30-second rundown
- DO put your single most important insight in the first paragraph, not paragraph eight. If a reader — human or machine — stops after your intro, they should already have the answer they were looking for.
- DO rewrite your H2 headings as literal questions your reader would type into a search bar, then open the paragraph below with a direct answer that echoes the heading’s keyword in the first word.
- DON’T write sentences like “there are many powerful solutions available.” Name Salesforce. Name HubSpot. Name Pipedrive. Generic prose gets zero citations. Namedropping is survival.
Three takeaways from Kevin Indig’s ‘The science of how AI pays attention’ article.
Your AI overlord has the attention span of a deadline-drunk editor, but you’re still writing like Tolstoy.
Kevin Indig analyzed 1.2 million ChatGPT answers and isolated 18,012 citations to find out exactly where and why the machine pays attention.
What he found should terrify every SEO content drone still crafting “ultimate guides” with a slow-burn intro and a punchline buried on page four.
The machine is not your friend. It is a busy editor with a deadline and a hair-trigger delete finger.
1. The Ski Ramp Will Bury Your Best Ideas Alive.
The data is statistically indisputable, P-Value of zero, which in science means God himself couldn’t argue with it.
A full 44.2% of all ChatGPT citations come from the first part of your content, 31.1% from the middle, and 24.7% from the conclusion.
The AI does wake up at the end, but only for a summary or conclusion section — it skips the actual footer like a drunk avoiding his bar tab.
The model is trained on journalism and academic papers, both of which follow BLUF doctrine: Bottom Line Up Front, always, or you are dead to it.
But here is where it gets strange and interesting.
Zoom in to the paragraph level and the top-heavy bias dissolves.
A full 53% of citations come from the middle sentence of a paragraph, not the first. Only 24.5% come from the opening sentence.
The AI is not lazy. It reads deeply within a paragraph, hunting for the sentence with the highest information gain, the one that packs in the most relevant entities and expansive facts.
The implication: your best shot at a citation lives in the paragraphs of the first 20% of the page, but you do not need to torture that first sentence to carry the whole load.
2. Structure Your Page Like a Wire Report, Not a Novel.
There are five characteristics of cited content, and the first two are about architecture.
Citation winners are nearly 2x more likely to contain definitive language — phrases like “is defined as” or “refers to” — because in a vector database, “X is Y” is the strongest possible bridge between a question and an answer.
The model wants to resolve a query in a single sentence. Your slow atmospheric wind-up is not atmosphere. It is friction.
The second structural weapon is the question-and-answer heading.
Cited text is 2x more likely to contain a question mark, and 78.4% of those questions come from H2 headings.
The AI is treating your H2 as the user prompt and the paragraph below it as the generated response.
Rename your heading from “The History of SEO” to “When did SEO start?” and the paragraph that follows it becomes, in the machine’s cold little mind, an answer worth quoting.
3. Pack In Names, Balance Your Tone, and Write Like the Economist.
Normal English text runs at 5-8% entity density. Heavily cited text runs at 20.6%.
That is the difference between “there are many good tools for this task” and “top tools include Salesforce, HubSpot, and Pipedrive.”
The model prioritizes sentences with named anchors because they lower its perplexity — its confusion — and make the answer verifiable.
Name your competitors. Name the tools. Name the people.
The data says namedropping is not tacky. It is survival.
On tone, the cited sweet spot is a subjectivity score of 0.47 — the “analyst voice,” halfway between dry Wikipedia and unhinged opinion.
State the fact, then tell the reader what it means.
And on complexity, winners score Flesch-Kincaid 16, college level. Losers score 19.1, PhD level.
The AI prefers clean subject-verb-object structures it can extract facts from, not labyrinthine jargon sentences that make everyone feel stupid.
Write like the Economist. Not like a dissertation.
Like someone who respects the reader’s time and their own intelligence equally.
The old game was keeping humans on the page. The new game is making a machine trust you enough to quote you.
These are not the same game. They are not even the same sport.
But here is the consolation prize: front-loaded, clear, entity-rich content satisfies the algorithm’s architecture and the time-starved human reader who scans for insights.
You are not writing for robots instead of people.
You are finally being forced to write well for both at the same time.
