A good FAQ is the easiest thing to serve a generative model: a question someone would really ask, followed by a closed answer you can copy without touching it. The trouble is most FAQs don’t answer what the buyer asks —they answer what the company feels like telling—, and that’s why AI ignores them. This guide isn’t about markup: the FAQPage schema that declares those blocks is plumbing, covered elsewhere. It’s about the strategy of the questions: how to pick the ones your buyer actually asks and answer them so AI copies them. The general shape of citable text is in content structure to get cited by AI; here we drop down to the question-answer pair and which questions deserve one.
What the questions and answers that get you cited by AI are (and why it copies them verbatim)
The questions and answers that get you cited by AI are blocks where the heading is a real buyer doubt and the direct, closed, self-contained answer sits below it. They’re the format a generative model copies with the least risk: the buyer talks to AI in questions, and the model matches that question with the block that most resembles it and pastes the answer below. A FAQ isn’t a drawer of frequent doubts at the foot of the page; it’s a battery of citable fragments, each matched to a query someone actually types.
That’s why the question-answer block wins so many citations: it comes with the question already attached. You don’t force the model to deduce which doubt your paragraph solves; you tell it in the heading, in the user’s words, and hand it the answer packaged. What decides whether it cites you isn’t the markup —that’s technical plumbing—: it’s whether the question is the one the buyer asks and the answer stands on its own.
The real mistake: you answer the questions you like, not the buyer’s
The average FAQ is written from the inside: «Why choose us?», «What makes us different?», «What are our values?». They’re questions the company loves answering and no buyer ever types into a chat. AI doesn’t retrieve them because nobody asks them; and even if someone did, the answer is usually self-praise with no data, exactly what the model discards.
The buyer asks about their problem, not your brand. They ask «how much does X cost?», «does X work for Y?», «what’s the difference between X and Z?», «how do you do X without Y?». Those are the questions AI receives and tries to answer. If your FAQ doesn’t contain them, you have no block to match, however well marked up it is.
How to find the buyer’s real questions
Good questions aren’t invented in a meeting: they’re collected from where the buyer already asks them. Your job isn’t to imagine doubts, it’s to mine the ones that exist and put them on your page in the exact words they’re phrased.
- What people actually ask you. Support email, the sales chat, pre-sale calls: that’s the gold. The doubts that repeat before a purchase are, literally, the questions AI also receives.
- Autocomplete and «people also ask». Google shows you for free how people finish their questions about your category. Each suggestion is a real question, written the way the user writes it.
- Forums and communities. Reddit, Quora and industry forums are full of the raw question, unfiltered by marketing. It’s where the model drinks too.
- Ask the AI itself. «What do people usually ask before hiring X?» hands you a map of doubts the model already associates with your category.
Not all of them count. The filter is simple: would someone about to buy, who doesn’t know you yet, ask it? If yes, it’s in. Curiosity questions («who invented X?») inflate the count and bring no buyer. Prioritize intent ones —price, fit, comparison, how to start—: that’s where the buyer decides, and where a citation of yours is worth something.
How to write the answer AI copies: two or three sentences, done
The citable answer has a concrete shape: the full claim in the first sentence, the nuance or the data point in the second, and stop. Two or three sentences that stand on their own off the page. If your answer opens with «it depends on many factors» and develops over a ten-line paragraph, the model has nothing to copy: it has to summarize you, and to summarize it prefers someone else.
- The answer first, whole. «X costs from €79 a month» before «our pricing adapts to each client». The concrete figure is what gets copied; the dodge doesn’t get cited.
- The nuance after, not before. If there’s an «except when», it goes in the second sentence, not as an excuse to dodge the answer in the first.
- Don’t open with the brand. «At [company] we believe that…» chains the answer to you and makes it useless off your page. Start with the answer, not your name.
- The figure with its source, if you have it. A number of your own with its unit turns your answer into a primary source. And if you don’t have the real figure, describe the mechanism —never invent one to fill space, because when the model corrects it against another source it drops you from the answer for good—.
| Answer that doesn’t get cited | Answer that gets copied |
|---|---|
| «The price depends on your needs. Contact us for a custom quote.» | «The service starts at €79/mo and includes X. From there it scales with your volume of Y.» |
| «At our company we’ve been helping clients with this for years.» | «Yes, X works for Y as long as you have Z. If not, you need W first.» |
The right-hand column can be pasted into an answer without a retouch. The left one forces the model to keep looking —and it finds whoever actually answered—.
How many questions, in what order, and where to put them
A FAQ that gets cited isn’t a wall of forty questions: it’s a well-chosen handful of the ones that actually decide the purchase. More isn’t better; relevant is better. And its spot isn’t only the foot of the page.
- Six to ten intent questions per page, not forty of filler. Every extra question nobody asks dilutes the ones that matter. Quality of doubt over quantity.
- Buyer ones up top. Price, fit and comparison first; the operational or detail ones after. Order is also a signal of what you prioritize.
- Inside the body, not only at the end. The question-answer pair works the same as a section inside the article as it does as a FAQ block at the foot. Put the doubts where the text prompts them.
And a golden rule: one question per block, one answer per block. If you pack three doubts into one answer, the model can’t extract any cleanly. Six well-answered pairs are six citation shots; a paragraph with six doubts mixed together is zero.
Mistakes that turn your FAQ into theater
A badly planned FAQ isn’t neutral: it wastes your time and tells the model you’re not the source. These are the failures that empty it of citations.
- Questions nobody asks. «Why are we different?» is self-praise disguised as a doubt. Out.
- Answers that don’t answer. «It depends», «contact us», «every case is unique»: the model can’t copy a dodge.
- Markup with no content behind it. Building the FAQPage schema over empty answers is plumbing wrapping theater: flawless outside, useless inside.
- Copying the competitor’s FAQ. If you repeat their questions and answers, you add nothing new and the model already has that source. Information gain —saying something the other sources don’t— is what makes you citable.
If you’d rather this work got executed on your money pages —mining your buyer’s real questions and rewriting each answer so AI copies it— GEO optimization does the labor, question by question, and GEO monitoring measures whether you start showing up when the buyer asks.