Most advice about getting cited by AI engines focuses on what to write — formats, schema, entity consistency. This post is about where: the same fact, on the same page, has a different probability of being quoted depending on how far down it sits. We’ve covered how to structure content so AI agents can parse it; this is the companion question — once your page is parseable, which parts do engines actually lift?

The circulating claim — and who actually ran the sample

If you read AEO roundups, you’ve met the figure: roughly 55% of AI Overview citations come from the first 30% of the cited page — usually unattributed, or attributed to “a study”. Since we’ve written about reading AEO benchmark reports without fooling yourself, we traced this one to its owner before repeating it.

The owner is CXL, the marketing training company. In March 2026, CXL’s content lead Tarek Reslan published a structured analysis of 100 Google AI Overview citations, mapping where within each source page the cited snippet appeared relative to article length. The finding, in CXL’s own words: “55% of citations came from the first 30% of content, while only 21% came from the bottom 40%.” A further 24% came from the middle section (30–60%).

Here is CXL’s full positional breakdown, reproduced from the study:

Position on page (relative to article length)	Cited snippets (of 100)
0–10%	21
10–20%	27
20–30%	7
30–40%	9
40–60%	15
60–80%	11
80–100%	10

Two honest caveats before you build strategy on this. First, n=100 citations is a small sample from one engine (Google AI Overviews), and CXL publishes little about how the query set was chosen — treat the 55% as an observation, not a law. Second, the distribution isn’t a clean slide to zero: the 10–20% zone is the densest single segment, yet a fifth of citations still come from the bottom 40%. CXL’s explanation for those deep citations is worth more than the headline number: “A meaningful share of bottom-of-page citations in our analysis came from FAQ blocks addressing specific, discrete questions.” Self-contained question-and-answer units get cited from anywhere, because each behaves like a tiny page of its own.

Is there corroboration beyond one 100-citation sample? Directionally, yes — with a trace problem we should name. Search analyst Kevin Indig published a much larger analysis in his Growth Memo newsletter in February 2026, publicly described as: “I analyzed 1.2 million search results to find out exactly how AI reads. The verdict? It’s a busy editor, not a patient student.” Secondary reports of that study — CXL’s article among them, citing 44.2% of 18,012 verified ChatGPT citations landing in the first 30% of a document — put ChatGPT’s citation bias in the same territory as CXL’s Google numbers, but the full study sits behind Growth Memo’s paywall, so we could not verify those percentages at the owner’s page and won’t present them as established. Take the direction (top-heavy, two engines, two independent samples), not the decimals.

Why position matters: what the engine owners actually document

No AI engine publishes a statement like “we prefer the top of pages” — the positional bias is a third-party observation. What the owners do document is the mechanism that makes position matter: none of them treats a page as a single unit. They retrieve, score and rank pieces of pages.

Perplexity is the most explicit. Announcing its Search API in September 2025, Perplexity described its production infrastructure: “It is insufficient to operate simply at the document level. Our indexing and retrieval infrastructure divides documents up into fine-grained units. These sub-document units are individually surfaced and scored against the original query parameters.” Your page is not competing for citation — your paragraphs are, individually.

OpenAI documents chunking in its retrieval tooling. Its developer-platform retrieval guide — the machinery behind the file_search tool, not ChatGPT search itself, but the clearest public window into how OpenAI slices documents — states that by default every uploaded file “is indexed by being split up into 800-token chunks, with 400-token overlap between consecutive chunks”, and that a file_search query returns at most 10 results by default (configurable up to 50). Pipelines built this way never hand the model your whole article; they hand it a shortlist of fragments that scored well against the query.

Google says the least about position and the most about eligibility. Its “AI features and your website” documentation states: “There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary.” It also documents that “Both AI Overviews and AI Mode may use a ‘query fan-out’ technique — issuing multiple related searches across subtopics and data sources — to develop a response”, and that to be a supporting link “a page must be indexed and eligible to be shown in Google Search with a snippet”. Note the unit: a snippet. Google’s preview controls (nosnippet, data-nosnippet, max-snippet) likewise operate on passages, not pages — the same sub-document logic as everyone else.

Put the owner documentation together and the third-party position data stops being mysterious. If a system splits your page into scored fragments and forwards only the top scorers, a direct, self-contained answer wins — and an answer that only emerges across your final three sections was never a candidate, because no single fragment contains it.

Why front-loading works (and what it actually means)

Front-loading is the old journalism discipline of the inverted pyramid — conclusion first, support after, background last — applied at three levels of the page:

Page level: the primary answer to the query your page targets should appear inside the first 150–200 words, as a capsule that survives being quoted alone — not a teaser, the actual answer. CXL’s practical read of its own data: audit where your core answer sits, and “if it’s below the 30% mark, that’s your first optimization target.”

Section level: each H2 should open with its conclusion. A question-shaped heading followed by a direct answer is a retrieval-shaped unit; a heading followed by three paragraphs of wind-up is a wasted chunk. This compounds with the fan-out behaviour Google documents — sub-queries land on sub-topics, so every section is a landing zone — and with how reasoning-grade models fetch more sources and quote less from each, a pattern we unpacked in how ChatGPT’s thinking mode cites sources.

Block level: FAQ entries, definitions and single-fact tables are position-independent because they’re self-contained — the structural reason FAQ blocks punched above their position in CXL’s sample. You don’t have to cram everything into your introduction if your deep content is packaged as discrete answer units.

None of this means abandoning depth or narrative. It means re-ordering: lead with the finding, then earn it. Skimming humans benefit from exactly the same structure.

The restructure guide: before and after

Here’s what re-ordering an existing page actually looks like, element by element:

Element	Typical structure (answer-last)	Front-loaded structure (answer-first)
Opening 150–200 words	Anecdote, market context, “why this matters in 2026”	A capsule stating the page’s core answer, quotable on its own
Core definition	Emerges gradually across sections	Stated once, precisely (“X is…”), inside the first 20% of the page
Key finding or number	The payoff of the final section	In the capsule and repeated with support where it’s argued
Section openings	Build-up first, conclusion at the end of the section	Conclusion in the first sentence under each heading
Comparisons	Discursive prose spread across paragraphs	A table early in the relevant section, one claim per row
Deep-page content	Long dependent narrative that assumes everything above	Self-contained FAQ and definition blocks that survive extraction alone
Background and caveats	Interleaved throughout, delaying every answer	Grouped after the answers they qualify

A working sequence for an existing library: pull your ten most important pages and note where the sentence that most directly answers the target query sits. Anything below the 30% mark gets a rewritten opening — move the answer up, leave the supporting case where it is. Rewrite each H2’s first sentence to state that section’s conclusion, then convert your most-asked deep-page material into FAQ blocks with the answer in the first sentence. You’re not producing new content; you’re re-ordering what already ranks.

This is the same discipline we apply to outbound at Zian: our autonomous AI sales agents are built to lead with the point — who’s calling, why, and what’s on offer — because buyers, like retrieval systems, decide on the first fragment they process. If you want outreach across phone, SMS, email and WhatsApp built that way, Apply For Partnership.

FAQ: page position and AI citations

Where on a page do AI engines quote from?

Disproportionately from the top. The best-documented public sample, CXL’s March 2026 analysis of 100 Google AI Overview citations, found 55% of cited snippets came from the first 30% of the page, 24% from the 30–60% zone, and 21% from everything after the 60% mark — with the 10–20% zone the single densest segment. It’s one small sample from one engine, but larger paywalled work on ChatGPT citations points the same direction.

Is the “55% from the first 30%” statistic reliable?

It’s real and owned — CXL ran the sample and published the zone-by-zone data — but it’s 100 citations from one engine with a thinly described query set. Treat it as a directional finding that matches the documented retrieval mechanism, not a precise constant. Any roundup quoting it without naming CXL, the n of 100 or the engine hasn’t traced it.

Do AI engines read the whole page?

They fetch the whole page, but they don’t weigh it as one unit. Perplexity states its “indexing and retrieval infrastructure divides documents up into fine-grained units” that are “individually surfaced and scored”; OpenAI’s documented retrieval tooling splits files into 800-token chunks and returns at most 10 results per query by default. Selection happens at fragment level, which is why position and self-containment matter.

Does Google say you should front-load content for AI Overviews?

No. Google’s official position in its “AI features and your website” documentation is that “There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary.” The front-loading case comes from third-party citation measurements plus the snippet-level eligibility Google does document — not from any Google instruction.

Does front-loading mean writing shorter pages?

No — it means answering earlier on pages of whatever length the topic needs. Long pages remain citable when structured as a stack of self-contained units: an answer capsule up top, conclusion-first sections, and FAQ blocks deep down. CXL’s data showed deep-page FAQ blocks getting cited despite their position, because each question-answer pair works as a standalone answer.

How is this different from optimising content structure generally?

Structure (headings, tables, schema, server-rendered HTML) determines whether an engine can parse your page at all — we cover that in writing for agentic parsing. Position determines which parts of a parseable page get quoted: a beautifully structured page with its answer at the 70% mark is still unlikely to be cited for its main claim. Both sit inside a broader answer engine optimisation programme.

Source-ownership table

Every external figure and quote in this post, traced to the organisation that owns it:

Figure / claim	Owner	Verified at
55% of 100 sampled AI Overview citations from first 30% of page; 24% from 30–60%; 21% after 60%; zone table; FAQ-block finding	CXL (author Tarek Reslan, published 6 March 2026)	cxl.com/blog/google-ai-overview-citation-sources/
“1.2 million search results… busy editor, not a patient student” (study description, verified at owner; the 44.2% / 18,012-citation positional figure is reported secondhand by CXL, paywalled at the owner, and stated in this post as unverified)	Kevin Indig, Growth Memo (16 February 2026)	growth-memo.com/p/the-science-of-how-ai-pays-attention
“Query fan-out” quote; “no additional requirements” quote; supporting-link snippet eligibility	Google (Search Central documentation)	developers.google.com/search/docs/appearance/ai-features
file_search chunking defaults: “split up into 800-token chunks, with 400-token overlap”; max 10 results per query by default (up to 50)	OpenAI (developer documentation, retrieval guide)	developers.openai.com/api/docs/guides/retrieval
“Divides documents up into fine-grained units… individually surfaced and scored” quote	Perplexity (Search API announcement, 25 September 2025)	perplexity.ai/hub/blog/introducing-the-perplexity-search-api

Front-loading is a re-ordering job, not a rewriting job — the rare optimisation that serves human skimmers and retrieval systems identically. If you’d rather spend your team’s hours on conversations than restructuring, Zian’s autonomous AI sales agents run prospecting, qualification and booking end to end. Apply For Partnership to join the waitlist.