What is an extractable block?

An extractable block is a passage of roughly 40 to 60 words that opens with a direct answer, names its own subject, carries at least one specific claim, and makes complete sense read in isolation from the rest of the page. Engines retrieve pages as chunks and score each chunk on its own, so a block that needs its surroundings to mean anything will not be cited.

How should B2B headings be written to improve AI citation?

Phrase every H2 and H3 as the question a buyer would type into an AI engine, and make the first sentence underneath resolve that question directly. Label headings such as "Features" or "Our Approach" match no query pattern, so the sections under them are hard for an engine to pair with a user prompt.

Does FAQPage schema still matter now that Google has removed FAQ rich results?

The markup is still worth adding. Google stopped showing the FAQ rich result on May 7, 2026 and withdrew the documentation the following month, but that removed a search feature, not the schema type. FAQPage remains valid Schema.org, crawlers that read structured data still parse it, and the visible question and answer pairs underneath are what engines lift.

What is the fastest way to retrofit an existing page for AI citation?

Convert label headings into question headings, then promote the answer sentence to the top of each section, then replace vague quantifiers with specific numbers, and add a short FAQ section last. That sequence produces the most extraction improvement per editing hour without a full rewrite or any developer involvement.

Why do AI engines cite review sites instead of a product's own pages?

Review sites tend to publish structured, specific, question-answering content: named features, stated prices, direct comparisons. A product page that opens with brand claims and vague capability statements has no extractable block. The engine cites whichever page holds the most specific self-contained answer to the query, whoever published it.

Answer-First Content | The GEO Playbook

An AI engine cites a page by lifting a passage out of it, not by summarizing the page as a whole. To be cited, a page needs at least one self-contained block of 40 to 60 words that opens with a direct answer, names its own subject, carries a specific claim, and makes complete sense read in isolation. Every technique in this part exists to produce that block.

Where Part 1 stops and this part starts

Part 1 covered which index each engine retrieves from, which crawler user-agents to allow, and how citation formats differ across ChatGPT, Claude, Perplexity and Gemini. That layer decides whether your page enters the candidate pool at all. This part goes one level down: once a page is in the pool, what makes one passage the one that gets extracted rather than skipped.

The mechanism is passage retrieval. Each engine splits your page into chunks, typically 150 to 300 words per block according to practitioner analysis published by Ziptie, then scores those chunks for relevance and extractability. A page full of strong prose but no extractable passages loses to a shorter page that leads every section with a direct answer. The writing decisions below change how your chunks score.

Extraction happens at the passage level, not the page level. Write for the chunk, not for the article.

The extractable block: what it is and what disqualifies one

An extractable block is a passage that can be lifted out of its page and dropped into an AI answer without editing and without additional context. It has four properties.

Self-contained. It does not depend on a preceding paragraph to make sense. Pronouns without antecedents ("This approach works because..."), references to earlier sections ("As discussed above...") and dependent clauses that assume the reader has read the page all disqualify a block.
Answer-first. The first sentence resolves the question the heading raised. Engines often keep only the first sentence of a retrieved chunk, so a first sentence that sets context instead of answering fails.
Specific. Generic claims ("Our platform improves efficiency") are not citable. Named entities, dates, percentages and product names give the engine something concrete to attribute. A block without at least one specific claim is rarely cited.
40 to 60 words. This is a practitioner convention rather than a published research finding, and the mechanical reason it holds is chunking: a block in that range is long enough to carry a real claim and short enough to survive inside a single retrieval chunk instead of being split across two. Conbersa is one of several GEO practitioners who land on the same range. Treat it as a target, not a rule: a 70-word block that answers cleanly beats a 45-word one that does not.

What disqualifies a block

The following patterns are common in B2B copy and reliably prevent extraction.

Opening with a pronoun: "It integrates with..." What does? Name it.
Opening with a transition: "Additionally," "Furthermore," "Building on this..."
Burying the answer behind two sentences of context.
Vague quantifiers: "many," "several," "significant," "various."
Cross-references: "see the table below," "as mentioned in the previous section."
Promotional language: superlatives, brand claims and calls to action inside the block.

A useful test: cover the heading and read only the first sentence. If it does not answer a real question on its own, the block is not extractable.

Heading hierarchy: write the question, answer it immediately

A heading is not a label. It is the question your buyer typed into an AI engine, and the first sentence under it is the answer. This two-part structure is the most reliable way to produce extractable blocks at scale, because it forces every section to open with a direct response instead of a warm-up paragraph.

How to phrase an H2 or H3 as a buyer question

Question-phrased headings map directly onto conversational queries. They also match the query patterns AI engines use to pair candidate pages with a user prompt. The rules are short:

Open with "How," "What," "Why," "When" or "Which."
Name the subject explicitly: "How does [Product] handle multi-tenant permissions?" not "How does it handle permissions?"
Use the vocabulary your buyers use in sales calls and support tickets, not your internal product language.
Keep the heading under 12 words so it reads as a natural query.

The first sentence rule

The first sentence under any question heading must resolve that question. Not introduce it, not contextualize it: resolve it. If the answer needs a caveat, the caveat goes in the second sentence. SKROL's content structure guide puts it plainly: "The model often keeps your first sentence and drops the rest, so make the first sentence the citable one."

A practical check: after drafting a section, read only the heading and the first sentence aloud. If a colleague would say "that answers my question," the block passes. If they would say "okay, but what is the actual answer?" the first sentence needs rewriting.

Every H2 and H3 is a question. The first sentence under it is the answer. Nothing else qualifies as answer-first writing.

Before and after: four B2B copy rewrites

The pairs below are patterns from real B2B product and content pages. Each "before" would not be extracted. Each "after" produces an extractable block by applying the rules above.

Pair 1: product feature description

Version	Copy	Why it fails or works
Before	Our platform offers robust integration capabilities that connect seamlessly with your existing tech stack, enabling teams to work more efficiently across their preferred tools.	No named subject, no specific claim, vague quantifiers, promotional language. There is nothing here to attribute.
After	[Product] connects to Salesforce, HubSpot and Slack through native API integrations that sync data in under 60 seconds. No middleware is required. Setup is one authenticated connection per tool and does not need a developer.	Named integrations, a specific time claim, a named action. A model can cite this and attribute it to the page without inference.

Pair 2: security and compliance section opening

Version	Copy	Why it fails or works
Before	Security is a top priority for us. We take a comprehensive approach to protecting your data, with multiple layers of defense and regular audits to ensure compliance with industry standards.	Opens with a brand claim, names no standard, quantifies nothing. "Industry standards" is not citable.
After	[Product] is SOC 2 Type II certified and GDPR compliant. Data is encrypted at rest with AES-256 and in transit with TLS 1.3. A third-party firm runs an annual penetration test, and enterprise customers can request the report under NDA.	Named certifications, named encryption standards, a named audit process. Every claim is specific and attributable.

Pair 3: pricing page opening

Version	Copy	Why it fails or works
Before	We offer flexible pricing options designed to scale with your business. Whether you are a startup or an enterprise, we have a plan that fits your needs and budget.	No numbers, no plan names, no differentiators. Interchangeable with every competitor's pricing page.
After	[Product] starts at $299 per month for up to 10 users on Starter, $799 per month for up to 50 users on Growth, and custom pricing above 50 users. Annual billing takes 20 percent off each plan.	Named plans, specific prices, specific seat limits, a specific discount. An engine can answer "how much does [Product] cost?" straight from this passage.

Pair 4: use case or customer problem section

Version	Copy	Why it fails or works
Before	Many marketing teams struggle with fragmented data across multiple platforms. This makes it difficult to get a clear picture of campaign performance and leads to inefficiencies in reporting and decision-making.	Opens with "many," names no subject, makes no specific claim. It could describe any product in any category.
After	B2B marketing teams that report on paid, organic and CRM performance pull those numbers from three separate tools. [Product] consolidates all three into one dashboard that refreshes every 15 minutes, so a weekly report is one export instead of three manual pulls.	Named audience, named data sources, a specific refresh interval, a specific outcome. Nothing in it has to be taken on faith.

Replace every vague quantifier with a number. Replace every "our platform" with the product name. If a claim cannot be made specific, cut it rather than dress it up.

Markup that changes what an engine can extract

A prose paragraph makes the engine parse sentence structure to find the answer. Structured markup removes that work. The elements below change what can be extracted without inference.

FAQ sections

A visible question and answer section is the most extractable shape of content on a web page, as Space and Story's analysis puts it. Each answer is a discrete, self-contained passage that maps to a discrete query. Keep three to five questions per page, each answer 40 to 60 words, and make sure none of them depends on the surrounding content to make sense.

Pair the visible section with FAQPage JSON-LD. Google has deprecated the FAQ rich result itself: its documentation records that the feature stopped appearing in Google Search on May 7, 2026, and the FAQ documentation was withdrawn on June 15, 2026. That removed a search feature, not the markup. FAQPage is still a valid Schema.org type, it is still crawled by the bots that read structured data, and the visible pairs underneath it are what engines actually lift. Which leads to the rule that matters more than the schema: every question and answer in the JSON-LD must also exist in visible HTML on the same page.

Comparison tables

A table lets an engine pull a specific cell without parsing prose. "Plan A costs $299 and supports 10 users, while Plan B costs $799 and supports 50 users" has to be parsed. The same values in labelled columns do not. Use a table wherever two or more items are compared across two or more dimensions, and label every column and row explicitly so an extracted value carries its own context.

Numbered lists for procedures

When content describes a sequence, a numbered list produces a chunk per step rather than one dense paragraph. Each step should open with an action verb, stay under 50 words, and be executable without reading the steps around it. Nobori's AEO schema guide notes that HowTo schema pairs well with procedural content for AI extraction. Google withdrew the HowTo rich result from both desktop and mobile results in September 2023, so there is no search feature left to win, but the markup still describes the sequence to any parser that reads it.

Definition blocks

When a page defines a term, put the definition in its own short paragraph immediately after the term appears, before any elaboration. The definition paragraph is the extractable block, and the elaboration can follow underneath it. An engine looking to define the term will lift the definition, not the elaboration.

Every markup decision should cut the inference work the engine has to do. Tables, lists and FAQ sections are not formatting choices, they are extraction architecture.

How to retrofit an existing page without rewriting it

Most B2B sites have dozens of pages that are structurally sound but not extractable, and a full rewrite is rarely necessary. The order below produces the most extraction improvement per hour of editing. Work it in sequence, because each step sets up the next.

Audit the headings first. Read every H2 and H3. If a heading is a label ("Features," "Our Approach," "Why Choose Us") rather than a question, rewrite it as a question. It is the fastest change and it forces every edit that follows.
Fix the first sentence under each heading. Once the headings are questions, read the first sentence under each one. If it does not answer the question, move the answer sentence to the top of the section. Do not rewrite the section yet, just promote the answer.
Replace vague quantifiers with numbers. Search the page for "many," "several," "significant," "various," "robust" and "comprehensive." Replace each with a specific number or a named entity, or delete the claim if no specific version of it exists.
Add a FAQ section at the foot of the page. Write three to five questions a buyer would actually type about this page's topic and answer each in 40 to 60 words. Add matching FAQPage JSON-LD and check it with the Schema Markup Validator rather than Google's Rich Results Test, which dropped FAQ support when the rich result was deprecated.
Convert prose comparisons to tables. Find any paragraph comparing two or more things across two or more dimensions and convert it. Keep a one-sentence introduction above the table so the table has context.
Update the dates. Freshness is a retrieval signal for Perplexity and ChatGPT. If the page has been substantively edited, update dateModified in the Article schema and the visible date on the page. A 2023 date competes poorly against a 2026 page on the same topic even when the content is equivalent.

Do not add FAQ content that is not genuinely visible on the page. Engines cross-reference JSON-LD against rendered content, and schema with no matching visible element gives them nothing to lift.

The retrofit order is headings, then first sentences, then specificity, then FAQ, then tables, then dates. Skipping ahead produces less improvement than working the sequence.

The self-containment test: one check before you publish

Before publishing any page written or retrofitted for extraction, run this on every section you intend to be cited.

Copy the heading and the first two sentences of the section into a blank document.
Read them without the surrounding page.
Ask whether the passage answers a real question, names its subject, carries a specific claim, and makes complete sense on its own.
If the answer to any part of that is no, the passage is not extractable. Rewrite the first sentence before publishing.

The test takes under two minutes per section and it catches the three most common failure modes: opening pronouns, buried answers, and vague claims that look acceptable in context and fall apart when lifted out. BrandCited's FAQ schema guide makes the same point about FAQ answers specifically: "Read each answer without seeing the question, page title, or surrounding text. If it does not make complete sense on its own, rewrite it until it does."

Apply the same standard to every section opening, not just the FAQ answers. The engine does not know which paragraphs you intended to be extractable. It scores all of them, and the ones that pass this test are the ones that get cited.

Putting it together: the anatomy of a citation-ready page

A page built for extraction has a predictable shape. The elements below are not optional additions to an existing template. They are the template.

Page element	Extraction function	Common mistake
Opening paragraph, 40 to 60 words	The primary extractable block. Answers the page's title question directly.	Used as a brand introduction or a context-setter instead of an answer.
H2 and H3 headings phrased as questions	Maps each section to a discrete query.	Label headings ("Features," "Benefits") that match no query at all.
First sentence under each heading	Secondary extractable block. Resolves the heading's question.	A transition sentence that restates the heading without answering it.
Comparison tables	Allows cell-level extraction for comparison queries.	Prose comparisons that have to be parsed to get one value out.
FAQ section, three to five questions	Purpose-built extractable blocks for high-frequency queries.	Generic questions that match no real buyer query.
`FAQPage` JSON-LD	Declares the question and answer pairs in machine-readable form for any parser that reads structured data.	Schema questions that do not appear in visible HTML on the same page.
Visible published and updated dates	Freshness signal for Perplexity and ChatGPT retrieval.	No visible date, or a date left untouched after the content changed.
Author name and organization	E-E-A-T signal for Gemini and Google AI Overviews.	Anonymous pages, or an author not tied to an Organization entity.

Every element in that table is something a content editor can fix without a developer, with one exception. The JSON-LD needs a <script type="application/ld+json"> tag in the page head, which usually means a CMS template change. Do the content-layer work first: it improves extraction immediately and does not wait on a deployment.