Question banks: how to build one that makes good tests

Most question banks start as a folder of old tests. That works until you need a second version of an exam, a retake, or a fair test for a group that already saw last year's paper. Then you discover that you cannot find questions by topic, do not know which ones were already used, and cannot tell whether two versions are equally hard. This article explains what turns a pile of questions into a bank that produces good tests, and how to build one from what you already have.

A question bank is not a folder of old tests

What to record with each question

A question is only as useful as the information stored with it. At a minimum, record:

• Topic: the chapter, unit or learning objective it tests. Use the same list of topics as your syllabus, organised as a tree (subject → chapter → topic) so you can draw from a whole chapter or a single topic.
• Level: an intended difficulty or cognitive level, such as recall, understanding and application, or easy, medium and hard. Keep the number of levels small; three to five is usually enough to build balanced tests.
• Type: single choice, multiple answers, true/false, fill-in, essay, or a group of questions on a shared passage or audio clip.
• Answer key and, ideally, a short rationale: why the answer is right and why the strongest distractor is wrong. The rationale makes reviews faster and helps whoever edits the question in two years.
• Status: draft, in use or retired. Draft questions are written but not reviewed; retired ones stay in the bank for the record but are never drawn.
• Usage: how many times the question appeared on a test. Frequently used questions are more likely to have been memorised or shared.

After each exam, add what you learned: the difficulty and discrimination from the item analysis, and any complaint or correction. Over time this turns the label "medium" from a guess into a measured fact.

The blueprint: how a bank turns into a test

A blueprint (or test specification) says how many questions of each topic and level a test contains. For a 20-question test it might look like this:

Each cell is a small draw: "take 3 medium questions on percentages". Every version built from the same blueprint covers the same content at the same levels, which is what makes versions comparable and the test valid for its purpose. Write the blueprint first, from the syllabus and the decision the test supports, and then check that the bank has enough questions in every cell.

How many questions does a bank need?

Think per cell, not per bank. If you want versions that share no questions, each cell needs at least the number of questions the cell asks for multiplied by the number of versions. In the blueprint above, three versions with no shared "medium percentages" questions need 3 × 3 = 9 such questions in the bank.

In practice:

• For a single class test, a bank two to three times the length of the test gives you a fresh version for absentees and a retake.
• For tests given to many groups over time, such as course finals, hiring tests or certification exams, aim for five times the test length or more, and watch usage, because questions leak through people who have already sat the test.
• Thin cells are the usual bottleneck. A bank with 400 questions can still fail to produce two versions if it has only three hard questions on one topic. Count by cell before an exam, not after.

Write questions that last

A question that will be reused for years deserves more care than one written for a single test. The most widely cited guidelines for multiple-choice questions come down to a few rules:

• Put one clear problem in the stem, so a candidate who knows the subject could answer before reading the options.
• Write one answer that is clearly correct and distractors that are plausible to someone who has not mastered the topic, ideally based on real mistakes.
• Keep options similar in length and grammar; avoid "all of the above" and avoid negatives such as "which is NOT" unless the negative is the point.
• Make each question independent: one question must not give away the answer to another, which matters when questions are drawn into different combinations.
• Avoid facts that expire, such as current prices or "this year", or record a review date with the question.
• Avoid references to position, such as "the option above", because options and questions may be shuffled.

Keep the bank healthy

• Review after every exam. Run an item analysis and act on it: fix wrong keys at once, rewrite questions with low discrimination, and replace distractors nobody chooses. The statistics are explained in item analysis: difficulty and discrimination.
• Retire, do not delete. Questions that leaked, became outdated or performed badly should be marked retired so they are never drawn, while the record of past tests stays intact.
• Watch exposure. Prefer less-used questions when drawing, and refresh the cells you draw from most often.
• Keep drafts out of exams. New questions, including those drafted with AI, should be reviewed by someone who knows the subject before they become usable.
• Limit access. A bank is only as secure as the people who can read it. Keep answer keys out of shared drives and printed copies.

Building a bank from what you already have

1. Write the topic list and levels, matching the blueprint of the tests you run most.
2. Import your existing tests, then tag each question with its topic and level. Expect to fix or drop some questions on the way; old tests often contain duplicates and outdated content.
3. Count questions per cell and fill the gaps: write new questions, or draft them with AI and review every one before use.
4. Build your next test from the blueprint, run it, analyse it, and feed the results back into the bank.

Frequently asked questions

What is the difference between a question bank and a test bank?

The terms overlap. "Test bank" often means the ready-made questions a publisher supplies with a textbook; "question bank" and "item bank" usually mean a collection you build and maintain yourself, organised for drawing tests.

How many questions should a question bank have?

Enough for every cell of your blueprint: at least the cell's question count times the number of versions you want without overlap. As a rule of thumb, two to three times the test length for a single class, five times or more for tests given repeatedly.

Should candidates see the question bank?

Not for a graded exam. If candidates can study the actual questions, scores measure memory of the bank rather than the subject. For practice, publish a separate practice set instead.

Can AI write the questions?

AI can draft questions quickly, especially to fill thin cells, but it makes factual mistakes and writes weak distractors. Treat AI questions as drafts and have a subject expert review each one before it can be drawn into a test.

Sources

• American Educational Research Association, American Psychological Association and National Council on Measurement in Education (2014). Standards for Educational and Psychological Testing, chapter 4: Test design and development. Washington, DC: AERA.
• Haladyna, T. M., Downing, S. M. and Rodriguez, M. C. (2002). A review of multiple-choice item-writing guidelines for classroom assessment. Applied Measurement in Education, 15(3), 309–333.
• Vale, C. D. (2006). Computerized item banking. In S. M. Downing and T. M. Haladyna (Eds.), Handbook of Test Development. Mahwah, NJ: Lawrence Erlbaum.