CT-GenAI exam questions: format, K1 to K3 levels, traps and time management

You know the syllabus, you have practiced prompting, and you still lose points in practice tests. Most of the time, the problem is not knowledge: it is a negation read too fast, two "true" answers where you had to pick the best one, or an application question that ate five minutes.

This article takes apart the ISTQB® CT-GenAI exam questions (Testing with Generative AI): their format, what the K1, K2 and K3 levels test, with an original example of each, the wording traps, a time management strategy based on the numbers, and how to read an answer explanation properly.

The format of CT-GenAI exam questions

The exam has 40 multiple-choice questions worth 46 points, in 60 minutes (75 minutes if the exam is not in your native language). You need 30 points, which is 65%. Under the ISTQB® rules, each question asks you to select the best answer, or answers, from a list. Some questions explicitly ask you to select two: the sample exam published by ISTQB® includes them, and the number expected is stated in the question.

Not all points are equal:

Level	What it tests	Questions	Points per question	Total points
K1	Remember	8	1	8
K2	Understand	26	1	26
K3	Apply	6	2	12
Total		40		46

Source: ISTQB®, Exam Structure Tables v1.19.

K3 questions are 15% of the questions but 26% of the points. Five of them cover chapter 2 (prompt engineering) and one covers chapter 3 (risks). Finally, the sample exam states that real exams may include other question types and a different level of difficulty: train on the substance, not on one particular format.

The K-levels explained, with an example of each

The three questions below are illustrative examples written by QAcademy, using fictional data. They do not come from the exam and do not reuse the ISTQB® sample exam.

K1 — Remember: recognize a term

A K1 question checks that you know a definition or a keyword. The keywords listed at the top of each syllabus chapter are examinable, even when no learning objective mentions them.

K1 example (1 point). A tester puts a single example test case, in the team's format, into the prompt, then asks the LLM to produce test cases for a new user story. What is this technique called?

a) Zero-shot prompting b) One-shot prompting c) Meta prompting d) Prompt chaining

Answer: b. A single example is one-shot prompting. Zero-shot gives no example; meta prompting asks the LLM to produce or improve a prompt; prompt chaining splits the task into a sequence of prompts. No trap here: either you know the term or you do not. That is why keyword flashcards pay off.

K2 — Understand: explain, compare, choose

A K2 question checks that you understand why: the difference between two concepts, or the advantage of an approach in a context. It is the bulk of the exam: 26 questions.

K2 example (1 point). The test team of a fictional health insurer uses an LLM-based assistant to propose test conditions. The benefit specifications are published in an internal wiki and change every sprint. The team wants the proposals to be based on the latest version of the specifications. Which approach BEST meets this need?

a) Fine-tune the LLM on the current specifications, so that the business rules are built into its parameters. b) Set up RAG: the relevant passages of the specifications are retrieved at query time and given to the LLM as context. c) Raise the LLM's temperature so that it explores more variants of the business rules. d) Use a larger foundation LLM whose training data covers more insurance rules.

Answer: b. RAG retrieves information at query time, so each wiki update is taken into account as soon as it is indexed. Fine-tuning (a) freezes knowledge at training time: you would have to retrain every sprint, and the syllabus presents it mainly as a way to adapt a model to a domain, a vocabulary or an output format. Temperature (c) affects the variability of outputs, not how current they are. A larger model (d) still knows nothing about the company's internal specifications. To go further: our article on RAG for software testing.

K3 — Apply: solve a scenario (2 points)

A K3 question puts you in front of a scenario and asks you to apply a skill: choose a prompting technique, improve a prompt, evaluate an LLM output.

K3 example (2 points). Requirement: "A claim can be filed online no later than 5 working days after the date of the incident. After that, the claim must be filed by phone." Working days are Monday to Friday; there are no public holidays in the period. For an incident on Friday, March 6, 2026, a tester asks the LLM for two boundary value test cases. The output:

Reasoning: the weekend accounts for 2 non-working days,
so 5 - 2 = 3 working days to add from Monday, March 9.
Last accepted day: Wednesday, March 11.
TC-01: claim filed on Wednesday, March 11 → accepted online
TC-02: claim filed on Thursday, March 12 → refused, redirected to phone

Which assessment of this output is MOST accurate?

a) A hallucination: the LLM invented the redirection to the phone. b) A reasoning error: the rule is restated correctly, but the calculation is wrong. The boundary is Friday, March 13, and TC-02 expects a refusal for a date that must be accepted. c) A bias: the model favors invalid cases over valid ones. d) The output is correct: the two test cases sit on either side of a boundary.

Answer: b. Count: Monday 9 (1), Tuesday 10 (2), Wednesday 11 (3), Thursday 12 (4), Friday 13 (5). The last accepted day is Friday, March 13, and the first refused day is Monday, March 16. The LLM applied faulty logic (subtracting the weekend from the number of working days): that is a reasoning error as defined in section 3.1.1. The phone redirection is in the requirement, so (a) is wrong; nothing points to a systematic preference (c); and (d) is the trap: the cases do sit around a boundary, just not the right one. As a remedy: ask for the calculation step by step and check it (logical validation), or provide the calendar in the prompt. For the other ways LLM output goes wrong, see our article on LLM hallucinations in testing.

Wording traps

• "BEST", "MOST", "MOST LIKELY". Several options can be partly true. Do not look for a correct answer: rank them and pick the one that answers the stem most directly.
• Negations: "NOT", "INCORRECT", "EXCEPT". The question asks for the false option. Write T or F next to each option before choosing: this is where fatigue does the most damage.
• Absolute words: "always", "never", "all", "eliminates", "without any human intervention". With non-deterministic systems, an absolute statement is rarely right, and the syllabus stresses human verification throughout. It is not an automatic rule, but it is a signal to reread the option.
• True, but off-topic. A distractor can be an accurate statement from the syllabus that does not answer the question asked. Reread the stem before you commit.
• "Select TWO options". Judge each option on its own, and check that you select exactly the number asked.
• Neighboring concepts. Hallucination or reasoning error, RAG or fine-tuning, system prompt or user prompt: the K3 example above shows how an almost-right option catches a rushed candidate.

Time management: minutes per question

On average, you have 60 ÷ 40 = 1.5 minutes per question. The ISTQB® rules give an indicative time per level, which allows a finer calculation:

Level	Questions	Indicative time per question	Total time
K1	8	≈ 1 min	8 min
K2	26	1 to 2 min	26 to 52 min
K3	6	≈ 3 min	18 min

At 1 minute per K2 question, the total is 8 + 26 + 18 = 52 minutes, leaving an 8-minute margin out of 60 (23 minutes out of 75). But if you spend 2 minutes on every K2 question, you reach 78 minutes: impossible. The conclusion: K2 questions need to take around one minute, to protect the time that K3 questions require.

A practical strategy:

1. One complete first pass, at about 1.3 minutes per question (52 ÷ 40): checkpoints at question 10 around 13 minutes, question 20 around 26 minutes, question 30 around 39 minutes.
2. Flag and move on from any question that still blocks you after 3 to 4 minutes, even a K3. Losing two points beats rushing three questions at the end.
3. Keep the margin for flagged questions and for rereading negations.
4. Leave no question unanswered: get back to every flagged question before time runs out.

How to read an answer explanation

The review is where you learn the most, as long as you do not stop at "I got it wrong". For every missed question:

1. Reread the stem without the options. What exactly was it asking, and at which K-level?
2. Understand why the correct answer is right, by finding the relevant syllabus section.
3. Understand why each distractor is wrong: false statement, true but off-topic, too absolute, or a neighboring concept.
4. Classify your mistake: missing knowledge, reading too fast, vocabulary confusion, lack of time.
5. Log it in an error journal, with the learning objective. Sample entry: "LO 3.1.2 — reasoning error taken for a hallucination — cause: vocabulary — revisit: 3.1.1."

After two mock exams, this journal tells you whether your problem is knowledge or reading. Those call for different revision.

What to remember for the CT-GenAI exam

• Keywords = K1. The terms listed at the top of each chapter are examinable: one-shot, context window, vector database, shadow AI…
• 32 of the 40 questions are K2 or K3: the exam rewards understanding, not memorization.
• K3 points are concentrated on LO 2.2.1 to 2.2.5 (chapter 2) and LO 3.1.2 (identify hallucinations, reasoning errors and biases).
• A question can draw on several syllabus sections, like the K3 example, which combines boundary values and reasoning errors.
• 30 points out of 46: you need at least 30 of the 34 K1 and K2 questions if no K3 question is right, and only 18 if all six are. K3 questions change the picture.

The full scoring and how the exam runs are covered in our CT-GenAI exam guide. To organize your preparation weeks, see our CT-GenAI study plan.

FAQ

Are there questions with more than one answer?

Yes. Some questions ask you to select two answers, and the question says how many ("Select TWO options"). In the ISTQB® sample exam, the vast majority of questions expect a single answer.

Do the ISTQB® sample exam questions appear in the real exam?

No. The document states that its questions cannot be used as-is in a real exam. It is there to get familiar with the style, not to learn answers.

How much time should I allow for a K3 question?

About 3 minutes, based on the ISTQB® guideline. They are worth 2 points each: do not rush them, but do not stay stuck beyond 4 minutes on the first pass.

---

The best way to find out which question types cost you points is to face them. The free diagnostic gives you 10 exam-style questions in 5 minutes, with a result per chapter. It is indicative, but it is an honest starting point.

Sources

• ISTQB®, CT-GenAI certification page: exam format and downloadable documents.
• ISTQB®, Exam Structure Tables v1.19, 2026-08-19: questions, points and K-levels per chapter, indicative time.
• ISTQB®, Exam Structures and Rules v1.2, 2025-05-02: multiple-choice format, points for K3 questions, time per K-level, extra time.
• ISTQB®, CT-GenAI Sample Exam set A, v1.1, 2026-04-27: two-answer questions, limits of the sample exam.
• ISTQB®, Certified Tester Specialist Level Syllabus – Testing with Generative AI, v1.1, 2026-04-27, sections 0.5, 0.6, 3.1 and 4.1–4.2.