Back

How to Write FAQ Content an AI Answer Will Actually Quote

Tengence Team ·
ai-search GEO/SEO Schema Markup
How to Write FAQ Content an AI Answer Will Actually Quote

Summary: The value of an FAQ is not whether it earns a rich result but whether it is a passage that can be quoted directly. After Google narrowed FAQ rich results in 2023, many teams concluded FAQs were dead — the opposite is true. Generative engines prefer exactly this shape: one question, one short conclusion, clear boundaries. The difference is that an FAQ written for rich results and an FAQ written for extraction are not the same document.

Key Takeaways

1. Separate two jobs: rich results and answer extraction

FAQs serve two purposes and teams routinely conflate them, ending up with neither.

Rich results. Q&A markup produced an expandable panel in search. In August 2023 Google narrowed this: only authoritative government and health sites would routinely show FAQ rich results, and other sites would stop seeing them reliably even with correct markup. By 2024 the display had largely vanished for ordinary commercial sites.

Answer extraction. Generative engines assemble answers by finding passages that directly address a sub-question. This purpose was untouched by the rich-result change; if anything, the clear Q&A shape and conclusion-first phrasing make these passages the easiest to extract on a page.

So the correct read is: if your goal is a rich result, ROI is already low; if your goal is being cited when someone asks that kind of question, FAQ remains the most economical structure. Everything below targets the second goal.

Want to know how much of the ChatGPT and Perplexity answers your brand owns?

Explore TENGENCE GEO

2. Where questions come from: three real sources

Inventing questions is the most common self-deception. A writer sits in an office imagining "what users might ask", and the output sits one layer away from reality — real questions are shorter, more colloquial, more specific, and often carry a scenario.

Our question bank draws from three sources:

  1. Support tickets and after-sales records. Real wording and real emotion; highest conversion value.
  2. Internal search queries, especially zero-result queries. What users failed to find on your own site is usually the content gap.
  3. Questions sales and pre-sales answer repeatedly. Highest relevance to closing, and usually absent from tickets.

Together these cover most high-frequency asks. Organise them in one table: exact question, frequency, channel, whether an existing page answers it, and whether that answer is correct. The last two columns matter most — many questions are already answered somewhere, just scattered enough that a machine cannot extract the answer.

One category deserves separate treatment: comparison questions. When a user asks "which suits us better, A or B", an answer that only praises A cannot serve a comparison query. The extractable form is a decision table — which scale, which budget, which scenario picks which option, one row each. Such answers get cited far more often than "our solution is superior".

Source Typical ask Where it belongs Common flaw
Support ticket "Why do I have to log in again after switching devices?" Product help centre Too much jargon; the user described a symptom
Internal search "how much", "how to refund" Pricing page / after-sales page Too short, no scenario qualifier
Sales conversation "why is it more expensive than the competitor" Comparison page Written as marketing copy with no checkable number

3. Writing answers that survive extraction

Extraction slices; it does not read through. That yields three hard rules.

Rule one: conclusion first, within 30 to 60 words. Generative answers prefer passages whose first sentence answers the question. Open with two sentences of background and your passage is far more likely to be cut at the wrong point. Before: "As enterprise digital transformation deepens, data governance has become increasingly important. Regarding data retention periods, typically…" After: "Data is retained for 12 months by default, adjustable to 36 months in the console; anything longer requires compliance approval."

Rule two: the answer must stand alone. No dangling references. If a slice contains "this solution" or "the configuration above", the machine cannot resolve it and skips the passage.

Rule three: one question, one answer. A heading carrying three questions structures none of them. Split into three FAQs rather than merging.

Two empirical values: 60 to 120 words per answer, because past 150 a passage reads as an article rather than an answer; and every number carries its unit and basis, such as "12 months (calendar months, counted from activation)".

One more writing habit that gets skipped: append an applicability clause. "Refunds within 7 days (annual plans only, unactivated)". The purpose is not legal cover but preventing a truncated citation from misleading — a large share of complaints come from quotes that dropped the qualifier. Likewise, if an answer holds only in a specific region or version, put the condition inside that answer rather than in a footer disclaimer.

A final detail is the collapsed panel. Many help centres collapse every answer, so the rendered body contains only questions. That looks normal to a human and is hostile to extraction. The fix is not to remove collapsing but to ensure collapsed content exists in the initial HTML response: collapsed is a visual state, not missing content.

4. Should you still write markup?

Yes, but be clear about what it does and does not do.

What it does: tells a machine explicitly that this block is Q&A, this heading is the question, this paragraph is the answer. That genuinely helps extraction, because clearly bounded content slices correctly.

What it does not do: produce a reliable rich result. Since 2023, ordinary commercial sites rarely see an FAQ panel even with flawless markup.

So reposition markup from "competing for display" to "marking boundaries". Three details get missed most often:

The comparison method is simple: publish the same Q&A set as plain text first and record citation behaviour over a period; add the markup and, with prompts unchanged, observe for an equal period. One variable changed is the only way to attribute a result — if you also edited titles or structure in the same window, you have learned nothing.

5. Five failure modes and the update loop

Across projects we have reviewed, FAQs fail in five ways:

  1. The questions are not real. Invented, keyword-stuffed, phrased in a way nobody searches.
  2. The answers are too long. A 300-word answer is judged a body paragraph.
  3. Answers contradict each other. The same question has one version in help centre, another on the product page, a third in a blog post, with different dates and numbers.
  4. Markup does not match content. Markup declares question A while the page shows question B, or the answer field holds something else.
  5. The page is not indexable. Panels collapsed by default, noindex, or gated behind login.

The fifth is the most wasteful: perfect content that a machine cannot see is worth zero. Technically, collapsed content must be present in the first response and not depend on client-side async loading.

Update quarterly: review the quarter's new internal search queries and high-frequency support questions and add them; merge entries now fully covered by body copy; re-check every number that ages (price, deadline, version). The routine is simple and most teams have never run it once.

One benchmark: if your FAQ page has not changed a single word in a year, it is no longer an answer bank but a legacy section. Real questions move — product renames, price changes, policy updates and competitor shifts all generate new phrasings. We put FAQ maintenance on the quarterly content review agenda rather than in someone's memory.

FAQ

Q: Should FAQs sit at the end of an article or on a dedicated page?
Depends on the job. Detail questions raised by that article belong at its end; general product or service questions deserve their own page, because a dedicated page is a more stable citation source.

Q: How long should a single FAQ answer be?
60 to 120 words. When more explanation is needed, split into a short answer plus a longer explanation, with the short answer first.

Q: Is FAQPage markup still usable?
Yes, but accept that it no longer produces reliable rich results. Treat it as boundary markup rather than a display mechanism.

Q: Should multilingual FAQs be independent of each other?
Yes. Each language version should be rewritten around real questions in that language rather than translated sentence by sentence; translated phrasings are usually not what local users type.

Q: Do FAQs dilute the main keyword?
That is rarely the real problem. The actual risk is topic drift — ten unrelated questions on one page leave a machine unable to say what the page covers.

Related Reading

Data Sources

  1. FAQPage — schema.org type definition — the current Q&A vocabulary and its intended semantics.
  2. Changes to HowTo and FAQ rich results — Google Search Central, August 2023 — the official narrowing announcement and its scope.
  3. Search Quality Rater Guidelines — Google — how answer quality and evidence are assessed.
← Previous: SaaS Content Marketing: How to Build a Blog That Actually Acquires Customers Next: AI Search Result Page Optimization: Treat It as Four Modules, Not One Ranking Slot →
Consult
QR code

WeCom