How to Do AEO for Your Website: Answer Blocks, Schema, FAQ Pages

Quick answer

Canlah AI does AEO for a website by rewriting three blocks on every page a brand wants quoted: the labelled answer block above the first H2, the first sentence of each H2 and the table that carries the numbers. In a Canlah AI teardown of 526 citations and 92 cited pages from 60 buyer queries run on ChatGPT and Google AI Mode in September 2026, 32 of 75 located copied facts, or 43%, sat above the first H2. A further 23 of 75, or 31%, came from the first or second H2, and about 55 of 75, or 73%, sat inside the first 30% of the page.

Canlah AI ships this work as the on-site stage of a GEO retainer, meaning answer capsules, schema, a crawlable facts page and comparison pages. This guide publishes the manual version so a team with no tool can apply the same edits.

Treat the techniques below as directional. A teardown of cited pages shows which shapes engines copied once a page had been retrieved; it cannot prove that reshaping a page causes a citation, because the sample holds no uncited control group. Rewrite three pages, record the answers before and after with dates and engines, then decide whether to roll the pattern out.

Disclosure: Canlah AI publishes this guide and operates the paid measurement and on-site service described in the measurement section. The position and format figures come from a Canlah AI teardown of 92 pages cited by ChatGPT and Google AI Mode in Singapore in September 2026. Google, Schema.org and vendor details were checked on public pages on September 23, 2026, not through paid accounts.

What does AEO mean for a website?

AEO for a website means making each fact on a page copyable by answer engines, and six properties decide whether a block survives that copy step: liftability, position, container, precision, visibility in HTML and agreement. The engines in question write an answer instead of listing links: Google AI Overviews and Google AI Mode, ChatGPT, Perplexity and Microsoft Copilot.

• Liftability: one sentence cut from the page still states the whole answer, including its own subject.
• Position: the block sits above the first H2, where 43% of located copied facts sat, or inside the first two H2s, where a further 31% sat.
• Container: the fact lives in a table row under a judgment header rather than in a loose bullet.
• Precision: the figure is exact, such as S$3,900 to S$7,900, rather than rounded into a word such as affordable.
• Visibility in HTML: the number sits in text a fetch can read. In the searchVIU test of October 30, 2025, a price living only inside JSON-LD was found by none of five AI systems.
• Agreement: the numbers in the answer block match the numbers in the table below it.

Grade each block on its own against those six properties. That is the difference between an AEO strategy and a redesign: the unit of work is a copyable block that passes or fails, not a page that looks better.

Six AEO techniques, ranked by what engines copied

The labelled answer block is the AEO technique engines copied most, supplying 32 of 75 located copied facts, or 43%, in the Canlah AI sample. The liftable H2 opener follows with 23 of 75, or 31%, so the first two techniques together supplied 55 of 75, or 73%.

AEO techniques against the Canlah AI citation study: 92 cited pages, ChatGPT and Google AI Mode, study published September 23, 2026.

#	Technique	What to ship on the page	Share of located copied facts	Caveat before you roll it out
1	Labelled answer block	A 40 to 60 word Quick answer, key-figure strip or price card between H1 and the first H2, carrying the exact number	32 of 75, 43%	A position share, not proof that the block caused the citation
2	Liftable H2 opener	A first sentence under each H2 that names the subject, states the verdict and contains the digit	23 of 75, 31%	Works only when the opener carries the number; scene-setting openers were skipped
3	Decision table	Header cells that state what the column decides, body cells written as complete facts	17 of 75, 23%	The header row is copied with the cells, so a vague header travels into the answer
4	Exact figure in visible HTML	Prices, dates and eligibility conditions as readable text, duplicated in schema afterwards	Not separable from rows 1 to 3	The JSON-LD finding rests on one searchVIU test of eight prices, dated October 30, 2025
5	One exact sentence plus one range sentence	For an unregulated business: your own confirmed figure, then the market range with the conditions it depends on. Law firms and clinics publish their own figure and its conditions only, with no market range	Not measured by position; Step 6 gives the engine split	The split quoted in Step 6 covers dental and family-law queries; regulated sectors follow their own advertising rules rather than this pattern
6	Restating FAQ	Conclusions repeated as self-contained sentences, plus FAQPage markup	About 6 of 75, 8%	FAQ answers were copied mainly where the body never stated the conclusion

Source: Canlah AI teardown of 92 cited pages, September 2026. Positions were assigned by hand against the rendered page. Only 75 of the 526 citations resolved to a sentence because stored answers were truncated at 600 characters; rows 1 and 2 together cover 55 of 75, or 73%.

How to do AEO on a page, step by step

Doing AEO on a page takes six steps in a fixed order, starting with the sentence you want quoted, because a page with schema, a FAQ block and no answer sentence gives an engine nothing to copy.

Step 1: Write the sentence you want quoted before writing the page

Draft the sentence you want an engine to hand back, with your subject as its grammatical subject and the number inside it. A dental page reading “Invisalign at Advanced Dental costs S$3,900 to S$7,900, depending on how complex your case is” was copied by ChatGPT from about 1% into the page. “Prices vary by case” cannot be copied, because copying it leaves the engine nothing to tell the user.

Step 2: Put the exact figure in visible text before putting it anywhere else

Write the number into HTML a fetch can read, then duplicate it in structured data. One tuition provider in the Canlah AI sample was cited for its results numbers while its S$420 monthly fee never appeared in the page HTML, so ChatGPT sourced fees elsewhere. Schema would not have rescued it, because in the searchVIU test the direct-fetch pass did not read prices held only in JSON-LD.

Step 3: Open every H2 with the verdict

Move the conclusion to the first sentence of the section and leave the context for sentence two. This is the rule with the most evidence behind it: 15 of 18 cited family-law pages had the copied sentence sitting first in its paragraph. The counter-example in the Canlah AI sample is a prenuptial-agreement page whose H2 opened with “is a complicated matter” and was skipped, while its FAQ answer, which stated the conclusion, was copied instead.

Step 4: Give the table a header row that carries the judgment

Write header cells such as Best for, Winner, Market range or What to verify, and body cells as complete facts. Engines copied the header row with the cells, returning strings such as “Level | Part-Time Tutor | Full-Time Tutor | MOE Teacher” almost verbatim, so a header reading “Notes” arrives in the answer carrying no meaning.

Step 5: Make the answer block and the table agree

Check that every number in the opening block appears identically in the table under it. One cited tuition page quoted S$35 to S$55 for primary level in its top card while its own table quoted S$25 to S$35, and the engines used the table and skipped the card. A mismatch wastes the page’s most valuable block.

Step 6: Write one exact sentence and one range sentence

Give the page two sentences, because the two engines cite two different page types. In dental queries ChatGPT took 72% of its citations from government and public-institution pages while Google AI Mode took 88% from clinic price pages, and in family law the split was 78% government against 73% law-firm pages. For an unregulated business, write one sentence carrying your own exact figure, which is what ChatGPT copies from source pages, and one carrying the market range “depending on A, B and C”, which AI Mode copies from secondary syntheses. Law firms and clinics should publish only their own confirmed figure and the conditions it depends on, with no market-range or peer-comparison sentence: regulated sectors follow their own advertising rules, and in Singapore that means the Legal Profession (Professional Conduct) Rules for law firms and MOH’s advertising guidelines for clinics, dental practices and aesthetics providers, both of which restrict publishing peer fee comparisons.

A worked AEO example: one service page, before and after

A worked AEO example on a Singapore tuition service page turns one unquotable page into three separately copyable units without changing the design. It is reconstructed from failure patterns in the Canlah AI sample rather than from a single client page, and the figures inside the quoted copy are illustrative.

Before, the opening block: “We are a Singapore tuition centre with experienced teachers and a proven track record. Our fees are competitive and depend on level, group size and subject. Contact us for a quote.”

After, the opening block: “Primary group tuition at this centre costs S$25 to S$35 an hour in 2026, and secondary group tuition costs S$35 to S$55 an hour, depending on level, group size and subject. Across 18 Singapore centres with published rate cards reviewed in September 2026, the market range for primary group tuition was S$20 to S$45 an hour.”

Three things changed. The figure moved into visible text above the first H2, where 43% of copied facts sat. The sentence became liftable, because it names its subject and carries both numbers without needing the heading above it. The page also gained a market-range sentence with its conditions, the shape Google AI Mode copies from synthesis pages.

Before, the fee section H2 and opener: “Our fees. Every family’s situation is different, and we believe in flexible arrangements that suit your child.”

After, the fee section H2 and opener: “How much does primary tuition cost here? Primary group tuition costs S$25 to S$35 an hour, billed in four-lesson blocks, with no registration fee.” The table underneath repeats those numbers, one row per level, under headers reading Level, Group rate, One-to-one rate and What is included. An AEO example is finished only when the number in the block, the number in the opener and the number in the table are the same number.

What to put in schema, and what not to

Schema is a one-time template job and an entity signal, not a per-page AEO lever: in an Ahrefs difference-in-differences test across 1,885 treated pages and 4,000 matched controls, adding JSON-LD moved citations by −4.6% for AI Overviews, +2.4% for AI Mode and +2.2% for ChatGPT, and only the decline was statistically significant. Ship Organization once, the relevant service or product type once, Article on editorial pages and accurate dates.

{
  "@context": "https://schema.org",
  "@type": "FAQPage",
  "mainEntity": [{
    "@type": "Question",
    "name": "How much does primary group tuition cost?",
    "acceptedAnswer": {
      "@type": "Answer",
      "text": "Primary group tuition costs S$25 to S$35 an hour, billed in four-lesson blocks, with no registration fee."
    }
  }]
}

Two rules follow. Every fact inside the markup must already exist word for word in the visible page, because the JSON-LD copy is a duplicate rather than a home. FAQPage markup should also not be added expecting a search feature: Google restricted FAQ rich results to well-known government and health sites in August 2023, removed the rich result on May 7, 2026 and deleted the documentation in June 2026. Canlah AI validates syntax with the Schema Markup Validator and eligibility with Google’s Rich Results Test.

When does a separate FAQ page earn its place?

A separate FAQ page earns its place only when each question is a buyer query no other page answers, because FAQ blocks supplied only about six of 75 located copied facts, or 8%, in the Canlah AI sample. The cases where the FAQ was copied share one feature: the body had failed to state the conclusion.

Write each answer as a complete sentence naming its own subject, since an answer reading “It depends on the plan” is unusable outside its question. Never let a conclusion exist only in the FAQ.

How to interpret five common results

Five results come back when a page is re-checked after AEO edits, and each one points to a different block to fix: not mentioned, the wrong sentence copied, the table copied over the summary, cited by one engine only and cited with a stale number.

Not mentioned and not cited: Confirm the page answers a question buyers type, then check whether an answer sentence exists at all. A page built from sub-headings and bullets often holds no sentence an engine can lift.

Cited, but the wrong sentence was copied: The copied sentence is almost always first in whichever paragraph held a number. Move the number into the sentence you want quoted and take the competing number out of the paragraph above it.

Copied from the table instead of the summary: This is the Step 5 mismatch, and the signal is clean: the engines trusted the table. Correct the summary, never the table.

Cited by Google AI Mode but not by ChatGPT: The page is being read as a secondary synthesis rather than as a source. Add the first-party exact figure, the date it was verified and its denominator, because ChatGPT in the Canlah AI sample preferred official pricing pages and government schedules.

Mentioned with an out-of-date number: Fix the figure in the answer block, the table and the FAQ in one edit, then record the date. A stale number surviving in any one of the three blocks keeps reappearing, because engines copy the block rather than the page.

What AEO cannot do

AEO cannot do three things: put a page into the candidate pool, be graded by length or substitute for off-site evidence.

• It cannot put a page into the candidate pool. Google states there are no additional requirements to appear in AI Overviews or AI Mode and no special optimizations necessary, and describes a query fan-out that issues several related searches before an answer is written.
• It cannot be graded by length. Cited pages in the Canlah AI sample ran from 40 words to 9,000, median near 1,300 to 1,500, so adding words is the one change the evidence never rewards.
• It cannot substitute for off-site evidence. Canlah AI is the counter-example: its own site carried three JSON-LD blocks, zero parse errors, 80 of 80 pages crawled and all five AI answer-time crawlers allowed on September 7, 2026, and was still named in seven of 177 grounded non-branded answers, a 4.0% mention rate.

When to move from AEO edits to measurement

AEO edits need a measurement layer once a buyer question is worth re-checking on a schedule, and four vendors reviewed on September 23, 2026 publish one: Otterly AI, Peec AI, Profound and Ahrefs Brand Radar. A single answer cannot separate a real change from model volatility.

• Otterly AI. Best for: a small team that wants daily prompt tracking at a published price, since the Lite plan lists $29 a month for 15 prompts across ChatGPT, Google AI Overviews, Perplexity and Microsoft Copilot. What to verify: whether 15 prompts cover your buyer questions once each one runs on four engines. Public source: Otterly AI pricing.
• Peec AI. Best for: a team that expects to scale, with tiers listed from 50 prompts on three models to 350 prompts across five projects. What to verify: which three models the entry tier covers for your market. Public source: Peec AI pricing.
• Profound. Best for: a first look at no cost, through the free AEO Report it publishes. What to verify: the paid plan price, which was “Not found on reviewed page” because the pricing page routes to a demo. Public source: Profound pricing.
• Ahrefs Brand Radar. Best for: a team already working in Ahrefs, since Brand Radar runs on Ahrefs’ own index. What to verify: whether that index holds the buyer questions your market types, in the engines you care about. Public source: Ahrefs Brand Radar.
• Canlah AI. Best for: a team that wants the six on-site techniques delivered and the answers re-measured by one provider, reported as ranges with timestamped evidence. What to verify: the limitation stated below, and whether a managed service suits the team for the first three months; after three months of the GEO service, clients can run the work themselves with Canlah All-Stars. Public source: Canlah AI GEO services and facts page.

A grade for a page’s block structure was “Not found on reviewed page” for all four vendors, and Canlah AI measurement does not grade block structure either: each option records whether the brand appeared and which URL was cited.

“Not found on reviewed page” means the vendor’s public materials did not name that capability when reviewed on September 23, 2026. It is not proof that the capability is absent.

Canlah AI is a Singapore-based SEO + GEO agency that measures and improves brand visibility inside ChatGPT, Gemini, Google AI Overviews and Google AI Mode, reporting AI visibility as re-verifiable ranges with timestamped evidence. Its on-site stage is the six techniques above, delivered as answer capsules, schema, a crawlable facts page and comparison pages. Its limitation is the one its own audit exposes: Canlah AI can rebuild a page’s blocks in days, but its own site was named in only seven of 177 grounded non-branded answers with the same foundations in place, so it cannot promise that an engine will retrieve a client page. It starts as a managed service: a team that wants its own dashboard from day one, with no agency involved, is closer to the model Otterly AI and Peec AI sell, while Canlah AI clients can run the work themselves with Canlah All-Stars after three months of the GEO service.

Do not rewrite a page library from this guide alone. Rewrite three pages and record the exact answers before and after with engine, date and denominator. Keep the old evidence so the comparison survives the next model update, then decide whether the pattern has earned a wider rollout.

Frequently asked questions

How do you do AEO for a website?

You do AEO for a website by rewriting three blocks on each page you want quoted: the answer block above the first H2, the first sentence of each H2 and the table that carries the numbers. In the Canlah AI sample of 75 located copied facts, those three blocks supplied 43%, 31% and 23% of the copies. Schema and FAQ markup come after the visible text, never instead of it.

What is an AEO strategy in practice?

An AEO strategy is a page-by-page specification, not a campaign. For each page you decide the sentence you want quoted, put the exact figure in visible text above the first H2, open every H2 with the verdict, give the table a judgment header row and restate the conclusion in the FAQ. Everything else in an AEO strategy is design.

What are the AEO best practices for 2026?

The first AEO best practice for 2026 is to put the answer and its number in the first 30% of the page, where about 73% of copied facts sat in the Canlah AI sample. The others are to write the exact figure into HTML before schema, make the summary and the table agree and, for an unregulated business, carry one exact sentence for ChatGPT plus one market-range sentence for Google AI Mode; a law firm or clinic instead publishes its own figure and conditions only, since regulated sectors follow their own advertising rules rather than a market-range comparison. Length is not a best practice: cited pages ran from 40 to 9,000 words.

Does schema markup improve AEO results?

Schema markup is worth configuring once and is not a citation lever. Adding JSON-LD moved citations between −4.6% and +2.4% across 1,885 pages in the Ahrefs test, and a price placed only inside JSON-LD was retrieved by none of five AI systems in the searchVIU test. Keep schema for entity resolution and rich-result eligibility.

Can AEO guarantee that ChatGPT quotes my page?

No. Page structure decides what an engine copies once your page is in the candidate set, and retrieval is decided by the index, the query fan-out and the sources the engine already trusts. Google states that no special optimizations are required to appear in AI Overviews or AI Mode.

How is AEO different from GEO?

AEO is the on-page half: the blocks, figures, tables and markup that make a fact copyable. GEO covers the whole problem, including which third-party sources an engine retrieves before your page is considered. In the Canlah AI self-audit, on-site foundations were sound and the brand still appeared in 4.0% of grounded non-branded answers, which is a GEO problem rather than an AEO one.

Method and sources

The position and format figures come from a Canlah AI teardown published on September 23, 2026, covering 60 buyer queries across dentistry and aesthetics, ecommerce and consumer goods, family law, tuition and B2B SaaS. The queries were run through ChatGPT and Google AI Mode from Singapore, producing 526 citations across roughly 300 domains, of which 92 cited pages were read line by line. Answers were stored truncated at 600 characters, so only 75 copied facts could be matched to a sentence, and query terms came from Google autocomplete. The study has no uncited control group, so it shows which shapes were copied, not which shapes cause a citation. The worked example is reconstructed from failure patterns in that sample and is not a client page. Vendor and Google pages were checked on September 23, 2026 without paid accounts.

• Canlah AI cited-page format study, internal export, September 23, 2026. Position shares, container types, word-count range and the engine-by-industry splits.
• Canlah AI self-audit, September 7, 2026. The 4.0% mention rate, three JSON-LD blocks, 80-page crawl and crawler-access check.
• Ahrefs, we tracked 1,885 pages adding schema. Effect sizes for AI Overviews, AI Mode and ChatGPT.
• searchVIU, what ChatGPT, Claude, Perplexity and Gemini really see. The eight-price direct-fetch test, October 30, 2025.
• Google Search Central, AI features and your website. The no-special-optimizations statement and the query fan-out description.
• Google Search Central, documentation updates. FAQ rich result restriction in August 2023, removal on May 7, 2026 and documentation deletion in June 2026.
• Schema.org FAQPage and the Schema Markup Validator. Type status and validation.
• Otterly AI pricing, Peec AI pricing, Profound pricing and Ahrefs Brand Radar. Plan tiers, prompt counts and engine coverage, September 23, 2026.
• Canlah AI GEO services and facts page. On-site stage scope and measurement protocol.

See where your brand stands in AI answers. Run the free AI visibility audit →

Related articles

• SEO vs AEO vs GEO
• Best AEO agencies in 2026
• AEO pricing in Singapore