Research dispatches on AI citation behaviour and content evidence — plus a learning series that turns the same published studies into practical skills.
AI systems cite content that contains verifiable claims. Here's what statistical density is, why it matters, and how to measure yours in under five minutes — the clearest entry point into the evidence-first approach.
AI can cite your content without absorbing it into the answer. The Zhang et al. 2026 study shows why. Learn the two-stage mechanism and test your own content.
Keyword density and evidence density are different signals. AI systems respond to the second. Here's the distinction, why it matters, and how to measure your own content in under five minutes.
A peer-reviewed ACM study found that prepending a newer date to identical content reversed AI ranking preference by up to 25%. Here's what the recency signal is, why it matters, and how to audit your own content for it.
A practitioner dataset of 23,387 AI citations found that content type is sharply unequal in citation share. Before optimising the quality of what you produce, ask whether the format you're producing is competitive in your target context.
Only 6.82% of ChatGPT's citations overlap with Google's top 10. The two systems are not re-ordering the same competition — they are drawing from largely separate source pools. Here's what gets you into the one AI draws from.
A May 2026 preprint tested 14 models across search depths from 2 to 150 tool calls. Factual accuracy dropped an average of 42 percentage points at maximum depth. One model went from 79% accuracy at 2 sources to 17% at 150.
A May 2026 preprint measured 761,495 citation pairs across ten models and five providers. It found that 30.6% of citations in commercial search-augmented LLMs structurally distort what their sources actually say.
A 2026 preprint finds that adding statistics and quotations — the core advice in the original GEO research — reduces citation visibility by 14–19% when applied to already-fluent LLM-generated content. The same tactics improve visibility on human-written pages.
A March 2026 preprint found that standard GEO rewriting methods apply uniform changes to all documents regardless of context — and for long-tail, specialist content, that uniformity can actively lower citation rates.
Zhang et al. 2026 identified a measurement gap at the core of GEO research — the difference between citation selection (whether your content is chosen) and citation absorption (how much your content shapes the answer). Most tools only measure the first.
A March 2026 preprint tested structural changes alone — no semantic content change — and measured a mean 17.3% improvement in citation rates across six LLM engines. Here is what Yu et al. measured and what it means for your content structure decisions.
Two independent preprint studies — one measuring absorption depth, one measuring source-set stability — converge on the same platform characterisation. ChatGPT cites fewer sources more deeply. Perplexity cites more sources more shallowly. The difference is structural.
Three preprints from early 2026 changed the question. The field started by asking "what works?" and ended up asking "what works, for what content type, serving which audience?" The structural tier, substrate conditionality, and audience conditionality each depend on the others.
Six research papers validated, challenged, and reshaped Psytable's three GEO diagnostic tools. Two confirmed what we built. Four complicated it. Three gaps remain — and we don't have tools for them yet.
Only 6.82% of the pages ChatGPT cites overlap with Google's top-10 results for the same queries. Chalkidis et al. 2024 measured source selection across 55,936 queries — and found that LLMs and traditional search engines are drawing from largely separate pools.
A practitioner dataset of 23,387 citations shows thought leadership takes only 5.4% of branded-query citations — while reviews and listicles take 57%. Here's why the Layer 1 question (content type) comes before the Layer 2 question (structural quality).
A peer-reviewed ACM study found that prepending a newer date to identical content reversed AI preference by up to 25% — across seven models. Here's what Fang et al. 2025 means for your update-versus-create decisions.
Most advice treats AI visibility as one goal. The research shows two different targets: citation selection and answer absorption. Here's what the data shows about which one actually matters.
A close reading of Aggarwal et al. (2024) — the only peer-reviewed study in our tool set. What they measured, how they measured it, and what the +31% figure actually means for content you're writing today.
Zhang et al. 2026 analysed 21,143 citations across ChatGPT, Google AI Overviews, and Perplexity. Here's what they found about the structural properties of high-influence pages — and what the preprint status means for how you should use it.
Traditional SEO optimised for keyword density. AI citation optimisation requires something fundamentally different — evidence density. Here's why the two diverge, and what the research says about the signals that actually matter to AI systems.
The Evidence Density Score is the highest-confidence starting point. Peer-reviewed source. No signup.