I pulled our own Search Console data this week the way I do most weeks, and one pattern kept repeating: our FAQ-cluster pages get clicked, our long-form explainer pages mostly don't. Same site, same author, same publishing cadence. One of our FAQ pages sits at a healthy click rate while a diagram-heavy explainer nearby pulled in over 950 impressions and a single click. The gap wasn't topic quality. It was that one page had answers a model could lift whole, and the other one made a reader — or a crawler — work for the point.
That's not a hunch. It's the same mechanism Google documents for its own FAQ rich results and the same extraction behavior AI answer engines describe when asked how they select quotable spans [1]. A FAQ block is not a compliance checkbox you bolt onto the bottom of a page because a template says so. It is, sentence for sentence, some of the highest-leverage copy on your site — if you write it like something meant to be extracted, not summarized.
FAQ answers get lifted verbatim by AI engines when they are self-contained, 40-60 words, lead with the direct answer in the first sentence, and are backed by FAQPage schema that matches the visible text exactly. Answers that bury the point after a setup, or that need the surrounding paragraph for context, get paraphrased or skipped instead of quoted.
Why do most FAQ blocks get ignored by AI engines?
Because most of them are written for humans skimming, not for a model extracting a self-contained span. A typical FAQ answer opens with a throat-clear — "Great question, this depends on a few factors" — before it says anything. Extraction models don't reward patience. They pull the passage that answers the question most directly, and if your first sentence is a preamble, the model either paraphrases around it or moves to a competitor's page that answered faster.
The second failure is length drift. Answers that run 150+ words read fine to a human scanning past them, but they stop being a single quotable unit — they become a paragraph with a quotable sentence buried somewhere inside it. The model has to decide which part to lift, and when it has to decide, it usually just doesn't bother lifting your exact words. It restates the idea in its own language instead, which means you lose the citation and, often, the click.
What makes an answer extractable instead of just readable?
Extractable answers pass a specific, mechanical test: pull the answer out of the page, paste it somewhere blank with no surrounding context, and check whether it still fully answers the question a real person typed. If it does, a model can lift it whole. If it needs the sentence before or after it to make sense, it isn't extractable yet, no matter how accurate it is.
Three things make an answer pass that test consistently:
- The claim comes first. Sentence one states the answer. Everything after that is supporting detail, not the answer itself.
- It carries its own subject. Don't write "This depends on your plan" — write "A Starter plan includes X" so the sentence still makes sense with zero surrounding text.
- It's bounded. One question, one answer, one idea. Two questions crammed into one FAQ entry force the model to choose which half to quote, and it often quotes neither cleanly.
"Pricing for this kind of service can vary quite a bit depending on scope, and there are a lot of factors that go into it, but generally speaking most providers in this space tend to charge somewhere in a mid-range bracket once you account for everything."
"Most GEO monitoring tools cost $20-$95/month; done-for-you agencies typically start around $1,500/month per brand. The tool tells you the gap, the agency closes it."
How does FAQPage schema reinforce the visible answer?
Schema doesn't create the answer — it labels it. FAQPage structured data, marked up as Question and acceptedAnswer pairs in JSON-LD, tells a crawler explicitly which text on the page is a question and which exact span is its answer, removing the guesswork a model would otherwise do from layout alone [2]. Google's own guidance is specific here: the schema's answer text must match what a user actually sees on the page, word for word, or the markup is treated as unreliable and can be ignored entirely [3].
That single rule — visible text and schema text must match exactly — is where most implementations quietly break. A developer ships the FAQPage schema once at launch, a copywriter edits the visible answer three months later for tone, and now the two have drifted. The schema still validates. It just isn't true anymore, and a model that samples both and finds a mismatch has a real reason to trust neither.
Where should the FAQ block live on the page?
Below the core content, not buried under three unrelated sections and a newsletter form. Crawlers and extraction passes weight earlier, more prominent content slightly higher, so a FAQ block that a reader has to scroll past a gallery, a testimonial carousel, and a related-posts grid to find is a FAQ block a model may never fully process either [4]. Put it where a real visitor would actually look for it: right after the section that raises the objection the FAQ answers.
Order matters inside the block too. Lead with the question your actual search and support data show people ask most — not the question that's easiest to answer, or the one that makes your product look best. If you don't already know which questions people ask, your own sales calls, support tickets, and site-search logs are a better source than a keyword tool guess.
How many FAQ questions is enough, and when is it too many?
Four to eight focused questions beats fifteen shallow ones. Every extra question in the block is fine structurally, but each one has to independently pass the extraction test — self-contained, answer-first, bounded — or it's just diluting the ones that do work. A FAQ block padded out to hit a "content depth" quota with restated or overlapping questions doesn't add citation surface area; it adds noise the model has to sort through to find the two or three answers actually worth lifting.
A useful check before publishing: read only the bolded questions down the page, ignoring the answers. If two of them could plausibly be answered the same way, merge them. If a question wouldn't survive being typed into a search bar or an AI chat box by an actual buyer, cut it.
What breaks the pattern most often, even on sites that know better?
Schema-answer drift is the most common failure we find when auditing sites that already believe they've "done" FAQ schema. The JSON-LD was written once, correctly, at launch. The visible copy has been edited since — a price changed, a policy softened, a sentence got friendlier — and nobody updated the markup. The schema still validates against Schema.org's spec, so nothing throws an error. It's just no longer true, and Google explicitly treats mismatched or manipulative structured data as invalid [3]. The second most common failure is duplicate FAQ blocks reused verbatim across dozens of pages on the same site — a pattern crawler documentation and site-quality guidance both flag as a low-value, near-duplicate signal [5]. If your FAQ answers are identical on every service page, they're not helping any one of them get cited.
Questions people ask
Roughly 40 to 60 words, front-loaded with the direct answer in the first sentence. Answers under 25 words are often too thin to stand alone; answers over 90 words usually bury the extractable claim under qualifiers, so the model paraphrases instead of quoting.
Yes, moderately. Engines weight earlier content on a page slightly more during extraction passes, and a FAQ block placed after the core content (not buried at the very bottom below unrelated sections) tends to get parsed more reliably. Order questions from the highest-intent, most-searched phrasing down.
Not every answer, but the ones you most want cited should. A specific figure, date, or named source gives the model something concrete to attribute back to you, which is what turns a paraphrase into a verbatim quote with your brand attached.
Yes. Both systems favor the same underlying pattern: a plain-language question matched to a self-contained, factual answer with FAQPage schema reinforcing the pairing. There is no separate format to write for each engine; write one clean answer and both can lift it.
Sources cited
Want your FAQs audited, not guessed at?
We'll show you exactly which of your existing FAQ answers a model can lift today, and rewrite the ones that can't.
Get a free AI Visibility Audit →