Precision is the share of items RankGear sorted into a given bucket that human review confirms actually belong there. It answers a single question about every positive call the system makes: when it says “yes,” how often is it right?
| Term | Precision |
|---|---|
| Category | Statistics and Evidence |
| Also known as | Positive Predictive Value |
| Where it appears | Semantic enrichment evaluation |
What it means in RankGear
When RankGear enriches a page semantically, it assigns candidates — entities, topics, or classifications drawn from the content — into labelled buckets. Precision measures how trustworthy those assignments are. Take every candidate the system placed in a bucket, compare each against a human-confirmed label, and precision is the fraction that the human agrees with. A precision of 0.90 means nine of every ten items the system flagged for that bucket genuinely belonged; the tenth was a false positive that slipped in.
How to interpret it
Read precision as a rate between 0 and 1 (or 0–100%), and always read it alongside the sample it was computed on. A high precision figure drawn from only a handful of candidates is fragile — one or two corrections can swing it sharply. Precision also tells you nothing about what the system missed: a model that labels only the candidates it is certain about can post near-perfect precision while ignoring most of the real matches. That blind spot is exactly what recall captures, which is why the two are read as a pair rather than in isolation. Before comparing precision across runs, confirm the value is genuinely measured rather than missing, and that a reported zero means “no correct positives” and not “not evaluated.”
| Precision value | What it tells you |
|---|---|
| Near 1.0 | Almost every positive call is confirmed — few false positives, but says nothing about coverage. |
| Mid-range | A meaningful fraction of positive calls are wrong; treat the bucket’s assignments with caution. |
| Near 0 | Most positive calls are unconfirmed — the classifier is over-flagging, or the sample is too small to trust. |
Example
Suppose an enrichment run tags 50 candidates as belonging to the “product feature” bucket. A reviewer checks them and confirms 44; the other six were unrelated phrases the model over-matched. Precision for that run is 44 ÷ 50 = 0.88. A second run over the same page might report the same 0.88 average but concentrate its mistakes differently — all six misses in one closely related sub-topic rather than scattered — which changes how confidently you can rely on that bucket even though the headline number is identical.
Important considerations
- Precision only judges the items that were flagged. It cannot tell you about correct matches the system never surfaced — pair it with recall to see both sides.
- Small evaluation sets make precision unstable; a single relabelled item can move the number several points, so weigh the sample size before drawing conclusions.
- Precision is a comparative indicator of classification quality inside RankGear’s evaluation, not a Google ranking score. A higher figure does not make a page rank — statistical association is not causation, and small SERP-derived samples can be noisy.
Related terms
Part of the RankGear glossary · how RankGear measures · the 870 factors.