GES-C02 study guide: SnowPro Specialty Gen AI exam

The SnowPro Specialty: Gen AI exam (GES-C02) tests whether you can build and govern generative AI workloads inside Snowflake using Cortex AI functions, Cortex Search, Cortex Analyst, Cortex Agents and document processing. It is a specialty certification for practitioners with at least a year of gen AI work on Snowflake. You answer 55 questions in 85 minutes, mostly multiple choice and multiple select with some interactive items, and need a scaled score of 750 out of 1000.

GES-C02 at a glance

Details as of September 2026 — confirm on the official exam page before booking.

Who this exam is for

Snowflake describes the candidate as someone with one or more years of gen AI experience with Snowflake in an enterprise environment. Python proficiency is helpful, and data engineering and SQL knowledge are assumed. There are no formal prerequisites.

That last point shapes how you should study. The exam does not teach SQL or pipelines; it assumes them. Most questions put a Cortex feature inside a realistic data task, so fluency with roles, warehouses, stages, streams and tasks is part of the ground you stand on.

What the exam covers

The study guide behind the current exam is dated April 2026. Snowflake ships Cortex features quickly, and several items in the guide were in preview when it was written, so check the current documentation for anything you plan to rely on. Two names have already changed since the guide: Snowflake Intelligence is now Snowflake CoWork, and Cortex Code is now Snowflake CoCo. The guide uses the older names and the current documentation uses the new ones, so learn both.

Snowflake for Gen AI Overview (18%)

This domain asks you to place each Cortex capability correctly: AI functions for model calls in SQL, Cortex Search for retrieval over unstructured data, Cortex Analyst for text-to-SQL over semantic views, Cortex Agents for orchestration, and Snowflake Intelligence as the end-user surface. It also covers Cortex Code, the interfaces (AI Studio, SQL, REST API), cross-region inference, MCP, Cortex Knowledge Extensions, and bringing your own models through Model Registry and Snowpark Container Services. Learn what problem each piece solves and how they connect.

Snowflake Gen AI Functions (38%)

The heaviest domain. It covers the general AI_COMPLETE function with structured outputs, the task-specific functions (classify, extract, parse, sentiment, summarise, translate, embed, filter, aggregate, similarity, transcribe, redact), vector distance functions and helpers such as token counting and text splitting. It then moves to use: RAG with Cortex Search and chunking choices, text-to-SQL with Cortex Analyst and verified queries, multi-turn chat in Streamlit in Snowflake, Cortex inside data pipelines, and running third-party models in Snowpark Container Services. Performance questions ask you to trade model size, latency, accuracy and provisioned throughput.

Snowflake Gen AI Governance (29%)

A large share for governance, and much of it is precise. You restrict which models can be used (account-level allowlist, RBAC, application roles), handle cross-region inference and sensitive data, and authenticate REST calls. You grant the right privileges and database roles for Analyst, Search, Agents and Snowflake Intelligence. You manage costs: what Cortex Search bills for, how to limit agent token use, which usage history views track which service, and how object tagging attributes spend. It ends with AI observability (evaluation, tracing, event tables) and the TruLens SDK.

Snowflake Document Processing (15%)

This domain covers AIPARSEDOCUMENT in its OCR and layout modes with page options, AI_EXTRACT with response formats and prompting, uploading documents and their format and size requirements, and automating processing with streams and tasks. Troubleshooting questions cover query errors, presigned URLs, privileges, cost and fine-tuning extraction models.

A 4-week study plan

This plan assumes six to eight hours a week in a Snowflake account where you can use Cortex. The functions domain is 38%, so it gets most of two weeks.

Week 1: Overview (18%) and the start of functions. Draw the Cortex map: which service handles retrieval, which handles text-to-SQL, which orchestrates, and which the business user sees. Then call AI_COMPLETE and three task-specific functions on a real table. Finish with a mixed paper on the overview domain.

Week 2: Functions in practice (38%). Build a small RAG flow: split text, create a Cortex Search service, retrieve, and answer with AI_COMPLETE. Build a semantic view and ask Cortex Analyst questions against it, then add a verified query. Put a Cortex call inside a pipeline. Use topic papers on the function family you use least.

Week 3: Governance (29%). Grant and revoke the Cortex database roles and watch what breaks. Set the model allowlist. Read the usage history views and match each to a service. Learn the cost components of Cortex Search. Set up tracing for one application and look at what observability records.

Week 4: Document processing (15%) and rehearsal. Parse a scanned PDF in both modes and compare the output, then run AI_EXTRACT with a response format. Automate it with a stream and task. Sit a full-length mock in 85 minutes, drill weak topics, and sit a second. The [AI certification study plan template](/blog/ai-certification-study-plan-template) adapts this to a longer schedule.

Where candidates lose marks

• Mixing up Cortex Search and Cortex Analyst. Search retrieves passages from unstructured text; Analyst turns questions into SQL over structured data via a semantic view. Scenarios describe the data, and the data type decides the service.
• Privileges and roles. Governance is 29% and much of it is exact. Know which database role grants which capability, and which extra privileges each service needs. Guessing from names is unreliable.
• Cost questions. Cortex Search bills in more than one way, and each usage history view tracks a different service. Candidates who have only used Cortex rarely know which view to query.
• Choosing AI_COMPLETE for everything. When a task-specific function exists (classify, extract, sentiment, redact), it is usually the better answer for cost and simplicity. The general function is for open-ended generation.
• Multiple-select questions. Read how many answers are required and choose exactly that many. Each option must hold up against the scenario's constraints, not just be true in general.
• Trusting preview-era notes. Features named in the guide as preview may have changed. Check current docs before memorising parameter names or limits.

Practising for GES-C02 on CertifyAIx

CertifyAIx has 806 original GES-C02 practice questions written against Snowflake's published study guide and weighted by its four domains, each with explanations. Most are single- and multiple-answer; a minority use matching and ordering to rehearse sequences such as pipeline setup. Snowflake's guide mentions interactive questions without naming their kinds, so these formats are our choice rather than a claim about how the real exam's interactive items look. See [how CertifyAIx practice questions are written](/blog/how-certifyaix-practice-questions-are-written).

• 3 full-length mock exams, timed at the real exam's 85 minutes
• 34 mixed domain papers and 39 topic papers of up to 25 questions each (60 minutes per 25)
• Progress tracked by domain and topic

Start with the [GES-C02 practice questions](/certifications/snowpro-specialty-gen-ai-ges-c02).

Frequently asked questions

How much experience do I need for GES-C02?

Snowflake recommends one or more years of gen AI experience with Snowflake in an enterprise setting, with SQL and data engineering knowledge assumed. There is no formal prerequisite.

How long should I prepare?

If you already build with Cortex, three to five weeks is a reasonable range. If you know Snowflake but not Cortex, allow longer, mainly for the governance and cost details that only come from hands-on use.

Are CertifyAIx questions real SnowPro questions?

No. Every CertifyAIx question is original and written from Snowflake's public study guide. We never use content from a live exam.

What score do I need?

Snowflake lists a passing score of 750 on a scaled range of 0 to 1000. Scaled scoring means it is not a simple percentage of questions correct.

Does the exam include unscored questions?

Snowflake says exams may include unscored items for statistical purposes. They are not identified on the exam and do not count towards your score.

CertifyAIx is independent and not affiliated with Snowflake. SnowPro Specialty: Gen AI is a trademark of its owner. All CertifyAIx practice questions are original.