How to Structure Content So AI Actually Extracts It: 11 Patterns With Before and After Rewrites

AI systems do not read your page. They retrieve a fragment of it, usually a few hundred words, and decide whether that fragment answers a question on its own. Content that reads well top to bottom and falls apart when a single section is lifted out will lose to worse content that survives the lift. The Princeton GEO study (Aggarwal et al., KDD 2024) found that adding quotations, statistics, and citations each raised visibility by roughly 25 to 40 percent in generative engine responses, and every one of those three is a change in how a passage stands on its own.

TL;DR

• Retrieval operates on passages, not pages. Optimise the paragraph, then the section, then the page, in that order.
• The single test that catches most problems: can this paragraph be pasted into a stranger's document and still make sense? If not, it will not survive chunking.
• Eleven patterns below, each with a real before and after taken from security content.
• The Princeton GEO study is the one widely cited result with a published method. Most of the specific numbers circulating about ideal passage length and table citation rates trace back to analyses nobody has published. Treat them as folklore.
• The cheapest high-yield fix in most B2B security content is killing pronouns and naming the entity every time.

Why extraction is not reading

When a person reads your page, context accumulates. They saw the H1, they know what "the platform" refers to, they remember the vendor named three paragraphs up. When a retrieval system pulls a passage, none of that travels with it. The passage arrives in a context window alongside four competing passages from four competing sites, and the model picks whichever one most directly answers the question it was given.

That is the whole mechanism, and it produces one counterintuitive consequence: redundancy is a feature. Restating the subject in every section reads slightly clunky to a human and is the difference between being extractable and being noise. Editing that redundancy out, which is what a good copy editor does, is actively harmful.

The orphan-paragraph test

Take any paragraph from your page. Delete everything around it. Paste it into a blank document. Read it as a stranger.

If it contains an unresolved reference (this, it, the platform, the above approach, as mentioned) or assumes a definition established earlier, it fails. A failed paragraph does not get excluded from retrieval, which would at least be clean. It gets retrieved, evaluated as low-confidence, and passed over in favour of a competitor's paragraph that stands alone.

Run this on five paragraphs from your highest-value page before reading further. Most B2B security content fails on three of five, and the failures cluster in exactly the sections you most want cited: the comparison sections and the technical explanations.

The eleven patterns

#	Pattern	Fixes	Effort
1	Answer in the first sentence of every section	Passage relevance	Low
2	Headings that state the claim, not the topic	Retrieval matching	Low
3	No unresolved references	Standalone survival	Medium
4	One idea per paragraph	Chunk boundaries	Medium
5	Name entities, kill pronouns	Entity resolution	Low
6	Tables for multi-attribute comparisons	Extraction fidelity	Medium
7	Numbered lists for sequences	Procedural queries	Low
8	Fact-dense first 200 words	Page-level relevance	Medium
9	Inline attribution with source and year	Trust signals	Low
10	One-sentence definitions before first use	Definitional queries	Low
11	Semantic HTML with a real heading hierarchy	Chunk boundaries	Low

1. Answer in the first sentence of every section

Not the first sentence of the page. Every section. A retrieval system that pulls your fourth H2 has no access to the answer you gave under your first H2.

Before: "There are a number of considerations when evaluating whether passkeys are appropriate for enterprise deployment. Organisations should weigh several factors before proceeding."

After: "Passkeys are appropriate for enterprise deployment when your workforce is on managed devices with a supported OS and you have a fallback for shared workstations. They are a poor fit for kiosk environments and for contractors on unmanaged hardware."

The before paragraph contains zero information. It is a promise that information is coming. Retrieval systems do not wait for the payoff.

2. Headings that state the claim, not the topic

A heading is the highest-weight text in a chunk. Spending it on a category label wastes the strongest signal you have.

Before: "Pricing Considerations"
After: "Why per-MAU pricing gets expensive above 500,000 users"

Before: "Implementation"
After: "How to migrate from Auth0 without forcing a password reset"

The second form matches how people actually ask. It also constrains what you write under it, which is a secondary benefit worth having.

3. No unresolved references

Before: "This approach has the advantage of not requiring changes to the identity provider. However, it does introduce latency."

After: "Token exchange at the API gateway does not require changes to the identity provider. It adds roughly 20 to 40 milliseconds of latency per request, because the gateway makes an additional round trip to validate the token."

"This approach" is invisible to anything that did not read the preceding paragraph. Name it. The rewrite is longer and that is fine: a passage that is 30 percent longer and 100 percent self-contained wins every time.

4. One idea per paragraph

Chunking splits on structural boundaries. A 200-word paragraph covering three ideas either gets split mid-argument or retrieved whole and diluted, and both outcomes lose to a competitor's tight 60-word paragraph that covers one idea completely.

Target two to four sentences. When a paragraph needs a "meanwhile" or a "that said", it is two paragraphs.

5. Name entities, kill pronouns

This is the highest-yield change in most B2B security content and takes an afternoon.

Before: "The vendor supports SCIM. They also offer SAML and OIDC, and their pricing scales with usage."

After: "Okta supports SCIM 2.0 for user provisioning. Okta also supports SAML 2.0 and OIDC, and Okta's pricing scales with monthly active users."

Three uses of "Okta" in two sentences reads repetitively and is correct. Entity resolution is how a model connects your passage to a query about a named product. Pronouns break that link, and a broken link means the passage answers a question nobody asked.

6. Tables for multi-attribute comparisons

Any comparison across three or more attributes belongs in a table. Prose comparisons force the reader, human or machine, to reconstruct a grid from sentences, and reconstruction introduces errors.

Before: "Vault is open source and self-hosted with a steep operational burden, while Doppler is a hosted service that is easier to run but less flexible, and AWS Secrets Manager is fully managed but locks you to one cloud."

After:

Tool	Deployment	Operational burden	Cloud lock-in
HashiCorp Vault	Self-hosted	High	None
Doppler	Hosted SaaS	Low	Vendor
AWS Secrets Manager	Fully managed	Lowest	AWS only

A caution on the folklore here. Several widely shared guides claim tables are cited some specific multiple more often than prose. Those numbers trace back to analyses that were never published with a method. The structural argument for tables is sound on its own and does not need an invented multiplier.

7. Numbered lists for sequences, bullets for sets

Numbered lists signal order. Bulleted lists signal membership. Using bullets for a procedure tells a retrieval system that step three is interchangeable with step one, which is exactly wrong for a migration runbook or an incident response sequence.

8. Fact-dense first 200 words

The opening of a page is weighted heavily in page-level relevance. Spend it on facts, not framing.

Before: "Identity management has become increasingly important for modern organisations. As threats evolve, companies need to think carefully about how they approach access control."

After: "SCIM 2.0 is the provisioning standard that syncs user accounts between an identity provider and a downstream application. It was published as RFC 7644 in 2015, is supported by Okta, Entra ID, and Google Workspace, and handles create, update, and deactivate operations. It does not handle group nesting consistently across providers, which is where most implementations break."

9. Inline attribution with source and year

Write "according to Cloudflare's May 2026 bot traffic data" rather than "recent research shows". The named source and date travel with the passage when it is extracted; a hyperlink does not. This is the single pattern from the Princeton study with the clearest supporting evidence, and it costs nothing.

10. One-sentence definitions before first use

Define every acronym and category term in one clean sentence the first time it appears, in the form "X is a Y that does Z". Definitional queries are a large share of AI search volume, and a clean definition sentence is the passage that gets pulled to answer them.

Before: "Modern CIAM platforms handle this differently than legacy IAM."

After: "Customer identity and access management (CIAM) is identity infrastructure built for external users, customers and partners, rather than employees. CIAM platforms handle registration, consent, and progressive profiling, which workforce IAM tools generally do not."

11. Semantic HTML with a real heading hierarchy

Chunkers use heading tags to find boundaries. Styled divs that look like headings are invisible to them. One H1, H2s for sections, H3s nested under H2s, no level skipping, real <table> elements rather than CSS grids, real <ol> and <ul> rather than paragraphs starting with a dash.

Modern component libraries make this failure easy. If your design system renders headings as styled spans, you have an extraction problem no amount of writing fixes.

The seven-step audit

To apply this to an existing page rather than a new one:

1. Run the orphan-paragraph test on every section. Mark the failures.
2. Rewrite the first sentence of each section so it answers the section's heading.
3. Rewrite each heading as a claim or a question.
4. Search the page for "this", "it", "they", "the platform", "the above". Replace each with the named entity.
5. Find every prose passage comparing three or more things. Convert to a table.
6. Add source and year to every statistic that lacks one. Delete any statistic you cannot attribute.
7. View source. Confirm headings are real heading elements and tables are real tables.

Steps 4 and 6 produce most of the gain. Step 6 is where people stall, because converting prose to a table often reveals that the underlying comparison was never fully worked out.

Structure gets a passage evaluated. It does not get it retrieved in the first place, which is a question of crawler access covered in the AI crawler reference, and it does not tell an engine what your entity is, which is what schema markup handles. For where these fit in a full programme, see the AEO strategy playbook.

Frequently Asked Questions

How long should a paragraph be for AI extraction?

Two to four sentences, roughly 40 to 80 words, is a reasonable working target because it usually contains exactly one idea. Be sceptical of guides quoting precise optimal word counts with citation multipliers attached; those figures come from analyses that were never published with a method. The real constraint is one idea per paragraph, and length follows from that.

What is the orphan-paragraph test?

Take a paragraph, delete everything around it, and read it as a stranger. If it contains an unresolved reference such as "this approach" or "the platform", or relies on a definition given earlier on the page, it fails. Failed paragraphs still get retrieved, they just get evaluated as low-confidence and passed over for a competitor's self-contained passage.

Does repeating the product name hurt readability?

Slightly, and it is worth it. Entity resolution is how a retrieval system connects your passage to a query about a named product, and pronouns break that connection. Aim for naming the entity in the first sentence of every paragraph and in every heading, then use pronouns within a paragraph where the referent is unambiguous.

Do tables really get cited more than prose?

The structural argument is strong: a table preserves the relationship between attributes without requiring reconstruction, which reduces extraction error. The specific multipliers circulating in AI SEO content, such as tables being cited four times more often, trace back to unpublished analyses and should not be quoted as fact.

What did the Princeton GEO study actually find?

Aggarwal et al., published at KDD 2024, tested about 10,000 queries in a simulated pipeline where Google retrieved the top five sources and GPT-3.5-turbo synthesised a cited answer. Adding quotations, statistics, and citations each raised visibility by roughly 25 to 40 percent on their position-adjusted word count metric. The important caveat is that the source was already present in a fixed context, so the result measures how well you compete once retrieved, not whether you get retrieved.

Should I restructure old content or write new content?

Restructure first. An existing page already has crawl history, internal links, and whatever authority it accumulated. The seven-step audit takes about an hour per page and applies that hour to an asset that is already discoverable, which a new page is not.

Does this conflict with writing well for humans?

In two places only. Self-contained sections require repetition that a copy editor would cut, and explicit entity naming reads more mechanically than pronouns. Everything else on this list, direct answers, clear headings, one idea per paragraph, attributed claims, and defined terms, is ordinary good technical writing that happens to also be machine-legible.