How the readiness score is calculated

The sample follows the blueprint

Each of these exams has a published blueprint: its domains and how much of the exam each one is worth. A certification attempt here draws 30 questions and shares them out across those domains in proportion to their published weight, so the mix you answer has the same shape as the exam you are preparing for.

The share-out uses the largest-remainder method. Each domain first gets the whole-number part of its exact share of the 30 questions; the few slots left over go to the domains with the largest fractions, ties going to the heavier domain. That keeps the total at exactly 30 without favouring any domain by rounding. Questions inside each domain are drawn at random, so a retake is a different set.

For NVIDIA’s NCP-AAI, whose ten published weights add up to 98%, the full test draws this:

One thing can move those counts slightly. On Easy or Medium the pool is narrowed to easier questions first, and if a domain runs short at that level, the gap is filled with other random questions from the same exam so you still get as close to 30 as the bank allows.

A domain needs three questions

One or two questions say almost nothing about a domain. Miss the only question you were given and you would be told you are at 0% on a topic you may know well. So a domain gets its own bar on the results page, and can be flagged as a weak area, only when your attempt included at least 3 of its questions.

In the NCP-AAI draw above, that is 6 domains worth 76% of the blueprint. The other 4 still count: every answer feeds the overall score. They are simply not diagnosed on their own, because the sample is too small to say anything fair. A retake draws a fresh set, which is the honest way to dig into them.

A scored domain is flagged weak when you earned less than 60% of its points.

Harder questions count for more

Each question is tagged easy, medium or hard and is worth 1, 2 or 3 points to match. Inside a domain, your result is the share of available points you earned, so a hard question you get right counts 3 times as much as an easy one. Every question has 4 options and exactly one correct answer.

The difficulty you choose sets the pool. Easy draws easy questions only, Medium draws easy and medium, and Hard, the full test, draws from all three.

The overall score is weighted by domain

The overall score is not the share of questions you got right. Each domain’s result is multiplied by its blueprint weight, the results are added up, and the total is divided by the sum of the weights of the domains in your attempt. That gives a number from 0 to 100 in which a 15% domain moves the score three times as much as a 5% one, as it would on the real exam.

That has a consequence worth knowing. On NCP-AAI, an attempt that is perfect except for every question in Agent Architecture and Design (15%) scores 85: Exam Ready. The same attempt missing all of Human-AI Interaction and Oversight (5%) instead scores 95. An overall score can look comfortable while hiding a domain you know nothing about, which is why the results page shows a bar for every scored domain and flags a weak one by name.

The four bands

• 85–100: Exam Ready, shown with a pass likelihood of 88%
• 70–84: Almost Ready, shown with a pass likelihood of 65%
• 50–69: At Risk, shown with a pass likelihood of 42%
• 0–49: Not Ready, shown with a pass likelihood of 20%

Be clear about what the pass-likelihood figure is. It is one fixed number per band, set by hand. It does not come from your individual answers beyond the band you land in, and it is not calibrated against real exam results. Read it as a plain-language restatement of the band, not as a prediction.

The real exams do not score this way either. Microsoft, for one, sets its pass mark at a scaled 700 out of 1,000, which cannot be converted into a share of questions answered correctly. Aim to land clear of a band boundary rather than on it.

Keeping the question bank honest

Multiple-choice questions leak their answers in two predictable ways: the correct option sits in the same position too often, or it is visibly the longest, most carefully qualified option. Either one lets someone score well without reading the question, and a readiness score that can be gamed is worse than none.

This is not a hypothetical. When the NCP-AAI bank was measured on 5 September 2026, always answering B scored 70%, and always picking the longest option scored 98%. The bank was fixed, and scores from before that date overstated readiness. Two bank-wide limits have been enforced since:

• Answer position. No one of the four positions may hold more than 35% of the correct answers in a bank. An even spread would be 25%.
• Longest option. The correct answer may be the longest option in no more than 45% of questions, so “pick the longest” stops being a strategy.

Question by question, a correct answer more than 1.3 times the length of the longest wrong option, and more than 25 characters longer, is sent for review. Options such as “all of the above” and “none of the above” are rejected outright, because they test the format rather than the subject.

The checks run in three places: in the tests every newer certification bank must pass before it ships, on newly generated questions before they can reach the bank, and against the whole live bank once a week, because the same biases come back whenever questions are written in bulk.

What the score cannot tell you

• It is a sample. 30 questions are fewer than a real exam sitting, so one score carries noise. Two or three attempts, each a fresh draw, tell you more than one.
• The questions are not exam questions. They are written to the published blueprint. Vendors do not release their live exam items, so no practice test can show you the real ones.
• It measures knowledge, not exam conditions. Timing, nerves and question formats beyond multiple choice are not part of the score.

What it does do is tell you, domain by domain, where your knowledge stands against the weights the vendor says matter, and it is built so that a high score has to be earned.

Which tests work this way

Every certification test on the site uses the method above, each with its own vendor’s blueprint:

• NVIDIA NCP-AAI readiness test
• NVIDIA NCP-AAI practice test
• Microsoft AI-103 readiness test
• GitHub GH-600 readiness test
• CompTIA SecAI+ readiness test
• Microsoft AI-901 readiness test

The older subject quizzes, such as Python and SQL, have no blueprint to weight against. Their score is simply the share of available points you earned, using the same points and bands, and any topic under 60% is flagged without the three-question minimum. For the exams themselves, the AI certifications comparison lists every figure each vendor publishes.