About the questions
How the questions are written
Every question here is original. We write it from the vendor's own exam guide and documentation and link the page that backs the answer. If one is wrong, you can report it from the question itself.
462 questions across 9 tracks. Every one links its source. Page updated .
Written against each vendor's exam guide
Each certification track follows its vendor's published exam guide. Questions are filed under the guide's own domains. A practice set shorter than the whole bank draws from each domain by the weight the guide gives it, so the set has the shape of the exam.
- AWS Certified AI Practitioner (AIF-C01): 65 questions, written against the AIF-C01 exam guide.
- Microsoft Azure AI Fundamentals (AI-901): 50 questions, written against the AI-901 exam guide.
- Google Cloud Generative AI Leader (Gen AI Leader): 60 questions, written against the Gen AI Leader exam guide.
- AWS Machine Learning Engineer Associate (MLA-C02): 30 questions, written against the MLA-C02 exam guide.
- Databricks Generative AI Engineer Associate (GenAI Engineer): 45 questions, written against the GenAI Engineer exam guide.
- Claude Certified Architect: Foundations (CCAR-F): 59 questions, written against the CCAR-F exam guide.
- Claude Certified Developer: Foundations (CCDV-F): 105 questions, written against the CCDV-F exam guide.
- Forward-Deployed AI Delivery Skills: 24 questions. No vendor exam exists for this skills lab, so its questions are written from the documentation it covers.
- OpenAI API Solutions Skills: 24 questions. No vendor exam exists for this skills lab, so its questions are written from the documentation it covers.
AI models draft and check the questions. Each check is a separate run with one job, and the fact checks open the cited page and test the answer against it.
A source on every question
All 462 questions link the documentation page that supports the answer. The explanation shows that link after you answer, so you can check the reasoning yourself instead of taking our word for it.
How the bank was checked
The first bank of original questions, each written from public documentation and linked to its source.
The bank grew to 450 questions. For each track, one pass wrote questions only from the vendor's published objectives and documentation, a second opened every cited page and checked the answer against it, and a third compared the wording with the vendor's published item style. Three questions were rejected outright.
An audit of every question against a published set of multiple-choice writing rules (Haladyna, Downing and Rodriguez, 2002). Each wrong option had to be a mistake a real candidate makes. 537 wrong options did not meet that bar and were rewritten. A further pass argued against each rewrite and fixed the 44 problems it found. Three questions that had drifted close to Anthropic's published sample questions were rewritten.
A wording pass matched questions to the way each vendor's own sample questions read: AWS's MOST and LEAST closings, Microsoft's current product names, Google's direct questions. 109 questions in the current bank carry that rewrite.
A crawl of every cited source found links that had moved or broken. Each one now points at a page that still states the fact the question turns on.
The longest answer is not a giveaway
A common flaw in multiple-choice questions is a correct option that runs longer than the rest, because it carries the reason it is right. A September review found it here: on the CCDV-F track the correct option was the longest of the four in 94% of questions.
The reasons moved into the explanations and the options were rebalanced. Today the correct option is the longest on 9% of questions (42 of 462) and the shortest on 10% (46). Chance is 25% for each. A test runs on every change and fails if any track has the correct option longest, or shortest, on more than 40% of its questions.
What we do not claim
- We are not affiliated with or endorsed by Anthropic, OpenAI, AWS, Microsoft, Google, Databricks or Pearson VUE. Their names identify the exams.
- These are not real exam questions. We do not use recalled or leaked exam items, and we do not copy the vendors' sample questions.
- A practice score is not a pass prediction. The 80% goal is a study target. How scoring works
- The banks do not cover every exam objective. Every question here has one correct answer, while several of these exams also use multiple-response and other item types.
Found a problem? Report it
Under every question, on its own page and in the answer review after a set, there is a "Report a problem with this question" link. Pick what is wrong (the answer, the wording, out of date, a broken link, or something else), add a note if you like, and send it.
Reports are grouped by question for the person who maintains the bank. A report stores the question, the reason, your note and the time. If you send it from a results screen or while signed in, it also stores that practice set or your account. It never stores your email or IP address. Privacy