citation density

Citation density counts units, not words. A unit is a span of roughly 100 to 400 tokens that makes a complete claim and reads correctly with nothing before or after it. The metric asks a blunt question: how many different ways can a model quote this page and have the quote still make sense? A page with one such span has one answer to sell. A page with seven has seven, and can win seven different queries without any of them competing.

The term it gets confused with is keyword density, and they are near-opposites in what they reward. Keyword density is a ratio — occurrences of a term divided by total words — invented for lexical ranking systems that matched strings. It is indifferent to whether the text means anything and it goes up when you repeat yourself. Citation density is indifferent to which words appear and goes up only when you add a genuinely separable claim. You can raise keyword density by padding. Padding lowers citation density, because it inflates the denominator without adding units.

It is also distinct from length, which is the assumption most content teams arrive with. Word count is the denominator here, not the numerator. Doubling a page's length while adding no new standalone claims halves its citation density and makes the page strictly less citable than it was. This is the single most counterintuitive consequence of the metric, and the one that breaks a decade of habit: the shorter of two pages is frequently the denser one.

Count the spans that pass the self-containment test, then divide by the page's word count over 500. A span passes if it names its own subject rather than relying on a pronoun pointing at the previous paragraph, states its claim inside the first sentence or two, and lands between roughly 100 and 400 tokens. Anything below that floor is a fragment with no supporting context; anything above it usually contains two claims tangled together and gets truncated at an arbitrary point when extracted.

The rubric grades it on a curve rather than a threshold. Full credit requires at least one extractable span per 500 words, and partial credit is awarded below that — a page at 0.6 is scored meaningfully above a page at zero, because the fix is different in kind. In practice, one per 500 is a floor rather than a target: pages that get cited repeatedly tend to run two to four, and the ceiling is set by how many genuinely distinct things the page has to say.

Take an 1,800-word buyer's guide. Its denominator is 3.6, since 1,800 divided by 500 gives the number of 500-word units. Now count the numerator honestly: the page has a 250-word introduction that says what the article will cover, four 300-word product sections that each begin with the words "it also", and a conclusion that recommends one option. Only the conclusion states a claim without depending on what came before it. That is one unit over 3.6, a density of roughly 0.28, and it fails the rubric floor.

Fixing it does not require more words — it requires the same words re-cut. Give each product section a question-shaped subhead, replace every opening "it also" with the product name and the claim, and move the recommendation's reasoning up into its first sentence. Nothing has been added. The page now has five spans that name their own subject and state a complete claim, which is five over 3.6, or about 1.4. It has gone from failing to comfortably above the floor with an edit rather than a rewrite.

The contrast worth internalising is what happens if you had instead followed the older instinct and expanded the guide to 3,000 words of additional background. The numerator stays at one, the denominator rises to 6, and density falls to 0.17. More effort, more words, a strictly worse page for retrieval. Both edits took roughly the same afternoon and they moved the metric in opposite directions.

• The wall of text. A page runs six hundred words without a break because the argument genuinely flows that way, and every sentence in it depends on the one before. Extraction cuts it at an arbitrary token boundary and produces a fragment nobody can quote, so a page full of good writing yields nothing. The fix is not to shorten the argument but to restate its subject at each turn, so any 200-token slice through the text still names what it is talking about.
• The subhead veneer. Someone reads that H2s improve extractability and sprinkles them through unchanged prose, so the page now has eight headings above eight paragraphs that still open with "this means that" and "as we saw above". The structure looks right in an outline and nothing underneath it survives being lifted out. The fix is to treat each subhead as a promise and rewrite the first sentence beneath it to deliver on that promise without referring backwards.
• The chunk-stuffing spiral. A team optimising for the number splits coherent explanations into artificially standalone stubs, repeating context in every paragraph until the page reads like a FAQ that lost its questions. Density rises and readability collapses, and the near-duplicate paragraphs start competing with each other for the same query. The fix is a cap on the ambition: aim for two to four units per 500 words, and stop when a paragraph would need to restate something a reader just read.

The relationship the table understates is that answer capsule and citation density are the same idea at two scales, and confusing them wastes effort. A capsule is the unit: one paragraph, pass or fail. Density is the count of passing units normalised by length. That means you never improve density directly — you improve individual paragraphs until they pass, and the ratio moves as a consequence. Keyword density sits in the table only because teams keep importing it, and the honest advice is to stop measuring it entirely.