Why a question and its answer is the most liftable unit on a website, and how to mark it up properly.
FAQ schema for AI search is FAQPage structured data that pairs each visible question on a page with its answer, giving retrieval systems a machine-readable block that maps directly onto how people phrase prompts.
An FAQ section is the only part of most websites that is already written in the shape a retrieval system is looking for: a question, followed immediately by a complete answer, with nothing in between. That is not a coincidence you should waste. It is also, in most companies, the most neglected page on the site.
Because it removes every step a retrieval system would otherwise have to perform. Somebody types a question. Your page contains that question, verbatim or close to it, as a heading. Directly underneath is a short block that answers it and nothing else. The match is at the level of the whole unit, not a guess about which four sentences out of a 2,000-word essay constitute the answer.
Compare that with a narrative page on the same subject. The answer is in there somewhere, distributed across three paragraphs, prefaced by context and qualified afterwards. A system extracting from it has to decide where the answer starts and stops, and anything it lifts risks carrying a dependency on the sentence before. The FAQ version has no boundary problem. It is pre-cut, which is why the practical AEO guide treats question-shaped headings as a structural fix rather than a content one.
There is a second effect. FAQ answers are short by convention, which forces the writing discipline that makes content extractable in the first place. It is hard to hedge for four paragraphs inside an FAQ. The format itself does some of the work.
Yes, but for a different reason than it used to, and being clear about this matters because most advice on the web still assumes the old reason.
Google removed the FAQ rich result feature from Search: it announced the deprecation in May 2026 and removed the FAQ documentation from Search Central in June 2026. Google’s guidance is that existing markup can be left in place for other search engines and services that may use it, but it will no longer produce a rich result in Google Search. If you implemented FAQPage schema to win the expanded accordion in the results page, that reason is gone and it is not coming back.
What has not changed is that FAQPage is a schema.org type, not a Google feature, and the vocabulary is consumed by parties other than Google’s rich-result renderer. Marking up a pair states unambiguously that this is a question, this is its answer, and the two belong together — useful to anything parsing the page for structure rather than rendering a snippet.
Do not rebuild your strategy around the schema. The gain from FAQ markup is clarity at the margin. The gain from the visible FAQ — real questions, answered directly, on a page a retrieval agent can fetch — is the substantial one. If you have an hour, spend fifty minutes on the questions and ten on the JSON-LD. Doing it the other way round is the most common mistake on this page’s topic.
JSON-LD in the head or body of the page. Each entry is a Question with a name, containing an acceptedAnswer of type Answer with a text. That is the whole structure, and almost every implementation error is a deviation from those four properties.
<script type="application/ld+json">
{
"@context": "https://schema.org",
"@type": "FAQPage",
"mainEntity": [
{
"@type": "Question",
"name": "How long does an AEO audit take?",
"acceptedAnswer": {
"@type": "Answer",
"text": "A surface audit is returned within 48 hours. A full audit takes five working days, because it includes three runs of every prompt in the set."
}
},
{
"@type": "Question",
"name": "Do you work with companies outside the UK?",
"acceptedAnswer": {
"@type": "Answer",
"text": "Yes. We work with UK and US B2B companies. All work is delivered in English."
}
}
]
}
</script>Three details that are easy to get wrong. The name must be the question text, not a label or a slug. The text may contain limited HTML, but it must be escaped — an unescaped tag or quote is the most frequent cause of a block that fails to parse at all. And keep mainEntity as an array even with one question.
A visible FAQ is text a human reading the page can see and read. A schema-only FAQ is JSON-LD describing questions and answers that appear nowhere in the rendered page. The second is a mistake, and it is worth understanding why rather than just being told not to do it.
Structured data describes the page; it does not add to it. When markup asserts content the page does not contain, three things follow at once. It is a guideline violation for the search engines that police it. The machine-readable version and the human version now disagree, which is exactly the ambiguity structured data exists to remove. And it is fragile: edit the visible copy and the JSON-LD asserts something the company no longer believes.
The rule is simple and has no exceptions worth taking. Write the FAQ on the page. Then emit schema that describes exactly what you wrote. If the answer in your JSON-LD is not on the page in substantially the same words, delete one of them.
The ones your buyers actually ask, including the awkward ones. This is where most FAQ pages fail, and the failure is a choice rather than an oversight: somebody decided it was safer to answer questions that flatter the product.
| Softball — nobody asks this | Real objection — everybody asks this |
|---|---|
| “Why should I choose your solution?” | “How much does it cost, and what is not included?” |
| “Are you committed to quality?” | “What happens if it does not work?” |
| “What makes your team different?” | “Who should not buy this?” |
| “Do you offer a range of services?” | “How long before I see anything change?” |
| “Is your process proven?” | “Can I do this myself instead of hiring you?” |
The left column is unquotable for a structural reason, not an aesthetic one. Nobody types those sentences, so nothing matches them, and the answers underneath contain no checkable claim a system would be willing to repeat on your behalf. The right column is quotable because each question is one somebody genuinely asks and each answer forces you to state something specific.
The hardest of them is “who should not buy this”, and it is also the most valuable. An answer that names the buyer you are wrong for is a claim no competitor’s marketing copy contains, which makes it distinctive, and it is the kind of statement that gets repeated because it is unusually informative. It also does the filtering work you would otherwise do on a call.
Where do you get the list? In order of quality: your last twenty sales calls, your support inbox, and the questions buyers are putting to assistants about your category. The first beats any keyword tool, and it is sitting in your own notes.
Enough to cover the objections, few enough that each one is answered properly. In practice that lands somewhere between six and fifteen for a focused page, and the number matters far less than two other things.
The first is that every question must be distinct. Three questions that are rephrasings of each other dilute all three: the retrieval system now has three partial matches for one query instead of one strong one, and a reader gets the impression the page is padded. If two questions have the same answer, they are one question.
The second is that each answer should be two to four sentences. One usually means you dodged the question. Six means you have written a section, and it belongs on its own page with its own heading rather than in an accordion. Forty questions answered in a line each is a page that answers nothing.
Mostly spread across the site, next to the thing the question is about. The standalone FAQ page has its uses but it is over-relied upon.
| Question type | Where it belongs | Why |
|---|---|---|
| “How much does X cost?” | The pricing page | The page already ranks for the intent; the answer belongs where the buyer is |
| “How long does onboarding take?” | The service or product page | It is an objection to that specific offer, not a general query |
| “What is AEO?” | Its own explainer page | Too substantial for an accordion; deserves a full answer |
| “Do you work with US companies?” | The central FAQ page | Applies to everything, belongs to no single page |
| “Why did my rankings not change?” | The relevant guide | Answered in context, where the surrounding page supports it |
The pattern is: general questions about the company go on the FAQ page, and questions about a specific offer go on the page describing that offer. A question answered next to what it is about carries more weight, because the whole page corroborates the answer rather than the answer sitting alone among twenty unrelated ones. The AEO checklist treats this as a placement item rather than a content one, which is the right way round.
Ranked by how often they turn up in a real audit rather than by how serious they sound.
text breaks the JSON, and a block that does not parse contributes nothing at all. This is the single most common failure and it is invisible unless you validate.acceptedAnswer. A Question with a name and no answer object is a question the parser cannot resolve. It happens when an answer field is empty in the CMS.QAPage is for a page with one user-submitted question and community answers, such as a forum thread. FAQPage is for questions the site owner asks and answers. They are not interchangeable and the distinction is checked.Validate before you ship. Paste the rendered page into a structured data validator, fix what it flags, and re-check after any CMS change that touches the FAQ fields. It is a five-minute step that catches the first and third items on that list outright.
By asking the questions and recording what comes back, repeatedly rather than once. Assistant output is non-deterministic: the same question asked twice can return different sources, so a single check tells you almost nothing.
Take the questions you put on the page, ask each one several times in the assistants your buyers use, and record how often your page is among the sources rather than whether it was. Appearing in one run out of three and appearing in three out of three are different results and a single observation cannot distinguish them. Keep the question list frozen so that next month’s run is comparable to this month’s, and log the runs rather than the impressions of them.
Do not expect a fast signal. Pages have to be re-crawled and re-encountered before anything shifts, and no page structure guarantees a citation from any assistant. Whether an agent can fetch the page at all is the gate above this one, and it is item one on the checklist; where FAQ work sits relative to ranking work is covered in AEO versus SEO. If you want the questions, the markup and the crawler access checked in one pass, that is what the free AI visibility check covers, and the guide to getting cited by ChatGPT sets out the other three gates the FAQ work sits inside.
It is worth adding, but not for rich results. Google announced the deprecation of the FAQ rich result in May 2026 and removed the documentation in June 2026, advising that existing markup can be left for other search engines and services. The value now is machine-readable structure, not an expanded snippet.
No. Structured data is meant to describe content that is on the page. Schema-only FAQs violate the guidelines that search engines enforce, and they create a page whose machine-readable version disagrees with what a human sees, which is the opposite of what markup is for.
Usually six to fifteen on a focused page, but the count matters less than distinctness. Every question should be one somebody genuinely asks, and no two questions should share an answer. If two entries would be answered the same way, they are one question.
FAQPage is for questions the site owner both asks and answers. QAPage is for a page built around a single user-submitted question with community answers, such as a forum thread. Using QAPage for a standard marketing FAQ is a type error and will be flagged.
Mostly throughout the site, next to what the question is about. Pricing questions belong on the pricing page and onboarding questions on the service page, because the surrounding content corroborates the answer. Keep the central FAQ page for questions that apply to the whole company.
It improves the odds because question-and-answer pairs are the most liftable unit on a site, but nothing guarantees a citation. Assistant output is non-deterministic and retrieval also depends on crawler access and on how clearly your company resolves as an entity.