How the test is built and scored

Summary

• What it is: 30 figural matrix puzzles measuring reasoning with unfamiliar patterns, taken unsupervised, with a 25-minute limit.
• Scale: deviation IQ, mean 100, standard deviation 15.
• What you get: an estimate, a 95% interval of about 11 points either side, and a percentile.
• Range: the test reports scores from 70 to 135. Beyond that it says “under 70” or “135+”.
• Norms: v1 (provisional), November 2026. See “The status of the norms” below. This is the most important limit.

Where the puzzles come from

Every puzzle is original. None is taken from Raven’s Progressive Matrices or the Wechsler scales, which are proprietary, or from the International Cognitive Ability Resource, whose items are licensed for academic research only.

Instead, each puzzle is generated from the rule types that research has identified in matrix problems (Carpenter, Just & Shell, 1990):

• Constant: a feature stays the same along a row.
• Progression: a feature steps by a fixed amount along a row (one more shape, one size larger, a quarter turn).
• Distribution: three values each appear once in every row.
• Overlay: the third figure combines the first two: everything in either, only what they share, or only what differs.

Easy puzzles use one rule. The hardest stack four or five across the shape, its size, its shading, how many there are, and a frame around it. The number and kind of rules is the main driver of difficulty in published matrix tests, explaining roughly half to four fifths of the variation between items (Schroeders et al., 2026; Preuß et al., 2025).

Wrong answers that cannot be eliminated by eye

Each of the seven wrong options is the right answer with one or two ruled features set to the value a plausible mistake would give. Every feature is therefore right in several options and wrong in several, so the answer cannot be found by looking for the odd one out. Tests built this way show better validity (Pallentin et al., 2023). The position of the right answer is balanced across the test.

The same test for everyone

The 30 puzzles are generated once from a fixed seed. Everyone answers the same items in the same order, which is what allows them to be calibrated.

How answers become a score

The test is scored with a three-parameter logistic item response model. Each puzzle has a difficulty; the chance of a correct answer rises with ability, and never falls below one in eight because there are eight options and a blind guess can be right.

• Ability is estimated as the average of its probability distribution given your pattern of right and wrong answers, starting from the assumption that abilities in the population follow a normal curve.
• IQ is 100 plus 15 times that ability estimate.
• The 95% interval comes from the spread of the same distribution, so it reflects your particular pattern of answers.
• Solving hard puzzles counts for more than solving easy ones. Getting hard ones right while missing easy ones looks like guessing and widens the interval.
• A blank answer is scored as wrong.

Discrimination is fixed at 1.3 for every item and the guessing parameter at 0.125.

The margin of error

Short reasoning tests are less precise than long ones. Published matrix tests of 20 to 48 items report reliabilities between about 0.80 and 0.93 (Pallentin et al., 2023, 2024). At 30 items, the interval this test reports is about 11 points either side of the estimate in the middle of the range, and wider toward the ends.

Your result page shows the interval for your sitting. Read the score as “probably somewhere in this range”.

The status of the norms

The item difficulties in this version are provisional. They are set from each puzzle’s design (how many rules, and which kinds) and have not yet been estimated from a calibration sample. That means the conversion from your answers to an IQ is a reasoned estimate, not yet an empirically normed one.

What this means in practice: the ordering of people is likely to be sound, because harder puzzles really are harder. The exact number could be off by several points in either direction on top of the interval shown, across the whole range.

The plan is to replace the provisional values with ones estimated from a quota sample of adults matched to census age and education, and to publish the sample size, reliability and date when that is done. Norming on the site’s own visitors alone would be wrong: people who seek out an IQ test are not a cross-section of the population, and a test normed on them would understate everyone.

What pushes a score up or down

• No supervision. Across 109 comparisons, unsupervised ability scores ran higher than supervised ones by about a fifth of a standard deviation, roughly three points (Steger et al., 2018). Figural puzzles are among the hardest to look up, which limits this.
• Small screens. In one national study, phone users scored lower than computer users on a cognitive battery by nearly half a standard deviation (Lynham et al., 2023). The result page flags a sitting taken on a narrow screen.
• Practice. Retaking the same puzzles raises the score without raising ability.
• Time. The 25-minute limit is generous, but rushing lowers scores and the result page says so when a sitting was very fast.

What the test cannot tell you

• It measures one ability. Figural reasoning is closely related to general intelligence, but a full assessment also covers verbal ability, memory and processing speed.
• It is not a clinical assessment and cannot diagnose giftedness, intellectual disability, or any learning or neurological condition.
• It cannot be used for Mensa. Mensa’s own guidance is that “online tests cannot be used for admission to Mensa”.
• It should not be used for hiring, school placement, or any legal or medical decision.
• It is for adults. The norms assume adult test-takers; it is not suitable for children.

Privacy

Answers are scored in your browser and are not sent to us. A result exists only in the link shown on the result page. Details are on the privacy page.

References

• Carpenter, Just & Shell (1990). What one intelligence test measures: a theoretical account of the processing in the Raven Progressive Matrices Test. Psychological Review.
• Lynham et al. (2023). Cardiff Online Cognitive Assessment in a national sample: cross-sectional web-based study. Journal of Medical Internet Research.
• Matzen et al. (2010). Recreating Raven’s: software for systematically generating large numbers of Raven-like matrix problems with normed properties. Behavior Research Methods.
• Pallentin et al. (2023). Construction and validation of the HeiQ: an operation-oriented figural matrices test. Journal of Intelligence.
• Pallentin et al. (2024). Validation of the short parallel and extra-short form of the Heidelberg Figural Matrices Test (HeiQ). Journal of Intelligence.
• Preuß et al. (2025). Using cognitive diagnostic models to evaluate the two-process theory of matrix reasoning. Journal of Intelligence.
• Schroeders et al. (2026). Developing BOLT: a matrix test using Boolean operations to assess logical thinking. Intelligence.
• Steger et al. (2018). A meta-analysis of test scores in proctored and unproctored ability assessments. European Journal of Psychological Assessment.