If you searched for average precision, you have probably met the term in machine learning, information retrieval, or search evaluation, and you want a clear explanation of what it measures, how it is calculated, and when it is useful. In simple terms, average precision is a way of scoring a ranked list of results: it rewards a system that puts relevant items near the top of the list, and it penalises a system that buries relevant items beneath irrelevant ones. This guide explains average precision step by step, with worked examples, common confusions, and practical advice on when to use it and when another metric fits better.
Average precision matters because many real systems return ranked lists, not single answers. A search engine returns pages in order, a recommendation system returns suggestions in order, and a detection system returns candidate objects in order of confidence. In all of those cases, the order matters as much as the set of items returned. A system that returns ten relevant items at the bottom of a long list is less useful than a system that returns the same ten items at the top. Average precision captures that difference in a single number, which is why it appears so often in research papers, evaluation reports, and model comparisons.
What Average Precision Means
Average precision is, at its core, an average of precision values taken at the positions where relevant items appear. Precision, in this context, is the fraction of items returned so far that are relevant. If the first item in a list is relevant, precision at that point is one out of one. If the first two items contain one relevant item, precision at that point is one out of two. Average precision looks at precision only at the ranks where a relevant item occurs, and then averages those values. That focus on relevant positions is what makes average precision sensitive to ranking order.
Consider a simple example. Suppose a search returns five items, and the relevant ones are at positions one, three, and five. Precision at position one is one out of one, or 1.0. Precision at position three is two out of three, or about 0.67. Precision at position five is three out of five, or 0.60. The average of those three values is about 0.76, and that is the average precision for that query. If the relevant items had instead appeared at positions one, two, and three, the average precision would be 1.0, because precision at each relevant position would be perfect. If they had appeared at positions three, four, and five, the average precision would be much lower. The same set of relevant items therefore gets a different score depending on order, which is exactly the behaviour we want from a ranking metric.
Precision and Recall in Plain Language
To understand average precision fully, it helps to understand precision and recall together. Precision answers the question: of the items the system returned, how many were actually relevant? Recall answers a different question: of all the relevant items that exist, how many did the system find? A system can have high precision but low recall if it returns only a few items and most of them are relevant, while missing many other relevant items. It can have high recall but low precision if it returns almost everything, including many irrelevant items. Average precision balances these ideas by looking at precision as recall increases, rank by rank.
A precision-recall curve plots precision against recall as you move down a ranked list. Average precision can be thought of as an approximation of the area under that curve, although the exact calculation depends on the definition used. Some definitions interpolate precision values, smoothing the curve, while others use the raw precision at each relevant rank. Different software libraries may implement slightly different variants, so when you compare average precision values from different sources, check which definition each source used. Small differences in definition can produce small differences in score, even for the same ranked list.
How Average Precision Is Calculated
The calculation of average precision follows a clear sequence. First, you need a ranked list of items and a judgement for each item: relevant or not relevant. Second, you walk down the list from the top, keeping a running count of relevant items found so far. Third, at each position where the item is relevant, you calculate precision at that position, which is the number of relevant items found so far divided by the position number. Fourth, you add up those precision values and divide by the total number of relevant items. That total is the number of relevant items that exist for the query, not just the number returned, in the standard definition.
That last point deserves emphasis, because it affects how average precision handles missing items. If a system returns only two of five relevant items, and both are at the top, the sum of precision values is two, but the divisor is five, giving an average precision of 0.40, not 1.0. The metric therefore penalises systems that fail to retrieve relevant items, as well as systems that rank them poorly. Some variants divide by the number of relevant items retrieved instead, which gives a different, more forgiving score. When you read a report, look for a note about which divisor was used, because it changes the meaning of the number.
A worked example makes the steps concrete. Imagine a query with four relevant documents in total. A system returns six documents, with relevance pattern relevant, irrelevant, relevant, irrelevant, irrelevant, relevant. Precision at the first relevant position is one out of one, or 1.0. Precision at the second relevant position, rank three, is two out of three, or 0.67. Precision at the third relevant position, rank six, is three out of six, or 0.50. The sum is about 2.17. Dividing by the total number of relevant documents, four, gives an average precision of about 0.54. The missing fourth relevant document, which was never returned, is what pulls the score down from a higher value.
Average Precision Across Many Queries
Average precision is usually calculated for a single query, a single class, or a single evaluation case. When you have many queries, you average those per-query values to get mean AP, often written as mAP. Mean AP is the headline number in many search and detection benchmarks, because it summarises performance across a whole test set in one figure. If you see mAP in a paper or a leaderboard, it almost always means the mean of average precision values across queries or classes, although the exact averaging method should still be checked.
In object detection, mean AP has its own conventions, including rules about how closely a predicted box must overlap a true box to count as correct, and how scores are averaged across object classes. Those conventions vary between benchmarks, so a mean AP value from one benchmark is not directly comparable to a value from another benchmark unless the conventions match. The same caution applies to search evaluation: different test collections, different relevance judgements, and different definitions of average precision can all shift the numbers. Always compare like with like.
Average Precision in Search and Recommendation
In search evaluation, AP is valued because it reflects the user experience of scanning results from the top down. Users rarely read an entire result list; they start at the top and stop when they have found enough, or when they lose patience. A system that front-loads relevant results therefore serves users better, even if another system eventually returns the same relevant documents further down. AP rewards that front-loading, and it does so without requiring you to choose an arbitrary cut-off point, such as the top ten results only.
In recommendation, the same logic applies, with a useful twist. Recommendation lists are often short, and users may only look at the first few suggestions. AP still works for short lists, but its behaviour at small scales deserves care: with only one or two relevant items, a single swap in order can change the score dramatically. That sensitivity is informative, but it also means that AP values from small recommendation tests can be noisy. Averaging across many users, and reporting results alongside other measures such as simple precision at a fixed cut-off, usually gives a more stable picture.
AP is also used in classification problems where the model outputs a ranked score rather than a hard yes-or-no decision. By sorting items by predicted score and treating the top items as predictions at different thresholds, you can compute AP as a summary of how well the ranking separates positive cases from negative ones. In that setting, AP is closely related to the area under the precision-recall curve, and it is often preferred over accuracy when positive cases are rare, because accuracy can look impressive even when a model finds almost no positive cases.
AP Strengths and Limitations
The main strength of AP is that it combines ranking quality and retrieval completeness in one number, without forcing you to pick a threshold in advance. It is widely understood, widely implemented in evaluation libraries, and easy to explain to non-specialists using the worked examples in this guide. It also handles queries with different numbers of relevant items in a principled way, because each query’s score is normalised by its own total of relevant items before averaging across queries.
AP also has limitations. It depends heavily on relevance judgements, and if those judgements are incomplete, noisy, or inconsistent, the scores will reflect those flaws. It gives equal weight to each relevant item, which may not match user preferences: in some applications, finding one excellent result matters far more than finding several mediocre ones, while in others, coverage matters most. It also summarises a whole ranking in one number, which hides detail: two systems can have the same AP with quite different strengths and weaknesses, so a single number should not be the only basis for a decision.
| Metric | What it captures | When to prefer it |
|---|---|---|
| AP | Ranking quality plus completeness | Comparing ranked systems without a fixed cut-off |
| Precision at ten | Quality of the top ten results | When users only see the first page |
| Recall | Share of relevant items found | When missing items is costly |
| Normalised discounted cumulative gain | Graded relevance and ranking | When relevance has degrees, not just yes or no |
For readers exploring related everyday concepts, see this guide to what an earned run average means and this overview of average returns over a long period. Both deal with averages in different fields, and both show why the definition of an average matters as much as the number itself.
Common Mistakes With AP
One common mistake with AP is confusing it with simple accuracy or with mean precision. Accuracy measures the share of all judgements that are correct, including correct rejections, while AP focuses on the ranking of relevant items. Mean precision, if used loosely, might mean the average of precision values at arbitrary points, which is not the same thing. Using the terms precisely matters, especially in reports that other people will rely on to choose between systems.
Another mistake is comparing AP values across different test sets, different relevance definitions, or different implementation variants without noting the differences. A score of 0.70 on one collection and 0.65 on another does not necessarily mean the first system is better, because the collections may differ in difficulty. A fair comparison keeps the test set, the relevance judgements, and the calculation method fixed, and changes only the system being evaluated. That discipline is what makes evaluation trustworthy.
A third mistake is reporting AP alone, without context. A single number cannot show whether a system fails on particular types of query, whether it performs unevenly across groups of users, or whether its errors are harmless or costly. Good evaluation reports AP alongside other measures, breaks results down by query type where possible, and includes examples of failures. That extra context turns a leaderboard number into genuine understanding.
Frequently Asked Questions About AP
Is AP the same as mean AP?
No. AP is the score for one query, class, or case. Mean AP is the average of AP values across many queries or classes. Mean AP is the headline figure in many benchmarks, but it is built from individual AP values.
Can AP be one?
Yes, AP reaches its maximum value of one when every relevant item is retrieved and every relevant item appears before any irrelevant item. In practice, perfect scores are rare on realistic test sets, and a perfect score on a small or easy test set should prompt a check of whether the test is challenging enough.
Does AP need a threshold?
No, and that is one of its advantages. AP considers the full ranked list, so you do not need to choose a cut-off such as the top ten in advance. If your application only ever shows a fixed number of results, however, a metric at that cut-off may be a useful companion measure.
Why do different tools give different AP values?
Different tools may use different variants: interpolated versus raw precision, dividing by total relevant items versus retrieved relevant items, or different handling of ties. Those choices produce small differences. Check the documentation of the tool you use, and use the same tool and settings when comparing systems.
Conclusion: AP
AP is a compact, widely used way to judge ranked lists: it averages precision at the positions where relevant items appear, rewarding systems that rank relevant items early and retrieve them completely. Its strengths are its sensitivity to order, its freedom from arbitrary thresholds, and its clear interpretation through worked examples. Its limitations are its dependence on good relevance judgements, its equal weighting of relevant items, and the fact that a single number hides detail. Used carefully, alongside other measures and with consistent definitions, AP remains one of the most practical tools for evaluating search, recommendation, and ranking systems.
