Most FAQ sections are written for a scanner, not a reader. Short, keyword-stuffed headers stacked over answers that assume you already read the page above them. That format worked fine when the goal was a featured snippet. It does not work when the reader is a language model deciding whether your sentence is complete enough to quote.
When ChatGPT, Perplexity, Gemini, Claude, or Copilot answer a question, they are not rendering your page and reading it in order. They are pulling a chunk of text that stands on its own, matches the question closely enough to be relevant, and does not require the rest of the page for context. FAQ sections are one of the highest-leverage places to give them exactly that. Get the format right and you hand the model a ready-made citation. Get it wrong and your FAQ becomes decoration nobody extracts.
This is a practical guide to writing FAQ content that survives the trip from your page into an AI answer: how to phrase the questions, how to write answers that stand alone, how to mark the whole thing up with FAQPage schema, and where most FAQ sections quietly fail.
Google retired FAQ rich results from search in 2026, which led some teams to conclude FAQ schema had become pointless. That conclusion mixes up two separate things: whether a feature earns a visual result in Google Search, and whether structured question-and-answer content is useful to a system trying to extract facts. FAQPage remains a valid, documented schema.org type, and AI assistants that crawl and parse the web still read it as a strong signal about what a page is answering.
Think about it from the model's side. An AI assistant answering a user's question is doing retrieval under time pressure: find a passage that answers this, verify it is trustworthy enough, cite it. A well-written FAQ entry is the easiest possible passage to retrieve, because you have already done the work of isolating one question and one complete answer. A page of flowing prose forces the model to do that isolation itself, and it will not always get the boundaries right.
We covered the broader mechanics of marking up a page for machine readers in our post on structured data for LLMs. FAQ content is the single most common place teams get that structure wrong, because it looks simple and usually is not treated with much care.
The biggest gap between a human-facing FAQ and a model-facing FAQ is the question itself. Most FAQ headers are written like search keywords: "Pricing," "Refund policy," "Integrations." Nobody types those phrases into ChatGPT. They type full questions, usually with context baked in.
Compare the two styles directly:
The second version in each pair is closer to what a person actually asks a chat assistant, and it is closer to what the assistant's own retrieval process is trying to match. If your question and a user's prompt share enough phrasing and intent, your answer has a real shot at being the one pulled forward.
A fast way to source real phrasing: pull actual questions from support tickets, sales call transcripts, and onboarding chats. Those are the words your buyers use, not the words your content team would choose. If you track which prompts already surface your brand or a competitor's, that same prompt list is a direct source of FAQ questions worth answering on your own site.
Self-contained means the answer makes complete sense if it is the only sentence a reader ever sees, with no page title, no surrounding paragraph, and no "as mentioned above" to lean on. This is the single most common failure mode in FAQ writing, and it is almost always invisible to the person who wrote it, because they wrote it with the rest of the page fresh in their head.
A few concrete rules get you most of the way there:
Google's own structured data guidance for FAQ markup has, historically, required exactly this: answers written as complete, standalone text rather than a pointer to context elsewhere on the page. That requirement predates the current wave of AI answer engines, but the underlying reason hasn't changed. Whether the reader is a crawler, a rich result, or a language model, an answer that depends on surrounding context is an answer that gets mangled or skipped when it's lifted out of the page.
Here is what this looks like applied to three real questions, written the way we write our own FAQ entries at Bold GEO.
Q: Does Bold GEO track citations across ChatGPT, Perplexity, Gemini, Claude, and Copilot?
A: Yes. Bold GEO runs a daily refresh across ChatGPT, Perplexity, Gemini, Claude, and Copilot, checking whether your brand appears in answers to the prompts your buyers actually ask. Results are grouped by assistant so you can see where you are cited and where a competitor is showing up instead.
Q: How much does Bold GEO cost?
A: Bold GEO offers pay-as-you-go scans at $1 each, plus monthly plans at $49, $99, and $249 that include a scheduled refresh volume and prompt tracking. Every plan starts with a 7-day free trial and no credit card is required to begin.
Q: Do I need FAQPage schema for AI assistants to cite my FAQ content?
A: Schema is not strictly required, but it removes ambiguity. FAQPage markup tells any system parsing your HTML exactly which text is the question and which text is the answer, so the pairing survives even when a page is scraped for a snippet rather than rendered and read top to bottom.
Notice each answer names its subject, resolves on its own, and fits in two to three sentences. None of them say "as we mentioned" or lean on a heading above them to make sense. Read any one of them in isolation and it still works, which is the actual test.
Once the question-and-answer text itself is solid, wrap it in FAQPage structured data so the pairing is machine-readable rather than inferred from HTML layout. The pattern is a JSON-LD block with one Question entry per FAQ item, each containing an acceptedAnswer of type Answer:
{ "@type": "Question", "name": "Your question, phrased the way a user would ask it", "acceptedAnswer": { "@type": "Answer", "text": "Your complete, self-contained answer." } }
A few practical notes on getting this right:
name field should exactly match the visible question text on the page. Don't optimize the schema question differently than the on-page question; that mismatch reads as manipulative to anything checking your markup against your content.text field should exactly match the visible answer, not a longer or shorter version written just for schema.If you haven't set up structured data on the rest of your site yet, our guide to structured data for LLMs covers the other schema types worth prioritizing alongside FAQPage, including Organization and Article markup that gives models basic facts about who you are.
Most FAQ sections we review fall into one of three traps. All three are fixable without a redesign.
"Why should I choose us over competitors?" answered with three sentences of adjectives is not an answer, it's a pitch. A model looking for a factual response to "does this tool support X" will pass over marketing copy in favor of a source that states a fact plainly. Save the persuasion for your homepage. In an FAQ, state the fact, then stop.
"It depends on your plan" is technically true and completely useless as an extracted answer. If the real answer varies, say how it varies, in the same sentence. "Storage limits range from 10GB on the free plan to 500GB on the Business plan" is extractable. "It depends on your plan" is not, because it doesn't actually answer anything on its own.
If the true answer to a question requires reading three separate parts of your page, your FAQ entry is misrepresenting itself as complete when it isn't. Either write a genuine summary answer in the FAQ and link out for depth, or don't include that question in the FAQ section at all. A partial answer that looks complete is worse than no answer, because it risks getting cited with missing context.
FAQ sections are one input into a page's overall trustworthiness to an AI system, not the whole picture. The same self-contained, fact-first writing standard applies to your About page, your pricing page, and your product pages. We wrote about applying this to the page that usually gets the least editorial attention in our guide to optimizing your About page for LLMs, which is worth pairing with FAQ work since both pages tend to be where models look for basic facts about a company.
The honest way to know whether any of this is working is to check whether it changes what AI assistants actually say about you. Rewriting a FAQ section is a bet, not a guarantee, and the only way to collect on that bet is to watch your citations before and after the change. That's the gap Bold GEO is built to close: a daily refresh across ChatGPT, Perplexity, Gemini, Claude, and Copilot that shows whether your brand is cited, how often, and against which competitors.
Start with the FAQ questions your buyers actually ask, written as complete answers a model could quote without you in the room. Then measure whether that work moved the needle.
Bold GEO monitors how your brand is cited across ChatGPT, Perplexity, Gemini, Claude, and Copilot on a daily refresh. 7-day free trial, no credit card.