Term frequency (TF) is the raw count of how many times a query term appears in a document — or, in BM25F, in the weighted blend of a document’s fields — measured before BM25’s saturation curve is applied.
| Term | Term Frequency |
|---|---|
| Category | Statistics and Evidence |
| Also known as | TF |
| Where it appears | BM25 and BM25F math |
What it means in RankGear
Term frequency is one of the raw inputs that feed the BM25 and BM25F relevance math RankGear uses to model how well a page’s text matches a query. In those formulas, a term that occurs more often pushes the relevance contribution up — but the count is fed through a saturation function first, so each additional occurrence adds less than the one before it. In plain BM25 the count comes from a single body of text; in BM25F it comes from a weighted combination of fields (title, headings, body and so on), where a match in a heavily weighted field counts for more than the same match buried in the body.
How to interpret it
Read term frequency as a raw ingredient, not a finished score. More is not linearly better: because BM25 deliberately saturates, the jump from one occurrence to three moves the contribution far more than the jump from twenty to forty, which barely moves it at all. Interpret TF alongside document length — BM25 normalizes for it, so the same count in a short page carries more weight than in a long one — and, for BM25F, alongside which field the term landed in. When you see a term frequency of zero, check whether the term is genuinely absent or whether the field simply was not measured.
| What you’re looking at | What it tells you |
|---|---|
| Raw term frequency | Straight count of occurrences before any adjustment; higher counts add progressively less. |
| After BM25 saturation | The count’s actual contribution to the score — flattened, so repetition has sharply diminishing returns. |
| After length normalization | The same count discounted for a long document or credited for a short one. |
Example
Suppose the query is “hydroponic nutrients” and two pages of roughly 800 words each are compared. Page A uses the phrase four times; Page B uses it twelve times. The raw term frequency is 4 versus 12 — a three-to-one gap — but once BM25 saturation is applied the difference in scoring contribution is much narrower than three to one, and if Page B is also considerably longer, length normalization shrinks it further still. The raw count looks decisive; the modeled relevance treats the two pages as far closer than the counts alone suggest.
Important considerations
- Term frequency is a pre-saturation input, not a ranking output. A high count does not translate into a proportionally high relevance contribution, and it does not make a page rank.
- BM25 length normalization means the same TF means different things in a short page versus a long one — always read it with document length in mind.
- In BM25F the count is a field-weighted blend, so where the term appears matters as much as how often.
- These are comparative modeling indicators, not Google’s scores. Statistical association with ranking does not establish causation, and small SERP samples can be unstable.
Related terms
Part of the RankGear glossary · how RankGear measures · the 870 factors.