FAQ markup is the only place AI quotes you word for word
Updated:
Quick answer: an answer inside FAQPage markup is already a complete, labelled, self-contained unit, which is exactly the shape a retrieval system is trying to build out of your page anyway. Everywhere else the machine has to guess where an answer starts and stops. In FAQ markup you hand it the boundaries, and the wording that comes back in the answer is often yours, unchanged.
Why does this one format behave differently?
Every other part of your page has to be reconstructed before it can be used.
A retrieval system takes your HTML, strips it to text, splits the text into fragments, and hopes the splits land in sensible places. It does not know where your answer to a particular question begins. It infers that from headings, paragraph breaks and position, and it is often wrong.
FAQPage markup removes the inference. Each entry declares a question and an answer, and the boundary between them is explicit rather than guessed. The answer is not a fragment that happens to contain an answer. It is an answer, labelled as such, by you.
That is the whole mechanism. It is unglamorous and it explains why the same sentence performs differently depending on where it sits.
- FAQPage
- A schema.org type that labels question and answer pairs on a page as such.
- Verbatim quoting
- Reproducing a passage word for word instead of paraphrasing it.
Which shape works?
<div itemscope itemtype="https://schema.org/FAQPage">
<div itemscope itemprop="mainEntity" itemtype="https://schema.org/Question">
<h3 itemprop="name">How long does a technical audit take?</h3>
<div itemscope itemprop="acceptedAnswer" itemtype="https://schema.org/Answer">
<div itemprop="text">
Between three and five working days for a site under fifty pages.
Larger sites take longer because crawl time scales with page count,
not with complexity.
</div>
</div>
</div>
</div>JSON-LD in a <script type="application/ld+json"> block does the same job and is easier to maintain. Either is fine. What matters is that the question and answer are declared, not implied.
Three rules that decide whether it works:
The visible text must match the markup. Markup that says one thing while the page shows another is the oldest schema abuse there is, it is detected, and it costs more than it gains.
Answers should run 40 to 80 words. Short enough to be lifted whole into a response, long enough to carry a complete thought. Below forty the answer usually needs context it does not have. Above eighty it gets truncated, and truncation happens wherever the system decides, not where you would have chosen.
Questions should be phrased the way people ask them. "Pricing" is a label. "How much does an audit cost?" is a question. The second matches how queries actually arrive.
What does our engine look for?
We score two related things separately, because they fail independently.
struct.faq-schema checks whether valid FAQPage markup exists and whether its entries correspond to visible content on the page.
cite.faq-block checks whether there is a real FAQ section at all, with five to eight question-and-answer pairs, regardless of markup. A page can have the section without the schema, which is common, or the schema without a proper section, which is worse.
The second is the one that carries more weight. Markup around nothing is nothing. A genuine FAQ section without markup still gets extracted, just less reliably.
Which mistake makes FAQ blocks useless?
Most FAQ sections on business sites answer questions nobody asked.
"Why choose us?" is not a question. It is a sales heading wearing a question mark, and the answer under it is always a paragraph of adjectives. Nobody types that into anything, and there is no query for a model to match it against.
The questions worth including are the ones you actually get. Not the ones marketing wishes people asked, but the ones that arrive by email, in the first five minutes of a call, in the message before someone books.
A test that works: read your FAQ and ask, for each entry, whether a stranger would ever type that sentence into a search box. Entries that fail this test are taking up space that a real question could occupy.
The second test: does the answer contain a fact? "Yes, we work with clients across Europe" is a fact. "We pride ourselves on flexible service" is not. An answer with no fact in it gives a model nothing to relay, and relaying is the entire mechanism.
Before they decide, clients ask the same six questions: what it costs, how long until results, whether we guarantee positions, how we differ from the rest, how much of their own time it takes, and what happens if we stop. An FAQ block built from invented questions instead of these answers nobody.
Where does FAQ markup not help?
Worth being direct about the limits, because schema is oversold routinely.
It does not make an unknown brand known. FAQ markup improves extraction. It does nothing for whether a model has a prior about your company. Those are separate problems, and no amount of markup touches the second one.
It does not rescue thin content. A page with nothing on it but a FAQ block is a page with nothing on it. The markup makes the emptiness easier to parse.
It is not a ranking mechanism. Google's treatment of FAQ rich results has changed more than once, and the visible snippet behaviour is not the reason to do this. The reason is retrieval, which is a different system with different rules.
It can be over-applied. Fifteen FAQ entries on a short page is the pile-up pattern that gets discounted. Five to eight real questions beats twenty invented ones.
Why do the answers get quoted rather than paraphrased?
This is the part that surprises people who have watched their content get rewritten in AI answers.
When a model uses a fragment it is unsure about, it paraphrases. Paraphrasing is a hedge: it lets the model smooth over ambiguity, merge two sources, and avoid committing to wording it cannot verify. Your phrasing disappears, and often your attribution with it.
When a model uses a fragment that arrives pre-labelled as a complete answer to a specific question, there is less to hedge. The unit is already the right size and shape. Reproducing it is the low-risk operation.
This has a consequence worth planning for: the wording in your FAQ is the wording that may end up in front of someone who never visits your site. That is the only place on most pages where you get to choose the exact sentence a machine repeats.
Write those sentences accordingly. Put the number in. Name the constraint. Say the thing you would want said if you were not in the room.
What did we find on our own site?
Our five scanned pages on 27 August 2026 have no FAQ section and no FAQPage markup anywhere. Both checks fail across the board.
That is not an oversight so much as a habit. The pages describe services in continuous prose, because that is how service pages have been written for twenty years. The format predates the reason to change it.
The fix is cheap and specific: take the six questions that arrive most often by email, answer each in fifty words with a fact in it, mark it up. It is perhaps an hour of work per page, and it is the only hour on the list where the output is text a model may repeat verbatim.
We have not done it yet at the time of writing, which is worth stating rather than implying otherwise.
Where do you get the six questions?
You already have them, written down, in places you do not think of as content.
Your sent mail. Search your outbox for question marks from the last three months. The questions clients asked before buying are the questions strangers are asking a model right now.
The first five minutes of any call. Whatever gets asked before the conversation becomes specific to that client is the general question. It repeats because it is general.
Your own internal confusions. If two people in your company explain something differently, customers are confused about it too, and nobody has written the answer down.
Objections. "Isn't this expensive" and "how is this different from X" are questions with question marks removed. Put them back and answer them honestly. An FAQ that engages a real objection is more convincing than one that avoids it, and it matches queries that actually get typed.
What you refuse. "Do you work with clients outside Europe?" answered with a clear no and a reason is a better entry than a vague yes. Specificity includes limits.
Six questions from those five sources takes an afternoon at most, and the answers are already in your head. The writing is transcription, not invention.
Frequently asked questions
Does JSON-LD work as well as inline microdata? Yes, and it is easier to maintain because the markup lives in one block instead of being tangled through your HTML. Keep the visible text and the JSON in sync.
How many entries should a FAQ section have? Five to eight real questions. More than that on a single page starts to look like padding, and padding is a pattern that gets discounted.
Can I reuse the same FAQ across many pages? Poorly. Duplicate blocks across a site add little and can read as templating. Each page should answer the questions specific to what it covers.
Should answers link to fuller explanations? Yes, when there is genuinely more to say. The answer itself should still be complete without the click, because the click is exactly what does not happen in an AI answer.
Does FAQ markup help with voice assistants? It is the same mechanism: a labelled, short, self-contained answer is the easiest thing to read aloud. Whether any particular assistant uses your markup is outside what we can verify.
What if my answers are longer than eighty words? Split the question. An answer that needs two hundred words is usually two questions that have not been separated yet.
The short version
Almost everything on your page gets reconstructed by a machine before it can be used, and reconstruction loses your wording. FAQ markup is the exception: you declare the boundaries yourself, and what comes back is often exactly what you wrote.
Which means the answers in that block are the only sentences on your site you can write knowing a machine may repeat them unchanged. Six real questions, fifty words each, a fact in every one. It is an hour of work and it is the cheapest thing on the list.
Sources: schema.org FAQPage specification; research digest compiled from 165 sources on AI search (August 2026); GeoScan v1.2 variable definitions; field scans of seo7.es, 27 August 2026. Details in sources/research-notes.md.