# Relevancy

Every [reference](/how-questions-are-answered/references/) carries a relevance
rating — how well that particular piece of material matches the question being
asked.

| Relevance | |
|---|---|
| **Excellent** | 80–100% |
| **Very good** | 60–79% |
| **Good** | 40–59% |
| **Ok** | 20–39% |
| **Low** | Below 20% |

## How it is worked out

Relevance is about meaning rather than wording. Candidate material is found by
semantic similarity — matching the *sense* of the question rather than its words
— and then scored on how well each piece actually answers what was asked.

This is why a policy passage about "encryption of stored data" can rate highly
against a question asking "is customer data encrypted at rest?", despite sharing
almost no vocabulary.

**Treat relevance as a rough guide:** Relevance tells you how close a piece of material looked to the question. It does
not tell you whether the resulting answer is any good.

A high relevance score on a passage that turns out to be about a different system,
or a low score on a passage that happens to contain exactly the right sentence,
are both entirely possible. Take it with a pinch of salt.

For judging an answer, [confidence](/how-questions-are-answered/confidence/) is a
much better indicator — it is assessed against the answer itself, not just the
material behind it.

## What it is useful for

Relevance is most useful when you are checking *why* an answer came out the way
it did. Scanning the references and their ratings shows you what ResponseHub
thought was pertinent, which usually explains an odd answer quickly — often it
has latched onto a passage covering a related but different topic.

It also drives one piece of behaviour directly: a knowledge base item matching
closely enough is used
[verbatim](/how-questions-are-answered/verbatim/).

## Next steps

- [Confidence](/how-questions-are-answered/confidence/)
- [Verbatim answers](/how-questions-are-answered/verbatim/)