1. The Concept, Plainly
"Write good content" is advice that skips the most important question: good by what measure? For AI citation, there are two distinct measures, and they are not the same thing. Keyword density counts how often a term appears in a piece of content. Evidence density counts how much of the content is verifiable — specific numbers, named sources, direct quotations. Research shows AI systems respond to the second signal, not the first.
2. Why This Matters Right Now
If you have ever followed SEO advice and added your target keyword more times, you have optimised for keyword density. That optimisation logic made sense for traditional search, which retrieved documents by matching terms. It does not transfer to AI citation. The content properties that research associates with AI citation probability are not about repetition — they are about how much of your content is specific and checkable. A writer who understands SEO does not automatically understand evidence density, and the gap between them is now measurable. You can run a score on your own content right now and see exactly where you stand.
3. The Mechanism
Aggarwal et al. (2024) — a peer-reviewed study published at KDD, one of the largest data science research conferences — measured which content properties predicted whether AI systems would cite a source by name. Keyword frequency did not appear as a predictor. The properties that did were different in kind: statistics presence was associated with approximately 31% higher citation probability; source attribution was associated with approximately 30% higher citation probability.
These are not soft associations. They describe a category difference: content that gives AI systems something specific and checkable to work from versus content that does not.
A second study — Zhang et al. (2026), a preprint not yet peer-reviewed, so treat its findings as directional — analysed 21,143 citations across three AI platforms and found that high-influence pages differed from low-influence pages across six dimensions, all of which map to evidence richness and structural completeness. High-influence pages were on average 11.44 times longer. They contained 12.50 times more headings. Statistics presence correlated with approximately 61% higher absorption in that study.
None of those properties describe keyword frequency. They describe the density of verifiable, structured information per piece of content.
This is the gap "write good content" does not close. The instruction is not wrong — it is just incomplete. Evidence density is the specific property of "good" that AI systems appear to respond to. You can have high keyword density and low evidence density at the same time: content that repeats a term frequently but contains few attributed statistics, few definitions, and no named sources. That content can rank in traditional search while being largely invisible to AI citation.
The working hypothesis — consistent with how AI retrieval systems function — is that AI systems reward evidence-dense content because they are retrieving to construct answers, not to match keywords. A keyword count does not indicate usefulness as a source. An attributed statistic does. The researchers note this mechanism is not yet experimentally confirmed; the studies measure correlation, not cause.
Evidence density is a measurable property. Raising it requires genuine additions: more attributed statistics, more source references, more definitional sentences that explain what things are. Unlike keyword density, you cannot improve it by repetition.
4. Try It Now
You can test the gap between keyword density and evidence density in your own content with a short AI prompt. The exercise takes under five minutes.
Paste this prompt:
"Read the following passage and rate its evidence density on a scale of 1 to 5, where 1 means the passage contains only general assertions with no verifiable claims, and 5 means the passage contains a high proportion of specific statistics, named sources, and direct quotations. Give your rating first, then explain what specific elements in the passage raised or lowered your score. Identify any sentences that make a claim but contain no verifiable information.
[Paste 200–400 words of your own content here]"
What to look for in the output: The AI will give a score and then explain which sentences counted as verifiable (those with numbers, named studies, specific dates, or direct quotations) and which did not. The sentences it flags as non-verifiable are your evidence gaps — claims your reader and the AI have no way to check. If your expected score and the AI's score diverge, that gap is the distance between what you assumed counted as evidence and what actually does.
Once you have your AI-generated rating, run the same content through the Evidence Density Score tool. It measures the same properties — statistics, source attributions, definitional sentences, comparative language, and structural richness — and returns a 0–100 score with a prioritised list of what to improve first.
5. The One Thing to Remember
Keyword density and evidence density are not the same metric — one counts term repetition, the other counts how much of your content is specific and verifiable, and AI systems respond to the second.
6. Go Deeper
The Field Notes post on evidence density versus keyword density covers the full research findings, the confidence-level caveats for each study, and the practical implications for content strategy: Evidence Density vs. Keyword Density: What Research Says About AI Citation Predictors.
Measure your evidence density now.
The Evidence Density Score measures statistics, source attributions, readability, and structural richness in a single 0–100 score — with a prioritised fix list ranked by evidence confidence. No signup.