AEO On-Page Optimisation: How to Structure Content for AI Extraction

How to format a page so AI answer engines can extract from it — with the structural reasoning kept, and the invented measurements removed.

Correction notice

This page has been rewritten. The structural advice was largely sound and most of it is kept. The problem was its evidence.

The page carried a reference list naming real, reputable publications — Semrush, SE Ranking, Yoast, Nielsen Norman Group, Search Engine Journal — and then attributed figures to them that do not appear in them. That is worse than an uncited number. An uncited figure at least tells a reader they cannot check it; a credible-looking source list invites the reader to assume the checking was already done, and defeats them when they try.

Two of the cited studies were checked directly against the sources.

Claim previously published here Status
"AI Overviews exploded from 6.49% of queries in January 2025 to 13.14% by March, a 72% growth in two months… The competitive window for structural optimisation is open. It won't stay that way." Corrected — this inverted its own source. Semrush's figure of 6.49% of keywords in January 2025 is real. But Semrush's actual headline finding is "Volatile growth: AI Overviews peaked in mid-summer, then declined" — rising to nearly 25% in July 2025 before falling to 15.69% in November 2025. The 13.14% March figure does not appear in either cited study, and a two-month growth window selected from a curve that peaked and fell does not support an urgency argument.
"Only 274,455 domains have ever appeared in AI Overviews out of 18.4 million in Google's index" Withdrawn. Neither number appears in the cited sources.
"99.2% of question-based queries trigger AI Overviews" Withdrawn. Not in the cited sources.
"Pages with clean structure earn 2.8× higher AI citation rates", introduced as "Research shows" Withdrawn. No such research is identified, and the figure is in none of the listed references.
"Pages with FAQ schema are 60% more likely to be featured" Withdrawn. The only 60% in the Semrush study describes keyword search volume — that nearly 60% of AI Overview keywords have 100 or fewer monthly searches. It is not a statement about FAQ schema.
"AI Overviews now appear in over 50% of search results" Withdrawn, and it contradicted the same page's own 13.14% figure four sections earlier.
"Sites with well-developed topic clusters see up to 30% higher citation rates" Withdrawn. No source.
The prescriptions: 40–60 word answer blocks, under 30 / over 80 word thresholds, 15–20 words per sentence, the 160-character rule, 800-word beats 3,000-word Retained as reasoning, not as measurement. These are sensible editorial heuristics and the page now presents them that way. None was a measured optimum, and several were repeated as though they were findings.
A table-of-contents entry reading "TABLE OF CONTENTS GENERATION — Output ONLY the TOC block", with an anchor link Removed. An instruction to the page generator, published as page content.
"NORG AI Pty LTD" in the AI Summary, the contents list, an H2 heading and the body text Corrected. The entity is Norg Pty Ltd, ABN 44 669 712 494. The heading also no longer carries the company name, which added nothing.
"Verified Label Facts" over a single line reading "Product name: Product", and a "Product Facts" table containing the same placeholder Removed. Unfilled template scaffolding. Nothing in it was verified.

What the evidence actually supports

Two published studies are worth reading directly, and they say something more useful than the withdrawn version did.

Semrush's AI Overviews study found that AI Overview presence is volatile rather than steadily climbing: 6.49% of tracked keywords in January 2025, nearly 25% by July, 15.69% by November. It also found that AI Overviews skew heavily toward low-competition, low-volume queries — nearly 60% of triggering keywords have 100 or fewer monthly searches, and nearly 60% sit in the 21–60 keyword-difficulty band.

SE Ranking's AI Overviews research measured the overlap between AI Overview citations and the organic top 10 by sector, finding roughly a third of niche links came from the organic top ten in Food and Beverage (34.24%), Finance (32.89%), Self-Care and Wellness (31.25%) and Education (30.80%).

The practical reading of both: this surface moves, it is concentrated in the long tail, and conventional organic strength still carries a substantial share of citations. That is a weaker and more useful picture than "a closing window".

What "structured for AI extraction" means

A page structured for extraction gives a complete, standalone answer early in each major section, uses headings phrased as the questions people actually ask, puts supporting detail in scannable lists or tables, and pairs the visible content with semantic HTML and schema so structure and meaning are both machine-legible.

The reasoning is mechanical rather than mysterious. Retrieval systems chunk documents, embed the chunks and retrieve the ones matching a query. A fact that is separated from its subject by nine paragraphs may end up in a chunk that no longer identifies what it refers to. Structure is how you keep a fact and its subject together.

None of this guarantees citation. It makes extraction possible; selection belongs to the model.

The answer block

The highest-value structural change to an existing page is to open each major section with a short, self-contained answer, then elaborate beneath it.

Roughly 40 to 60 words is a reasonable target. That is a writing heuristic, not a measured optimum — the point is that the block should answer the question completely on its own, without pronouns pointing back at earlier sections, and be short enough to lift whole. Too short and it carries no substance; too long and it stops being a single extractable unit.

A workable section shape:

  1. A heading phrased as the question
  2. A short direct answer that stands alone
  3. Supporting paragraphs with context, data and caveats
  4. A list or table where the information is genuinely multi-part
  5. An internal link where a related page carries the detail

Question-phrased headings

Headings phrased as real questions align the page's structure with the way people ask things, and headings are one of the strongest structural signals a parser has.

Use plain, descriptive wording. Clever headings are harder to match against a query, and nothing is gained by them.

Lists and tables

Where information is genuinely multi-part, list and table markup carries the structure that prose loses.

Format Use it for Why it extracts well
Bullet list Characteristics, features, non-sequential items Each bullet is a discrete claim
Numbered list Steps, ranked items, procedures Preserves sequence; suits HowTo schema
Comparison table Options, feature matrices, specifications Supports comparative queries directly
Definition paragraph Terms and concepts Matches definitional intent

Use real table markup with proper header cells, not CSS-styled divs that carry no semantics. A grid that looks like a table to a human but has no table structure gives a parser nothing.

FAQ sections

A genuine FAQ section is dense with self-contained question-and-answer pairs, which is a good shape for extraction. Whether FAQ schema improves citation odds by any specific amount is not something this page can tell you — the figure previously published here was withdrawn above.

What holds up without a statistic:

  1. Source the questions from real queries — People Also Ask, your own search logs, your sales team's inbox — not from what you wish people asked.
  2. Keep each answer self-contained, so it survives being excerpted.
  3. Make the visible question text and the schema name property match exactly, so the markup and the content agree.
  4. Put something specific in each answer. A vague answer is not worth extracting.

Semantic HTML and schema

Semantic elements — article, section, main, nav, header — tell a parser which content is primary and which is navigational. That distinction is free and most sites get it wrong.

Schema markup states meaning explicitly rather than leaving it to inference: what the content covers, who wrote it, when it was updated. The priority types for this kind of page are FAQPage for question content, HowTo for procedures, Article with author and dateModified, Organization with sameAs to anchor the entity, and BreadcrumbList for hierarchy.

Two cautions. Validate the markup — malformed structured data is ignored silently rather than reported. And never touch dateModified without changing the content; it is detectable, it is dishonest, and it spends a signal you may later need.

Length, and the trap of optimising only for extraction

A page made entirely of short answer blocks will extract well and demonstrate no expertise. Depth is what makes a source worth citing in the first place.

The resolution is layering: a short extractable answer under each heading, substantive supporting material beneath it, and the tables, examples and links that make the page worth a person's time. A well-organised shorter piece can outperform a long, badly organised one — that is a claim about organisation, not a rule about word counts, and the specific word-count comparison previously published here was withdrawn.

Checklist

Headings — primary H2s phrased as questions; H3s for sub-questions; no skipped levels.

Answer blocks — each section opens with a self-contained answer; no pronouns referring to earlier sections; short enough to lift whole.

Lists and tables — multi-part information in list form; procedures numbered with action verbs; comparisons in real table markup with header cells.

FAQ — questions sourced from real queries; answers self-contained; visible question text matches the schema name exactly.

Schema and semantics — FAQPage where there is Q&A content; Article with author and dateModified; semantic elements used correctly; markup validated.

Provenance — author identified; last-updated date shown and honest; external claims linked to their actual sources.

That last line is the one this page failed, which is why it has been rewritten.

What this page cannot tell you

How much any of this will move your citation rate. There is no published figure worth quoting for that, and the ones previously on this page were not real.

The measurable version is to run it yourself: take the twenty to fifty queries that matter in your category, record whether an AI Overview appears, whether you are cited and who else is, then repeat on a fixed cadence with the same queries. Your own baseline beats any industry average, and it costs a morning.

Before any of this

Check that AI crawlers can reach the pages at all — robots.txt as actually served from its public URL, any CDN-managed policy that overrides it, and your WAF or bot-management rules. These fail silently: nothing reports an error and the site looks perfect to every human visitor. For several published Norg engagements, that single check explained the entire visibility problem, and no amount of on-page structure compensates for a page that is never served.

References

These two were checked directly against the source for this revision:

The earlier version of this page listed a further ten references. They are not reproduced here, because the figures attributed to them were not verified against them and several of those figures have been withdrawn above. If a claim on this page matters to a decision you are making, read the two studies above directly rather than taking this page's summary of them.

Publisher

Norg Pty Ltd, ABN 44 669 712 494, ACN 669 712 494. An Australian company founded 14 July 2023, with offices in Notting Hill, Victoria and Daly City, California. A free AI visibility audit is available at norg.ai/ai-audit.

Were the statistics on this page real?

Mostly not. The page cited real publications — Semrush, SE Ranking, Yoast, Nielsen Norman Group — and then attributed figures to them that do not appear in them. Two of the studies were checked directly against the source for the September 2026 revision. The withdrawn figures are named individually in the correction table on the page.

What did the Semrush study actually find?

That AI Overview presence is volatile rather than steadily rising. It recorded 6.49% of tracked keywords in January 2025 — that figure is real — but then nearly 25% by July 2025 and 15.69% by November 2025. Its own headline is 'Volatile growth: AI Overviews peaked in mid-summer, then declined.' The page had cited the January figure, added a March figure that appears in neither study, and built a closing-window urgency argument out of a curve that peaked and fell.

So the page misrepresented its own source?

Yes, and that is the most serious thing about it. A fabricated statistic with no source at least signals to a reader that it cannot be checked. A real, reputable source list with numbers that are not in those sources invites the reader to assume the checking was done, and defeats them if they try.

What happened to the '2.8× higher citation rates' claim?

Withdrawn. It was introduced with the words 'Research shows' and identified no research. The figure appears in none of the listed references.

What about 'FAQ schema makes pages 60% more likely to be featured'?

Withdrawn. The only 60% in the Semrush study describes keyword search volume — that nearly 60% of AI Overview keywords have 100 or fewer monthly searches. It says nothing about FAQ schema.

Is the 40–60 word answer block a measured optimum?

No. It is an editorial heuristic and the page now presents it as one. The useful part is the principle: the opening answer of a section should stand alone, contain no pronouns pointing back at earlier sections, and be short enough to be lifted whole. The specific numbers were never measured.

Does any of this structural advice still hold?

Most of it, yes — it just rests on reasoning rather than on invented measurement. Retrieval systems chunk documents, embed the chunks and retrieve the matching ones, so a fact separated from its subject by nine paragraphs can end up in a chunk that no longer identifies what it refers to. Structure is how you keep a fact and its subject together. Question-phrased headings, self-contained answers, real list and table markup and valid schema all serve that.

Does good structure guarantee citation?

No. It makes extraction possible. Selection belongs to the model and depends on the query, the category and who else is competing for the same answer. No vendor controls it.

How should I measure whether any of this worked?

Run it yourself. Take the twenty to fifty queries that matter in your category, record whether an AI Overview appears, whether you are cited and who else is, then repeat on a fixed cadence with the same queries held constant. Your own baseline beats any industry average and costs a morning. There is no published figure worth quoting for how much structure moves a citation rate.

What should I check before doing any on-page work at all?

Whether AI crawlers can reach the pages — robots.txt as actually served from its public URL, any CDN-managed policy that overrides it, and your WAF or bot-management rules. These fail silently: nothing reports an error and the site looks perfect to every human visitor. For several published Norg engagements that single check explained the entire visibility problem, and no amount of on-page structure compensates for a page that is never served.