For two decades, keyword density was the dominant content optimisation metric. The logic was simple: if AI systems (in this case, search engines) retrieve content based on keywords, then including more keywords more frequently makes content more findable.
That logic is now outdated for a new generation of AI systems — and the research explains why. AI citation behaviour responds to fundamentally different signals than keyword frequency. Understanding the distinction is the starting point for any content strategy built for AI visibility.
What keyword density optimised for
Keyword density optimisation was a response to keyword-based retrieval systems. Traditional search engines retrieved documents by matching query terms to document terms. Documents with higher keyword frequency were retrieved more consistently — which is why keyword density became a metric worth optimising.
The signal was simple, measurable, and gameable. Content could be optimised by adding target keywords more frequently, placing them in headers, repeating them in the conclusion. The optimisation aligned reasonably well with the retrieval system's logic — even if it often degraded content quality.
What AI citation behaviour responds to
AI systems — large language models generating answers with citations — do not retrieve content based on keyword frequency in the same way. They assess content along multiple dimensions simultaneously. The research documents this clearly.
Statistics in content increased AI citation probability by approximately 31%. Source attribution increased it by approximately 30%. Keyword frequency does not appear as a significant predictor in either major study.
Per Aggarwal et al. (2024), the properties most associated with AI citation are: statistics presence, source attribution, readability at grade 8–10, and structural richness — headings, organisation, coherent paragraph structure. None of these are keyword frequency signals.
Zhang et al. (2026) adds to this picture with absorption signals: word count (substantive depth), heading density, paragraph density, definitional language, and comparative language. Again — no keyword frequency signals in the top differentiators.
Why the divergence makes sense
AI systems are selecting content for a different purpose than keyword-based retrieval systems. A search engine retrieves documents that contain the query terms. An AI system retrieves content that will support the construction of a confident, accurate, well-sourced answer.
Content that is confident and accurate tends to have: specific numbers (statistics), identified sources (attribution), clear explanations (readability and definitional language), and comparative context (comparative language). These are properties of well-evidenced writing — not keyword-dense writing.
Keyword-dense content, by contrast, often trades readability for repetition. It places keywords in positions that serve retrieval, not comprehension. It may actually score poorly on the evidence density dimensions that AI systems appear to weight.
The practical divergence. A piece of content can have high keyword density and low evidence density simultaneously. It can rank well in traditional search while being largely invisible to AI citation. The two systems are optimising for different inputs — and content strategy needs to choose which signal set it's building for.
What evidence density actually measures
Evidence density is a measure of how much verifiable, well-structured content a document contains — specifically the properties research links to AI citation and absorption.
Unlike keyword density, evidence density cannot be improved by repetition. You cannot add more statistics by repeating the same statistic twice. Evidence density requires genuinely more evidence: additional quantified findings, additional source attributions, additional research-backed claims.
This makes evidence density a fundamentally harder signal to game — and a fundamentally more honest measure of content quality. Content with high evidence density is, by definition, content that contains more verifiable, specific, well-sourced claims. That is content worth citing.
Three practical implications
Implication 1: Research investment pays differently. Keyword density optimisation could be outsourced to a tool or a writer who understood placement patterns. Evidence density optimisation requires actual research — finding statistics, identifying sources, reading studies. The investment is higher; the output is also harder for lower-quality competitors to replicate.
Implication 2: Content depth matters more than frequency. One substantive, well-evidenced 2,000-word piece on a topic — with statistics, definitions, comparisons, and source attribution — is likely to accumulate more AI citation signal than three keyword-optimised 600-word pieces on the same topic. The signals compound within a single piece of content, not across many thin ones.
Implication 3: Readability is no longer a trade-off. Traditional SEO sometimes forced a choice between readability and keyword placement. Evidence density optimisation has no such trade-off: readability at grade 8–10 is itself a high-confidence AI citation signal (Aggarwal et al. 2024). Writing clearly is writing for AI visibility — not a compromise away from it.
Measure your evidence density.
The Evidence Density Score quantifies all four evidence dimensions against research benchmarks. Paste any piece of content and see your score — plus a prioritised list of what to improve first, ranked by evidence confidence.