How do I structure content so AI can cite it

Structure a citation page as stacked extractable units: every H2 is the question a buyer asked, the next 40–80 words answer it as a standalone lift, then a table or numbered steps prove the claim. Keep entity names identical to About and schema, and make FAQ or HowTo JSON-LD match the HTML. Google publishes no word-count spec and no special AEO schema type. This skeleton lets a model quote a bounded passage instead of inventing one.

This is the page-outline spoke under the Answer Engine Optimization playbook. Which URLs to publish is cluster work. The identity packet across the web is entity architecture. This page is the HTML on one URL: headings, answers, objects, matching markup. I have been SEO certified since 2021. I will not pretend Google wrote a 60-word ranking rule.

The short answer

• Write each H2 as one buyer question. One question per heading. No metaphor titles that hide the job.
• Answer in the first 40–80 words under that H2. That range is an operator lift target, not a Google documented spec.
• Follow the answer with one bounded object: a table, numbered steps, a decision list, or a - [ ] checklist.
• Repeat the same legal name, product name, and offer name that About and Organization schema already use.
• Emit FAQ or HowTo JSON-LD only for questions and steps that already render. Matching visible text is Google’s rule. Extra markup is not an Overview ticket.

Layer	Job	Failure if you skip it
H2	Names the question the fan-out will ask	Model lifts a clever headline and a wrong paragraph
40–80 word answer	Standalone lift	Slogan, or a section pretending to be a lead
Table / steps	Bounded proof	Wall of prose; the cite becomes a paraphrase
Entity strings	Same name as About + schema	“Acme,” “Acme Inc,” and “we” in three sections
Schema	Labels the HTML	JSON-LD answers a question the reader never saw

If the page cannot survive a paste of one H2 plus its first paragraph into a blank doc, it is not structured for citation. Fix the outline. Do not buy a schema plugin.

What does “structure so AI can cite it” actually mean?

It means the page is a stack of passages a retrieval system can lift without inventing connective tissue. A citation is a named URL attached to a generated sentence. Rank is a list position. Those are different jobs. Structure serves the first: a heading that matches the question, an answer that still makes sense out of context, and an object with edges.

Google’s generative-AI guide tells you to organize pages with paragraphs, sections, and headings people can follow, and to keep important content in textual form (Google: optimizing for generative AI features; Google: AI features and your website). It also says you do not need to rewrite copy “just for AI systems,” that there is no ideal page length, and that there is no special schema.org type for generative AI search. Structure here is people-first outline work that also survives compression. It is not a secret markup file.

Claim	Honest status	What you do
“Google requires question H2s”	Not documented	Use them because humans scan and fan-out asks questions
“Google requires 40–80 words”	False	Operator range. Google says it has no preferred word count (people-first content)
“FAQ schema gets you cited”	False as a ticket	Visible Q&A can extract. JSON-LD is a label
“One ultimate guide wins the parent query”	Weak under fan-out	One question per H2; one URL per distinct intent
“Structure guarantees a citation”	False	Indexed + snippet-eligible is the floor. Citation is not a promise

This spoke is not the inverted-pyramid essay for the page lead, not the cluster map, and not the Wikidata / sameAs packet. Those jobs exist. They are different URLs. If you mix them, you will rewrite the homepage when the heading was the defect.

• The primary question is answered in the first breath of the page
• Each H2 owns one remaining question
• Each H2 is followed by a lift-ready answer, then one object
• Names on this URL match About
• Any JSON-LD on this URL can be found as HTML

A structured page is boring on purpose. Clever is a headline. Citation is a unit.

Why should every H2 be a question the buyer asked?

Because the heading is the retrieval handle. Google documents that AI Overviews and AI Mode may use query fan-out: the system issues related searches across subtopics, then attaches supporting links to the assembled response (Google: AI features and your website). A heading that already is that sub-question is easier to match than “Our Approach” followed by a paragraph that eventually mentions pricing.

Nielsen Norman Group’s inverted-pyramid work is recursive: the headline, the first paragraph, the first sentence of each section, and the first words of each sentence should carry information (NN/g: Inverted Pyramid). A question H2 is that rule applied to the outline. It is not a Google ranking factor I can cite. It is how you stop hiding the take under atmosphere.

H2 you shipped	What a buyer asked	What a model can lift
Our Approach	How do you run the engagement?	Atmosphere. Maybe a verb in paragraph four
Why Choose Us	Why hire this firm instead of a directory listing?	Adjectives. No constraint
Overview	What is this page for?	Nothing bounded
How do I structure a service page so AI can cite the offer?	The actual question	Heading + first paragraph as a unit
What does the audit include, and what does it exclude?	Scope	Table of in / out

Write the H2 as the question you would type into ChatGPT or Google, not as a brand line. Keep it to one question. Two questions in one heading is two answers fighting for sentence one.

Procedure for rewriting headings on an existing URL:

1. Paste the current H2s into a list. Ignore the body for one pass.
2. Next to each heading, write the question a buyer would actually ask.
3. If two headings map to the same question, merge the sections.
4. If a heading maps to no buyer question, it is decoration. Cut it or demote it.
5. Phrase the surviving H2s as those questions. End with a ? when the question is the whole heading.
6. Open each section with the answer, then the object. Do not keep the old wind-up.

Test	Pass	Fail
Stranger reads only the H2	Knows the question the section answers	Needs the H1 to guess
H2 copied into a search box	It is a real query	It is a slogan
Two H2s swapped	Bodies would break	Bodies would still work — duplicates
H2 plus first sentence	Complete enough to quote	“In this section we will cover…”

Google can match synonyms. You still should not make the model infer the question from a pun. The W3C heading tutorial exists for humans and assistive tech: headings describe the sections they introduce (W3C: Headings). That is enough justification. You do not need an AEO myth on top of it.

A question H2 that the body does not answer is worse than a dull H2. Do not title a section as a question you then dodge.

How long should the answer under each H2 be?

Aim for roughly 40–80 words, about two to four sentences. Long enough to name the subject, the answer, and one constraint. Short enough to lift whole. That is an editorial target for the section lead, not a ranking threshold.

Google is explicit that it does not have a preferred word count, that there is no ideal page length for generative AI features, and that you should not rewrite content just for AI systems (people-first content; generative-AI optimization guide). If a slide says “Google wants 60 words under every H2,” the slide is wrong. The rest of the section can be as long as the job requires. The lead is the lift.

Word count under the H2, before the table	What it usually is	What to do
Under 25	Slogan or category claim	Add the mechanism or the constraint
40–80	Lift-ready answer	Keep; move proof into the object
80–120	Two answers sharing a paragraph	Split; promote the second to its own H2
120+	A section pretending to be a lead	Cut or demote. The lead is not the article

Checklist for the section lead:

• Names the subject in plain language on first use
• States the direct answer without a throat-clear
• Includes one constraint, exception, or cost so it is not a slogan
• Would still make sense pasted into a chat answer with no H1
• Does not depend on a hero image, a previous H2, or “as we said above”
• Uses the same product and company names as the rest of the URL

The page-level 40–80 word opening is a different unit. That is the inverted-pyramid lead for the whole URL. This section is the recursive version: every H2 gets its own lift. If you only fix the hero paragraph and leave “Our Process” as a novella, you structured one passage and abandoned the rest.

Do not pad to 40 words with filler. A 32-word answer with a constraint beats an 80-word paragraph that restates the H2 three times. The band is a smell test, not a quota.

Section-lift test, one H2 at a time:

1. Copy the H2, the first paragraph, and the table or steps into a blank document.
2. Hide the rest of the page. Hide the H1.
3. Ask: would a stranger get the answer, the constraint, and the entity name?
4. If they need “the previous section,” you buried a dependency. Name the subject again.
5. If they get a slogan and no mechanism, you are under 25 words of actual answer. Add the constraint.
6. If they get two answers, split the H2.

Pass that test before you argue about schema. Schema cannot label a passage that does not exist.

What belongs under the answer — table, steps, or checklist?

One bounded object. Models extract edges better than atmosphere. A table is a set of rows. Numbered steps are a sequence. A checklist is a set of tests. Pick the object that matches the question type, then stop adding a second object that repeats the first.

Google’s AI-features page says to make sure important content is available in textual form (AI features and your website). A screenshot of a framework is not a citeable unit unless the same facts exist as HTML. Semantic lists and tables are text. A JPG of a funnel is not.

Question type	Object that extracts	Object that usually fails
What is X / what is not X	Comparison table vs near-terms	A metaphor paragraph
How do I X	Numbered procedure	A narrative of how it “feels”
Should I X or Y	Use-when / avoid-when table	A both-sides essay with no default
What breaks	Failure table: break, cost, replace	A scare sentence with no cost
Did we ship it	- [ ] checklist	A “we value quality” closer

Procedure for choosing the object:

1. Read the H2 out loud. If it starts with “how,” default to numbered steps.
2. If it starts with “what is” or “what’s the difference,” write a two-to-five-row table.
3. If it starts with “when” or “should,” write use-when / skip-when rows.
4. If it starts with “how do I know,” write a checklist a reviewer can tick.
5. If you want a quote block and a table, keep the table. Cut the quote unless it is a primary source.

Object	Citeable because	Abuse
Markdown table	Rows are discrete facts	Ten-column soup nobody can lift
Numbered steps	Order is the claim	Step 1 is a paragraph of philosophy
Checklist	Pass/fail tests	Fake checks that are slogans
Definition list	Term/meaning pairs	Keyword stuffing as <dt>

After the object, one closing line with a point. Do not end the section by restating the H2. End on the constraint the table just proved.

If the only honest object is “it depends,” the H2 is still too wide. Split the question until a table can hold a default.

How do I keep entity names consistent on this URL?

Pick one public string for the Organization, one for the person who stands behind it, and one for each product or offer on the page. Repeat those strings in the H1, the section leads, the table cells, and the schema name fields. Do not treat “the studio,” “SS,” and a former DBA as interchangeable in the same article.

This is not the full identity packet. Wikidata, sameAs, and Knowledge Graph disambiguation live on the entity architecture spoke. The job here is narrower: this URL must not teach a model three versions of you. Google’s structured-data quality rules already forbid marking up a performer the HTML does not describe (structured-data policies). The same honesty belongs in the prose. If the JSON-LD says Spurlock Studios and the H2 says “Spurlock Studio,” you handed retrieval a conflict.

String class	Lock to	Illegal variants on the same URL
Organization	Legal / public brand name	Missing “Studios,” old LLC string, nickname only
Person	Byline name	First name only in schema, full name in the lead
Product / offer	The SKU or offer you actually sell	Internal code names, retired packages, “our flagship”
Place	One HQ or service area claim	Three cities in three sections
Price / scope	The number on /contact or the offer page	“Starts at” in the lead, a different floor in the FAQ

Inventory before you rewrite:

1. Copy the Organization name from your live JSON-LD (or from About if schema is missing).
2. Copy the Person name from the byline and the author page.
3. Copy every product or offer mentioned on this URL into a list.
4. Search the draft for each. Highlight every mismatch, including pluralization that changes the entity (“Studio” vs “Studios”).
5. Fix the HTML first. Then change schema to match. Never the reverse.

Drift you will ship	What a model does	Fix
Three legal names	Hedges or picks a competitor with a cleaner packet	One name, used in H1, About, schema
Offer renamed in the hero, old name in the FAQ	Cites the retired package	Same string in hero, table, FAQ, JSON-LD
“We” with no noun for 400 words	Category answer, no brand	Name the Organization in the lead and once per major H2
Author byline ≠ Person schema	Splits the expert node	One byline string

A pronoun is fine after the noun exists. A page that never names the firm is a category essay. Category essays get cited as “experts say,” which is the opposite of the job.

Do not add a new brand nickname in this rewrite. Nicknames are an entity-architecture change. This URL only repeats the lock you already chose.

How should schema match the visible HTML?

If you emit structured data, every claim in it must already be visible to a reader of the page. Google’s general structured-data guidelines say not to mark up content that is not visible, and the AI-features page lists matching structured data to visible text as ordinary SEO hygiene (structured-data policies; AI features). The generative-AI guide is equally blunt: structured data is not required for generative AI search, and there is no special schema.org markup to add (optimization guide).

So the schema job on a citation URL is match, not invention. FAQPage is a WebPage with one or more frequently asked questions (schema.org FAQPage). If you use it, the Question.name values must be the FAQ headings on the page, and the acceptedAnswer text must be the answers the reader can open. HowTo steps in JSON-LD must be the numbered steps in the HTML. Article headline must be the H1 you actually rendered.

Google limited FAQ and HowTo rich results for most sites in August 2023 (HowTo and FAQ changes). Matching schema is still a quality rule. It is not a ticket into AI Overviews. Do not brief a plugin as if JSON-LD were the citation.

HTML on the page	Honest schema field	Mismatch you will ship
H1	Article.headline / WebPage.name	SEO title in JSON-LD, different H1 in the template
H2 question	Not automatically a FAQ. Only if it is also a visible FAQ item	Every H2 dumped into FAQPage
## FAQ H3s ending in ?	FAQPage Question.name	Schema questions from a tag manager, different wording
Numbered how-to	HowToStep.name / text	Seven steps in JSON-LD, five on the page
Organization name in the footer / About	Organization.name	DBA in schema, trading name in the H1
Price in a table	Offer.price if you mark an offer	Schema price from last year’s CMS field

Procedure for a schema match pass:

1. View the rendered page as a human. Copy the H1, the FAQ H3s, and any how-to steps into a scratch file.
2. Extract the JSON-LD. Diff question strings and step strings against the scratch file.
3. Delete JSON-LD that has no HTML twin. Do not “fix” it by hiding extra FAQ in the head.
4. If an accordion is collapsed by default but the answer is in the HTML, that can still be visible content. If the answer loads only after a click that never renders the text, it is not visible.
5. Re-test after the next copy change in the same commit. Schema drift is a weekly defect, not a launch defect.

Do	Do not
Write the HTML FAQ, then label it	Generate 12 questions in a plugin and hope the body catches up
One FAQPage block for the questions on this URL	Sitewide FAQ graph injected on every template
Keep Organization / Article / Product where they still describe the page	Treat leftover FAQ rich-result screenshots as a 2026 KPI
Validate that Googlebot can see the text	Assume a designer mock is what got crawled

Unused structured data is not, by Google’s 2023 wording, an emergency ranking penalty. Dirty, mismatched markup is a different problem: it is a false description of the page. Fix false descriptions. Do not add types “for AEO.”

What page skeleton should a citation URL share?

One H1 that states the job. A 40–80 word page lead that answers the primary question. A short-answer block. Then 8–14 H2s, each a question, each with a lift and an object. A ## FAQ with six question H3s if you run this chrome. A CTA that names the next human step. That is the skeleton. Formats change the objects inside the H2s. They should not change the stack.

Zone	What it is	Extract job
H1	The page’s one job, close to the primary query	Retrieval title
Page lead	40–80 words, standalone	The sentence a model quotes if it only takes the top
Short answer	3–5 bullets	Backup lift if the lead is skipped
H2 stack	One question each	Fan-out targets
FAQ H3s	Six questions ending in ?	Bounded Q&A; schema twin if you emit FAQPage
CTA	Lane page + intent	Human next step, not a citation unit

Numbered skeleton for a service or playbook spoke:

1. H1 = the primary question or a close paraphrase.
2. Lead paragraph answers it, names the Organization, names one constraint.
3. Optional one-line placement: which pillar this sits under, which adjacent jobs it is not.
4. ## The short answer with bullets a stranger could ship.
5. H2s as remaining questions in the order a practitioner would hit them: definition, method, object, names, schema, measurement, failure, week, skip.
6. ## FAQ with six H3s that end in ?, answers of two to four sentences.
7. ## CTA with /visibility and /contact?intent=visibility-audit when the offer is an audit.

Skeleton mistake	What it costs	Replace with
H1 is a pun, H2s are “Overview / Process / Benefits”	No handles	Question outline
FAQ is the only Q&A; body is a memoir	FAQ gets lifted, body is ignored	Same questions answered in the H2s, FAQ is the recap
CTA in the middle of the article	Interrupts the extract unit	One CTA after FAQ, per this chrome
Jump links as a substitute for leads	Table of contents without answers	Leads under the H2s
Duplicate H1 in an image	If the text is not in HTML, it may not exist for retrieval	Real H1 in the document

Google says perfect semantic HTML is not required, and that the web is not valid HTML, but that semantic headings help people including screen-reader users (optimization guide). Do not skip from H2 to H4 to look fancy. The outline is a tree. A skipped level is a broken tree for the reader you already have.

If a section cannot fit this skeleton, it is probably two URLs. Split it. That is cluster work, not a reason to revive “Part 1” as an H2.

How do definition, how-to, and comparison pages differ?

The skeleton stays. The object inside each H2 changes. A definition page earns tables that bound the term. A how-to earns numbered steps that a stranger can run. A comparison earns a default plus the exception. Do not force every URL into FAQ sludge, and do not force a case-study narrative into question H2s it cannot answer honestly.

Page type	H2 pattern	Lead under each H2	Object
Definition	“What is X?” / “How is X different from Y?”	Bound the term	Comparison table
How-to	“How do I X?”	State the sequence in one line	Numbered steps, one action each
Comparison / vs	“Should I X or Y?”	Pick a default	Use-when table
Failure	“What breaks if I X?”	Name the break and the cost	Failure table
Measurement	“How do I know X worked?”	Name the metric and the window	Log columns
Brand film / case study	Often not this skeleton	Story first is the point	Do not AEO-wash it

Worked pattern for a how-to H2:

1. H2: the question as asked.
2. Two to four sentences: the sequence in miniature, plus the constraint (time, tools, or what you skip).
3. Numbered steps with verbs. One action per step. No step is a lecture.
4. A small table for the decision inside the procedure, if a step branches.
5. Close on the failure if you skip a step.

How-to step	Good	Bad
1	Export the live H2s to a list	“Begin by considering your audience”
2	Rewrite each as a buyer question	“Kick off with a vision workshop”
3	Write the 40–80 word answer	“In today’s landscape, structure matters”
4	Add the table	Paste a screenshot of a whiteboard

Worked before / after for one service section (not the page lead):

Piece	Before	After
H2	Our Approach	What does the visibility audit include, and what does it exclude?
Lead	We believe in partnership and a tailored process for every brand.	The audit covers eligibility, extractable pages, and a frozen prompt panel. It excludes a 40-post calendar and schema plugins sold as Overview tickets.
Object	Three adjectives and a stock photo	Table: in scope / out of scope / owner
Name	“the studio”	The Organization name already on About

If the after column still needs the H1 to make sense, the lead is not standalone. Rewrite the lead. Do not add a fourth adjective.

Worked pattern for a comparison H2:

• Lead: pick A unless condition C, then B.
• Table: A / B / when / cost of being wrong.
• Do not “it depends” without a default. A default is a citeable sentence. An essay is a hedge.

If the page is a portfolio story, skip this skeleton. A film still and a narrative are allowed to be a narrative. The citation URL is the service, pricing, FAQ, or playbook page. Do not flatten every URL into the same question list. That failure has its own cost: the brand sounds like a FAQ generator, and the model still cannot tell who you are.

What fails first when teams try this?

They change the headings to questions and leave the first paragraph as a wind-up. The H2 now matches the query. The lift is still “In this section we explore.” Retrieval got a handle and no object. The second failure is emitting FAQPage from a tag manager so every template claims the same ten answers. The third is renaming the offer in the hero and forgetting the FAQ.

I have shipped hundreds of production sites. The defect is almost never “we needed more schema types.” It is a mismatch between the outline a human can use and the outline a model can lift.

Failure	What it looks like	Cost	Replace with
Question H2, buried answer	H2 is perfect; sentence one is throat-clearing	Cite goes to a competitor with a duller H2 and a real lead	Answer in sentence one
Schema-only FAQ	JSON-LD Q&A, no HTML twin	Policy violation; models that read HTML never see it	Visible FAQ, then label
Three names on one URL	Hero, FAQ, footer disagree	Hedged or wrong brand in the answer	String inventory
Ultimate guide, 28 H2s, no objects	4,000 words of memoir	Fan-out cannot find a unit	One object per H2; split the URL if needed
Screenshot as the only proof	Framework lives in a PNG	Not textual form	HTML table with the same cells
nosnippet or max-snippet:0	Perfect outline, zero snippet	Google’s floor for AI features is snippet-eligible	Fix robots / preview controls first
Word-count theater	Writers pad to 80	Google already said it has no preferred count	Constraint, not filler

Checklist when a “we restructured for AEO” ticket comes back with no citations:

• Is the URL indexed and snippet-eligible? If no, stop talking about H2s.
• Does each H2 still have a wind-up under it? Read sentence one only.
• Is there one table or step list per H2, in HTML?
• Do FAQ H3s match FAQ JSON-LD character for character, or did you skip FAQ schema (also valid)?
• Do Organization / offer strings match About?
• Did you measure on a frozen prompt panel, or did you refresh Search Console once?

The expensive version of this failure is a sitewide “AEO voice” rewrite that turns case studies into fake FAQs and leaves the money URLs as “Our Approach.” You spent the week on pages models were never going to quote, and you did not touch the ones they would.

If leadership wants a citation-rate multiplier for “question H2s,” I do not have one to invent. Public AEO blogs invent them. Google did not publish one. Log prompts instead.

How do I measure whether the structure is working?

You measure whether this URL starts getting cited on a frozen set of questions it was rewritten to answer — not whether sitewide impressions moved. Citation rate is noisy. Google’s AI features often do not trigger (AI features). OpenAI’s ChatGPT Search help says search citations can be incomplete, outdated, or incorrect, and that you should open the source (ChatGPT Search). One screenshot is an anecdote.

Search Console’s generative AI performance report, launched in June 2026, counts impressions in those features (Google: generative AI performance reports). It is not a citation log, and it is not a lead. Use it as a supporting series. The operator log is the panel.

Column	What you record	Pass signal	Fake cousin
Prompt	Exact string, frozen	Same 15–40 prompts for 28 days	Whatever someone typed this morning
Engine	AI Overview / AI Mode / ChatGPT / Perplexity	Named product	“AI”
Cited?	Yes / no / wrong-page	Your URL in the supporting links or inline cite	“They paraphrased us” with no URL
Passage	Which H2 + object got lifted	Matches the section you rewrote	Homepage cited for a spoke question
Name accuracy	Legal name / offer correct	Matches the string lock	Close enough
GSC AI impressions	If the report exists on the property	Directional only	Treated as ROI

Procedure:

1. Day 0: freeze 15–40 prompts that the URL is supposed to own. Include the H2 questions verbatim.
2. Log citation, wrong-page cites, and name errors before the rewrite. Screenshot the SERP or chat.
3. Ship the skeleton. Request indexing if the URL was blocked; do not expect same-day recrawl (AI features troubleshooting notes recrawl can take days to months).
4. Day 14: same prompts, same engines. Do not add new prompts mid-test.
5. Day 28: keep or recut. If eligibility was clean and the outline is extractable and you still have zero cites, the next diagnosis is corroboration and competition — not another FAQ plugin.

Window	Honest read	Too early
Same day	Eligibility and quote-test of the HTML	“We are cited now” after one Overview
14 days	Direction on the frozen panel	A Core Update week treated as your rewrite
28 days	Whether the section stack is doing work	Sitewide CTR
Tagged leads	Separate scoreboard	Impressions labeled as pipeline

If the model cites a different URL of yours, that is a cluster and cannibalization problem, not a heading problem. If it cites a competitor with a worse outline, you still may have an entity or corroboration gap. Structure is necessary. It is not sufficient. Do not sell it as sufficient.

What should I skip if I only have a week?

Do not start a cluster rebuild, a Wikidata campaign, or a sitewide FAQPage injection. Pick three money URLs that already rank or already get asked in sales, and put this skeleton on those three. Homepage identity sentence, one service or product URL, and the FAQ or offer page is enough for seven days.

Do this week	Skip this week
Rewrite H2s on three URLs into real questions	40-blog calendar “for AEO”
40–80 word answers + one table or steps per H2	llms.txt as the project
String inventory vs About	New schema types “for Overviews”
FAQ HTML that matches any JSON-LD you keep	Tag-manager FAQ on every template
Frozen 15-prompt panel, day-0 screenshots	A vendor GEO score as the KPI
Confirm snippet eligibility	Word-count padding to hit 80

Day-by-day if you actually have five working days:

1. Day 1: eligibility (indexed, snippet allowed, important text in HTML). Kill nosnippet mistakes. List the three URLs.
2. Day 2: H2 rewrite + section leads on URL 1. One object per H2.
3. Day 3: same pass on URL 2. String inventory across both.
4. Day 4: URL 3 + FAQ match + Organization name lock.
5. Day 5: freeze prompts, screenshot, request recrawl. Stop.

If a page cannot take question H2s without lying (a brand film, a photo essay), skip it. The week is for citeable URLs. Cluster planning and entity architecture are the next weeks, on their own spokes, not jammed into Friday.

A week is enough to make three URLs extractable. It is not enough to declare the program a success. Say that in the recap or someone will equate “we changed headings” with “we got cited.”

When is this structure not worth doing yet?

When the URL cannot appear as a supporting link anyway, when you do not have an offer you will take, or when the page’s job is narrative. Google’s floor for AI Overviews and AI Mode supporting links is indexed and eligible to show in Search with a snippet, with no extra technical requirements (AI features). If you fail that floor, H2 craft is theater.

Blocker	Why structure waits	What to do first
noindex, nosnippet, or max-snippet:0	Not snippet-eligible	Fix preview controls
Not in the index	Nothing to retrieve	Indexing, crawl, internal links
Mid-rebrand, two legal names in flight	You will teach both	Lock the name, then rewrite
No offer, no form, no calendar	A cite with nowhere to go	Ship the offer page
Case study / film page	Story is the product	Leave it; structure the service URL
Duplicate URL already answers the question	You will cannibalize	Merge or retire, then structure the survivor
Leadership wants a word-count SLA	You will pad	Show Google’s “no preferred word count” line

Also wait if the only planned change is schema. Matching JSON-LD to HTML is a hygiene pass on a page that already has answers. It is not a substitute for answers. If the body is still “Our Approach,” labeled FAQ will not save it.

If you are going to restructure, do it on URLs that can take a question outline without becoming a doorway. Thin pages with a perfect skeleton are still thin. Google’s people-first guide asks whether someone would leave feeling they learned enough to achieve the goal. A 40-word lead and a two-row table on a topic that needs a procedure is a doorway with nice headings.

When the blockers are gone, this structure is worth doing because it is cheap relative to a new content calendar and it is the outline humans already needed. You are not waiting for Google to bless a word count. You are making the page a machine can quote without lying.

FAQ

How do I structure content so AI can cite it?

Write every H2 as one buyer question, answer it in the first 40–80 words so the passage stands alone, then add a table, numbered steps, or a checklist. Keep Organization, Person, and Product names identical to About and schema, and make any FAQ or HowTo JSON-LD match that HTML. Google does not publish this as a spec and does not require special AEO markup. Citation still requires an indexed, snippet-eligible URL.

How do I measure whether this structure is working?

Freeze 15–40 prompts the URL should own, including the H2 questions, and log citation, wrong-page cites, and name errors on named engines for 28 days. Use Search Console generative AI impressions only as a supporting series if the report exists on the property. A sitewide CTR sparkline is not this measurement. One Overview screenshot is not a trend.

What usually fails first when teams try this?

They convert headings to questions and leave the first paragraph as a wind-up, so the handle exists and the lift does not. The next failures are schema-only FAQs that do not exist in HTML, and three brand or offer strings on one URL. Fix sentence one under each H2, then match schema, then lock names. Do not start with a plugin.

How long does this take to show results?

The HTML can pass a section-lift test the day you ship. Recrawl can take days to months, and AI features often do not trigger, so a provisional citation read wants about 14 days on a frozen panel and a locked read closer to 28. I will not invent a week-count SLA. If eligibility was broken, the clock has not started.

What should I skip if I only have a week?

Skip the cluster rebuild, Wikidata, llms.txt theater, and sitewide FAQ injection. Restructure three money URLs: question H2s, 40–80 word answers, one object per section, names matched to About, snippet eligibility confirmed. Freeze a 15-prompt panel on day 0. That is a week. It is not a program victory.

When is this not worth doing yet?

When the URL is not snippet-eligible, not indexed, mid-rebrand, or has no offer, and when the page is a narrative case study or film. Fix robots, index, and the name lock first. Do not AEO-wash a story page. Structure the service, FAQ, or playbook URL that a model should actually quote.

CTA

If the outline cannot be pasted as H2 plus first paragraph and still make sense, it is not ready to cite. Fix three money URLs, then get a second pair of eyes on the panel.

Lane: /visibility · Book a visibility audit.