Content structure: the part of the page AI engines actually quote

The 8,000-citation study: deep pages, not homepages

James Allen pulled almost 8,000 unique citations across 57 queries and four engines, ChatGPT, Google Gemini, Perplexity and Google AI Overviews, using tracking data from Rankscale.ai, and published the analysis in Search Engine Land on 12 May 2025. The headline structural finding: 82.5% of citations pointed to a deeply nested, specific page rather than a homepage or a top-level category page. Query intent shifted which kind of page won that slot. B2B queries favoured industry-specific sources and blogs, with official company blogs and sites accounting for roughly 17% of B2B citations, while B2C queries leaned on consumer review sites and mainstream media. Across both, the pattern held: a page built to answer one question outcompeted a page trying to represent an entire company or category.

That is the same instinct behind our guide to getting cited by ChatGPT, which argues for owning one specific answer per page rather than one sprawling guide gesturing at several. This study is the evidence for why: a homepage is competing with every other page on the internet that could plausibly answer the question, while a page written for the exact question has no such competition on its own site.

Where inside the page the quote comes from

Ryan Law at Ahrefs published a piece on structuring content for AI visibility on 8 May 2026, modified 22 August 2026, that draws on research by Kevin Indig into how AI models weight different parts of a page. Three figures from that research, as Ahrefs quotes them: 44.2% of citations come from the first 30% of an article's content, meaning the answer needs to sit near the top rather than arrive after several paragraphs of context-setting. Heavily cited passages carry an entity density of around 20.6%, meaning names, numbers, dates and specific things mentioned per sentence, against 5 to 8% in standard English prose. And sentences written in definitive language, stating a claim rather than qualifying it, get cited at 36.2% against 20.2% for sentences that hedge.

We could not fetch Kevin Indig's original piece directly to check the full methodology behind those numbers ourselves, so what is reported here is Ahrefs' citation of his research rather than an independent read of his own write-up, a gap we flag properly below. What the figures argue for is specific and testable regardless of the exact methodology: put the direct answer early, write with concrete entities rather than vague description, and state what you know plainly instead of qualifying every sentence. That last point lines up with a habit we already hold ourselves to. A guide full of "it's worth noting that" and "some experts suggest" is exactly the hedge pattern the data says gets cited less.

Retrieval is not citation

The structural findings above describe what makes a page a good candidate. A separate Ahrefs study, by Louise Linehan and Xibeijia Guan, analysed 1.4 million ChatGPT prompts from February 2025 and published the results on 15 April 2026, modified 27 August 2026, and it describes the step before that: getting pulled into the pool ChatGPT considers at all. ChatGPT retrieves dozens of URLs to answer a single query and cites roughly half of what it retrieves. Which channel a page comes through matters enormously to whether it survives that cut. URLs surfaced through a plain search channel got cited 88.46% of the time, 12.01% for news, and just 1.93% for Reddit, even though Reddit supplied 16.2 million of the retrieved URLs in the dataset, more than any other source, and 67.8% of everything retrieved but never cited was from Reddit. The authors describe the pattern as the model reading from the crowd and then citing a different, more citable source for the answer it gives.

That is a different measurement from the one in our piece on whether Reddit still dominates AI search citations, which used Semrush's cross-platform citation share data rather than ChatGPT's internal retrieval funnel, and the two are not directly comparable. What both agree on is the underlying point this section makes: appearing in an engine's source material and being the page it quotes are two different events, and a content strategy built around getting retrieved without addressing why so few retrieved pages get cited is optimizing for the wrong number.

The same study found two smaller structural details worth acting on. Titles that scored higher on semantic similarity to the original prompt, 0.602 for cited pages against 0.484 for non-cited ones, were more likely to be the page quoted, which argues for a title that states the actual question rather than a clever one that implies it. And pages with natural-language URL slugs, words rather than IDs or truncated fragments, were cited 89.78% of the time against 81.11% for pages without one, a small but consistent gap that costs nothing to fix.

A structure checklist that follows from all three studies

1. Put the direct answer in the first third of the page. Not the first sentence necessarily, but well before the halfway mark. 44.2% of citations come from that zone, and a reader or a model that has to dig for the answer is a reader or model likely to find it somewhere else.
2. Write with specific entities, not vague description. Name the actual company, tool, number and date rather than writing around it. Cited passages run at roughly three times the entity density of ordinary prose.
3. State what the evidence actually supports, plainly. Definitive language outperforms hedged language by close to two to one. Where the evidence is genuinely uncertain, say that directly instead of hedging every sentence as a reflex.
4. Build one page per specific question, not one hub for a whole topic. 82.5% of citations in the Search Engine Land dataset went to deep, specific pages. A page trying to be the answer to everything is rarely the answer to any one thing an engine can lift cleanly.
5. Use a descriptive URL slug and a title that states the question. Both showed a measurable, if smaller, citation-rate advantage in the Ahrefs retrieval data.
6. Do not mistake retrieval volume for citation odds. A channel that gets pulled into an engine's context window a lot, the way Reddit does, is not automatically a channel that gets quoted. Design the page to convert once it is in front of the model, not just to get there.

What we could not verify

Kevin Indig's original research on where citations fall within a page and on entity density and definitive language reached us only through Ahrefs' citation of it. Two attempts to fetch the source directly in this session returned a provenance check we could not clear unattended, so the 44.2%, 20.6% and 36.2% figures above are relayed as Ahrefs quotes his work, not independently confirmed against his own methodology section, sample size or the exact definition he used for a citation. Ahrefs' own 1.4-million-prompt study is more directly checked, but its "semantic relevance" scoring is computed from cosine similarity between embeddings generated by open-source models, an approximation the study's own authors describe as a proxy, not a disclosure of how ChatGPT itself selects sources. Search Engine Land's 8,000-citation figure comes from Rankscale.ai's own tracking panel across a fixed set of 57 queries, a specific and limited sample rather than a random cross-section of AI search traffic, and Rankscale sells AEO tracking tools, which is a reason to want independent replication rather than a reason to discount the finding. No AI company, OpenAI, Google, Anthropic or Perplexity, publishes its own account of how source selection actually works, so every number in this piece is an external, reverse-engineered observation rather than a disclosed mechanism.

Sources

The deep-page citation figure and the B2B and B2C source-type breakdown come from James Allen, How to get cited by AI: SEO insights from 8,000 AI citations, Search Engine Land, 12 May 2025, using tracking data from Rankscale.ai across 57 queries and four engines. The first-30%, entity density and definitive language figures are Kevin Indig's research on the science of how AI pays attention, as quoted in Ryan Law, On-page AEO: four writing frameworks for better AI visibility, Ahrefs, 8 May 2026, modified 22 August 2026. The retrieval-versus-citation figures, the channel conversion rates, the title similarity scores and the URL slug figures come from Louise Linehan and Xibeijia Guan, Why ChatGPT cites one page over another, Ahrefs, 15 April 2026, modified 27 August 2026, a study of 1.4 million ChatGPT prompts from February 2025.

Common questions