By Eric, BlinkBench Founder · Last reviewed: October 2026
Two people can finish the same word-recognition run with the same number of correct answers and very different memories. What separates them is how willing each was to say "seen" when a word only felt familiar. This guide explains why Verbal Memory reports d′ and a separate bias figure instead of a streak, using three worked runs computed with the site's own code.
The two ways to be wrong
Each word in the test is either a repeat or a new word, and you answer "seen" or "new". That gives four outcomes: a hit (a repeat you called seen), a miss (a repeat you called new), a false alarm (a new word you called seen) and a correct rejection (a new word you called new). Half of all trials are repeats (REPEAT_PROBABILITY is 0.5 in lib/verbalMemory.js), and three mistakes end a run, a miss and a false alarm costing the same.
A meta-analysis of 232 experiments on recognition memory splits what these answers carry into two parts. In its words, "Discrimination or sensitivity captures how well people can successfully discriminate old (studied) items", while response bias "captures the criterial level of evidence at which a learner terms an item old". Sensitivity is about the memory. Bias is about where you set the bar for saying yes.
A plain count of correct answers blends the two. Someone who says "seen" at the slightest familiarity collects more hits and more false alarms. Someone cautious collects fewer of both. Neither number alone says which memory was better.
What d′ and c do with the same four counts
The method behind this site's scoring is signal detection theory. In the formulas recorded in this site's sources, d′ is the z-score of the hit rate minus the z-score of the false-alarm rate. The bias figure, c, is minus half the sum of those two z-scores. A negative c means leaning toward "seen", a positive c means leaning toward "new", and zero is neutral. Because d′ subtracts the false-alarm rate from the hit rate, it is in principle unaffected by where the person set that bar.
Here are three imaginary runs of 20 repeats and 20 new words each, scored by lib/sdt.js, the same code the page uses. The code adds 0.5 to the hit and false-alarm counts and 1 to each trial total before computing rates, so a perfect hit rate does not produce an infinite score.
| Run | Hits | False alarms | d′ | c |
|---|---|---|---|---|
| Answers "seen" to every word | 20 of 20 | 20 of 20 | 0.00 | −1.98 |
| Cautious | 14 of 20 | 2 of 20 | 1.68 | 0.34 |
| Liberal | 18 of 20 | 8 of 20 | 1.42 | −0.47 |
The first row is the case a streak cannot see. Calling every word "seen" never misses a repeat, and the hit rate is a perfect-looking 20 of 20. The d′ is 0.00, because the false-alarm rate is just as high. In the real test that approach ends quickly, since every new word is a false alarm that costs a life. The table shows what the maths says about it.
The second and third rows are the interesting pair. The liberal run has the higher hit count, 18 against 14. The cautious run has the higher d′, 1.68 against 1.42, because it said "seen" to far fewer words that were new. The c column shows why the two differ: the liberal run sits at −0.47 and the cautious one at 0.34.
Why the correction leans low
The corrected d′ in the table is deliberately conservative. This site's source notes on a 1995 comparison of corrections say the log-linear rule used here "always UNDER-estimated population d′", whereas the alternative 1/(2N) rule "was more biased and could either over- or under-estimate". The effect is that a visitor looks slightly worse than they are rather than slightly better.
Reading your own number
A run with few trials has a wide band. The page prints a standard error beside d′ and shows no d′ at all below five trials of either kind (MIN_TRIALS_PER_CLASS). In the table above, the 20-and-20 runs carry standard errors of about 0.45 for the cautious and liberal runs. That is large enough that the 1.68 and 1.42 are not clearly different from each other, so the table shows how bias can move a streak, not that one imaginary person remembers better.
For context, the same meta-analysis reports a mean d′ of 1.85 for young adults, with older adults 0.46 d′ units worse at that point. That is a figure for the average of many experiments, with varying materials and trial counts. This site does not turn it into a percentile for you, because the spread across experiments is not the spread across people.
The result also reflects the setup as well as the person. The test is a continuous run in which every word you judge joins the list you must remember, so a long run and a short one are not measuring quite the same thing. Your screen and input device are part of what was measured too.
Frequently asked questions
Why not just count correct answers?
Because a count of correct answers cannot tell a good memory from a habit of saying "seen". Two people with the same count can have different d′ values and different bias, as the table above shows.
What does a c near zero mean?
It means you were not leaning either way: your hit rate and false-alarm rate were balanced around the midpoint. Negative c means you leaned toward "seen" and positive c toward "new".
Is a higher d′ always better?
A higher d′ means "seen before" and "new" felt further apart to you on this run, with the band printed beside it showing how firmly the trial count supports the figure. It is not a verdict on you, and this page states that it is not a cognitive assessment.
Why is the first row's d′ exactly zero?
When the hit rate and the false-alarm rate are equal, their z-scores cancel. Zero is the case where the two kinds of word were indistinguishable and every answer was a guess.
Sources
- Fraundorf, S. H., Hourihan, K. L., Peters, R. A., & Benjamin, A. S. (2019). Aging and recognition memory: A meta-analysis. Psychological Bulletin, 145(4), 339-371.
- Stanislaw, H., & Todorov, N. (1999). Calculation of signal detection theory measures. Behavior Research Methods, Instruments, & Computers, 31(1), 137-149.
- Hautus, M. J. (1995). Corrections for extreme proportions and their biasing effects on estimated values of d’. Behavior Research Methods, Instruments, & Computers, 27(1), 46-51.