Most advice treats AI visibility as one goal. We kept seeing this framing in the guidance — and the closer we looked at the research, the less accurate it seemed.
AI systems do two different things with your content. The first is citation selection — an AI includes your page in the list of sources it references. That's reach. The second is answer absorption — an AI actually draws from your content to shape the words and claims in its generated answer.
Most people optimising for AI visibility are targeting the first thing. They should be targeting the second. Here's what the data shows.
A preprint study by Zhang et al. (2026) analysed 21,143 citations across ChatGPT, Google AI Overviews, and Perplexity. One finding stands out. The finding that changes the optimisation conversation: high-influence pages — those visibly absorbed into AI answers, not just listed — shared a specific set of structural properties. Citation selection and answer absorption are not the same target. They require different content decisions.
What High-Absorption Pages Actually Look Like
The Zhang et al. data describes a gap, not a gradient.
High-influence pages were on average 11.44 times longer than low-influence pages. They had 12.5 times more headings. They had 5.69 times more paragraphs. These are not marginal differences between good and average content — they describe structurally different documents. When we built the Absorption Analyser, these were the properties we operationalised.
The content properties matter too. Pages with high definitional density showed approximately 57% higher absorption. Comparative content — sentences that draw direct comparisons between concepts, tools, or approaches — was associated with approximately 55% higher absorption. Statistics presence correlated with approximately 61% higher absorption in the same study.
One finding cuts against conventional optimisation advice: Q&A and FAQ-formatted sections showed approximately 5.74% lower relative absorption compared to narrative structure for ChatGPT. That matters. That is a directional finding from a preprint — it is current best-available evidence, not a settled rule — but it is a finding worth knowing before you build your next FAQ section.
Why This Matters More Than Citation Selection
Getting cited is a reach metric. Getting absorbed is a depth metric.
When an AI lists your page as a source, readers may or may not follow the link. That's a visit. When an AI draws from your content to construct its answer, your phrasing, your definitions, your comparisons shape what the reader sees — even if they never visit your site.
That distinction is The Citation Gap. Content that is structurally rich, evidentially dense, and written with definitional clarity is content that AI systems pull from. Not the same as being listed. Content optimised only for keyword presence may still be cited — but not drawn from. The Citation Gap is the distance between those two outcomes.
Platform Variance Is the Part Most Advice Gets Wrong
The picture is not uniform across platforms.
For Perplexity specifically, Aggarwal et al. (2024) — a peer-reviewed study published at KDD 2024 — provides the most evidentially robust data in the field. The numbers are direct. That study found quotation addition associated with approximately +22% citation probability, statistics addition with approximately +37%, and source citation with approximately +30%. These figures come from a real-world test on the live Perplexity platform. They are the most reliable numbers available.
For ChatGPT and Google AI Overviews, the Zhang et al. preprint provides directional guidance — structural richness (headings, paragraphs, word count) and content richness (definitions, comparisons, statistics) are the consistent positive factors. The average describes nothing. Optimising for "AI" as a single unified system is optimising for an average that may not match any actual platform. The evidence diverges. The platforms have different evidence bases, different study designs measuring them, and different findings about what works.
Measuring Absorption Properties Before You Publish
The structural properties Zhang et al. identified are measurable. Word count, heading count, paragraph density, definition sentence frequency, comparative sentence frequency, statistics presence — none of these require guesswork.
We built the AI Answer Absorption Analyser at psytable.com to operationalise exactly those properties. Paste your content. The tool scores it against six structural absorption dimensions drawn from Zhang et al. (2026) and returns a breakdown showing where your content is strong and where it falls short of the structural profile associated with high-influence pages.
It is free. No account. No email required.
The research gives you the target. The tool tells you how close your draft is to it.
Try the AI Answer Absorption Analyser at psytable.com.
Paste your content and get a breakdown of all six structural absorption dimensions drawn from Zhang et al. (2026). Free. No account required.
Research citations: Zhang et al. (2026), 'From Citation Selection to Citation Absorption,' preprint — not yet peer-reviewed; treat findings as directional. Aggarwal et al. (2024), 'GEO: Generative Engine Optimization,' KDD 2024 — peer-reviewed.