BM25 Document Length is the token count the BM25 ranking function treats as a page’s length when it normalizes term-frequency scores. RankGear reports it on its own, separate from the other page word counts, so the length BM25 actually uses is never confused with a plain body count.
| Term | BM25 Document Length |
|---|---|
| Category | Statistics and Evidence |
| Also known as | BM25 |D|, Document Length |D| |
| Where it appears | BM25 Drafts word-count table |
What it means in RankGear
You see this value in the word-count table of a BM25 Drafts run. It is the |D| that BM25 feeds into its length-normalization step — the count of tokens the scorer sees after it tokenizes the page, not the word count your editor or CMS would show. RankGear keeps it in its own column precisely because the two rarely match: BM25 length reflects how the retrieval model measures the document, while a body word count reflects how a human reads it. Compared against the drafts and reference pages in the same run, it tells you how long BM25 thinks the page is relative to everything else on the table.
How to interpret it
BM25 uses document length to damp the edge a long page would otherwise gain by simply repeating a term. A draft well above the run’s average length has its term frequencies discounted more heavily, so extra words buy progressively less; a draft near or below the average is discounted less. Read the number relative to the other rows in the same table rather than as an absolute target, and compare like with like within a single run — the average it normalizes against is set by the specific pages that run pulled.
| BM25 length vs. run average | What it tells you |
|---|---|
| Well below average | Term frequencies are discounted little; the page reads as short to BM25. |
| Near the average | Neutral normalization — term counts score close to face value. |
| Well above average | Term frequencies are discounted more; added length yields diminishing BM25 benefit. |
Example
A BM25 Drafts run compares three drafts against the pages currently ranking for the target query. Draft A comes in at 1,850 tokens, close to the set’s average length, so its keyword frequencies are scored near face value. Draft B is 4,200 tokens — well above average — so BM25 discounts its repeated terms and the extra length adds little to its fit. Reading the word-count table, you can see that padding Draft B further will not raise its BM25 score; trimming it toward the set’s typical length is the more useful move.
Important considerations
- This is a token count as BM25 measures it, not your editor’s word count — expect the two numbers to differ, and do not treat one as a correction of the other.
- The value is comparative within one run: the average length it normalizes against depends on which pages that run pulled, so small or shifting SERP samples can move it.
- BM25 is a retrieval-model indicator, not a Google score. Matching the average length does not make a page rank; statistical association is not causation.
Related terms
Part of the RankGear glossary · how RankGear measures · the 870 factors.