Recall is the share of a bucket’s human-labelled members that the classifier successfully retrieves — of everything that truly belonged, how much did the model actually find.
| Term | Recall |
|---|---|
| Category | Statistics and Evidence |
| Also known as | Sensitivity |
| Where it appears | Semantic enrichment evaluation |
What it means in RankGear
Recall surfaces when RankGear evaluates a semantic enrichment classifier against a set of human-labelled examples. Each labelled item either belongs to a bucket or it does not; the classifier makes its own call on every one. Recall counts only the items that genuinely belong and asks what fraction of them the classifier caught. A recall of 0.80 means the model retrieved four out of every five true members and missed the fifth. It is the direct measure of how much of the real signal an evaluated pass is leaving on the table.
How to interpret it
Read recall alongside its counterpart, precision, never on its own. Recall answers “how much did we find,” while precision answers “how much of what we found was right” — a classifier can score high on one by sacrificing the other, so a lone recall figure is easy to misread. Anchor the number to sample size before trusting it: recall computed over a dozen labelled members is noisy, and a single missed item can swing it by several points. Check whether a low value reflects a genuinely conservative model or simply a threshold set too high, and confirm the metric is measured rather than defaulting to zero because the bucket had no labelled members to retrieve.
| Recall value | What it tells you |
|---|---|
| Near 1.0 | Almost every true member is retrieved; few are missed, though precision may be paying for it |
| Mid-range | A meaningful share of true members slips through; weigh against precision and sample size |
| Near 0.0 | Most true members are missed — or the bucket had too few labelled members to measure reliably |
Example
A semantic enrichment run is scored against 200 human-labelled pages, 50 of which annotators marked as belonging to an “informational intent” bucket. The classifier tags 40 of those 50 correctly and misses 10, giving a recall of 0.80. A second run over the same labels reports the same 0.80 recall but retrieves a different 40, missing a harder-to-classify 10. The averages match, yet the spread of what each run catches differs, which changes how confidently you can pick a working threshold for production.
Important considerations
- Recall and precision trade off against each other; report and read them together, and consider the F1 score when you need a single balanced figure.
- Small labelled samples make recall unstable — a few missed items can move the number sharply, so treat figures from tiny buckets as directional at best.
- Distinguish a true zero (nothing retrieved) from a missing measurement (no labelled members existed to retrieve) before drawing any conclusion.
- Recall describes a classifier’s behaviour on labelled data, not a page’s standing with Google. Statistical association does not establish causation, and small SERP samples can be unstable.
Related terms
Part of the RankGear glossary · how RankGear measures · the 870 factors.