spark. Old names may still appear in older material.
How to read a question
Read the last sentence first, then find the business goal word:
| The question says | Think |
|---|---|
| Quickly, minimal change | Rehost, ready-made products, pre-trained APIs |
| No infrastructure to manage, scales to zero | Serverless: Cloud Run, Cloud Run functions, BigQuery |
| Avoid vendor lock-in, open | Open source, open standards, multicloud (GKE Enterprise, BigQuery Omni) |
| Differentiate, proprietary data | Custom or tuned model on your own data |
| Analysts who know SQL | BigQuery, BigQuery ML |
| Real time | Pub/Sub plus Dataflow, streaming into BigQuery |
| Never trust, always verify | Zero trust, BeyondCorp, Identity-Aware Proxy |
| See and fix misconfigurations | Security Command Center |
Elimination tricks
- Pick the most managed, most ready-made option that meets the need. Answers requiring you to build or run things yourself are usually wrong for this exam.
- Watch for absolutes (“always”, “only”, “never”) in a distractor.
- Business answers beat low-level technical ones. “Faster time to market” beats “uses TLS 1.3” when the question is about value.
- In select TWO questions each pick must be true and must answer the question separately.
Product matching
| Need | Product |
|---|---|
| Warehouse analytics on huge data, SQL | BigQuery |
| Files, media, backups, data lake | Cloud Storage |
| Standard relational (MySQL, PostgreSQL, SQL Server) | Cloud SQL |
| High-performance PostgreSQL-compatible | AlloyDB |
| Global relational with horizontal scale | Spanner |
| Huge low-latency NoSQL, IoT, time series | Bigtable |
| Mobile and web documents, real-time sync | Firestore |
| Governed self-service BI | Looker |
| Event ingestion and messaging | Pub/Sub |
| Stream or batch processing (Apache Beam) | Dataflow |
| Existing Spark or Hadoop jobs | Managed Service for Apache Spark |
| Models with SQL | BigQuery ML |
| Build, tune and run models and agents | Gemini Enterprise Agent Platform |
| Models with no code on your data | AutoML |
| Employee AI assistant and agents | Gemini Enterprise |
| VMs | Compute Engine |
| Managed Kubernetes | GKE |
| Serverless containers | Cloud Run |
| Event-triggered small code | Cloud Run functions |
| Manage and monetise APIs | Apigee |
| Manage Kubernetes across clouds | GKE Enterprise |
| Query data in AWS or Azure | BigQuery Omni |
| Metrics, alerts, dashboards | Cloud Monitoring |
Cloud Storage classes
| Class | Access | Minimum duration |
|---|---|---|
| Standard | Frequent | None |
| Nearline | About monthly | 30 days |
| Coldline | About quarterly | 90 days |
| Archive | Less than yearly | 365 days |
| Autoclass | Unknown or changing | Automatic movement |
Migration strategies (least to most change)
Retire (switch off) and retain (leave) -> rehost (lift and shift) -> replatform (move and improve) -> refactor (rewrite cloud-native) -> reimagine (new business model). Always start with discovery and assessment.
Security quick facts
| Topic | Fact |
|---|---|
| Shared responsibility | Google secures infrastructure; customer secures data, identities, configuration. Google adds “shared fate” |
| Encryption | At rest by default; in transit with TLS; in use with Confidential Computing |
| Authn / authz / audit | Who you are / what you can do / what you did (IAM, MFA, audit logs) |
| DDoS and web attacks | Cloud Armor |
| Private link to data centre | Cloud Interconnect (dedicated) or Cloud VPN (encrypted over internet) |
| Sensitive data discovery and masking | Sensitive Data Protection |
| Threat intelligence sources | Google visibility, Mandiant, VirusTotal |
| Posture and risk | Security Command Center |
| SIEM and SOAR with Gemini | Google Security Operations |
| Securing prompts and AI | Model Armor, AI Protection |
| Trust | Transparency reports, audits, sovereignty, residency, Access Transparency |
Operations facts
| Thing | Fact |
|---|---|
| CapEx to OpEx | Upfront hardware to pay-as-you-go; TCO is the full lifetime cost |
| Hierarchy | Organization -> folders -> projects -> resources; policy inherits downward |
| Budgets | Alert, don’t cap. Quotas limit use |
| Cheaper compute | Spot VMs (interruptible), committed use (1 or 3 years), sustained use (automatic) |
| Observability | Monitoring (metrics), Logging, Trace (latency path), Profiler (code), Error Reporting |
| Golden signals | Latency, traffic, errors, saturation |
| SLI / SLO / SLA | Measurement / internal target / customer contract with consequences |
| Error budget | 100% minus the SLO |
Pairs people mix up
| A | B | The difference |
|---|---|---|
| Hybrid | Multicloud | On premises plus cloud / more than one public cloud |
| Latency | Bandwidth | Delay of a request / capacity per second |
| Region | Zone | Geographic area / isolated site inside it |
| Data warehouse | Data lake | Modelled structured analytics / raw data of any type |
| Cloud SQL | Spanner | Standard relational / global scale relational |
| Bigtable | BigQuery | Operational NoSQL at huge scale / analytical SQL warehouse |
| Cloud Run | GKE | Serverless containers, no cluster / orchestration with control |
| Rehost | Refactor | Move as is / rewrite for cloud |
| Authentication | Authorization | Prove identity / grant permission |
| SLO | SLA | Internal target / customer contract |
| Looker | Looker Studio | Governed enterprise BI / lightweight dashboards |
| AutoML | BigQuery ML | No-code custom model / model built with SQL |
Defaults that are usually right
- Start with a business outcome, then pick technology.
- Prefer managed and serverless over self-managed.
- Use ready-made AI first; customise when it creates differentiation.
- Good data and governance come before good AI.
- Layer security (defense in depth) and apply least privilege and zero trust.
- Control cost with people, process and technology together.