Precision & Recall
How many picks were right, and how many real ones got caught
- Precision is the share of everything flagged that was actually real. It measures how few false alarms there were.
- Recall is the share of everything real that got flagged. It measures how few misses there were.
- The two count different totals underneath. Precision divides by the flagged pile, recall divides by the real pile.
- Loosen the cutoff and recall rises while precision drops. Tighten it and the opposite happens. The two sit on a seesaw.
- Which one to chase isn't decided by performance — it's decided by which hurts more, missing something or chasing a false alarm.
Contents
1The analogy
A metal detector gate stands at the entrance to an event. Every bag that passes through it gets a beep if it's carrying metal, and only a beeping bag gets opened and checked.
Out of a hundred bags that came through today, five actually had metal in them. The gate beeped for ten. Opening those ten, only four really had metal — the other six beeped over keychains and soda cans.
Two ratios come out of this. Four real hits out of the ten that beeped — that's precision. Four out of the five real ones caught — that's recall. Both ratios describe the same day, but they divide by a different total.
Turning the sensitivity dial up on the gate makes it beep more. Fewer bags get missed, but more get opened for nothing. Turning it down does the reverse.
2In detail
The totals underneath are different
The top number in both ratios is the same: how many were correctly flagged. What differs is the bottom.
Precision divides by everything that got flagged, so a false alarm drags it down. Recall divides by everything that was actually real, so a miss drags it down.
That difference means either ratio on its own can be misleading. Beep on every bag and nothing gets missed, so recall hits 100 percent. Beep on only the one bag that's obviously suspicious and precision hits 100 percent. Neither gate is any use at all.
One dial shakes both at once
An AI system usually hands back a score, like "80 percent chance of metal," and calls it a hit once that score crosses some line. That line is the sensitivity dial.
Lower the line and even borderline cases get flagged, so fewer things get missed and more false alarms happen — recall rises, precision falls. Raise it and only the sure things get flagged, so false alarms shrink and misses grow.
Saying "the precision is such-and-such percent" is only half a sentence, then. It needs to come with which line it was measured at, and the real skill of a model only shows up by tracing a curve of how the two move as that line slides.
Which one hurts more sets the rule
Which of the two to prioritize is decided by the situation, not the technology.
Where a miss is dangerous, recall comes first. A false alarm just costs someone a second look, but a miss never gets revisited. Screening for hazardous items and catching early signs of equipment failure fall here.
Where every flagged item costs someone real time, precision comes first. Too many false alarms wear a team down until warnings get ignored altogether. Filtering spam and pulling the relevant files out of a big stack of documents both fall here.
Put both numbers on the same screen
That's why a scorecard lists both side by side. A report that shows only one leaves room for the other to be quietly broken.
The F1 score is what gets used when the two need folding into one number. Once folded, though, it's no longer visible which of the two was low, so it helps to keep the combined score sitting next to the original pair.
With more than two categories, score each one separately
When there are three or more categories to sort into, both numbers get calculated separately for each one. It's common for one category to be caught well and another almost missed entirely, and a single overall average buries that gap.
A category that's thin in the data tends to score lower on recall in particular. Laying out the per-category table and reading from the lowest row up is a habit worth building.
3More precisely
Both ratios come from four possible outcomes: correctly flagged as real, wrongly flagged as real, missed while real, and correctly left unflagged. Precision divides correct flags by every flag; recall divides correct flags by every real case. A table of all four cells is called a confusion matrix.
Sliding the cutoff line and tracing the two values as they move produces a precision-recall curve, and the area beneath it is used to compare models. On data where one category is rare, this curve gives a far more honest picture than plain accuracy does.
The analogy has one limit. The metal-detector gate can measure precision just by opening the bags that beeped, but recall needs checking every bag that passed through quietly too. Real evaluation works the same way: counting the misses means having the true answer for every single case ahead of time, which is exactly why recall is harder to measure than precision. In practice this means recall often gets estimated on a smaller, carefully checked sample rather than the full flow of data, since labeling every single case by hand rarely scales.
4Try it yourself
5Common misconceptions
It's easy to think high precision means a detector that catches a lot, but actually flagging only a handful of sure things and missing everything else still scores high on precision.
It's easy to think both numbers can be pushed to 100 percent at once, but actually moving the cutoff line raises one while lowering the other.
It's easy to think precision is just another name for accuracy, but actually accuracy is the share of every call that was correct, while precision is the share of flagged items that were real — different totals from the start.
7One-line summary
In shortPrecision is the share of picks that were real, recall is the share of real ones caught, and moving the cutoff line trades one against the other like a seesaw.
Spotted an error or have a better analogy? Suggest an edit · Last updated2026-09-02