In August 2023, Google quietly dialed back FAQ rich results. The expanded accordion that once dominated the SERP for how-to and product queries collapsed almost overnight to a tiny pool of authoritative health, government, and well-known brand sites. Most SEO blogs read the announcement and declared FAQPage schema dead. They were wrong — or rather, they were right about the rich result and wrong about the markup. In 2026 the SERP slot is gone for almost everyone, but FAQPage JSON-LD is now arguably more valuable than it was at peak, because AI search engines — Claude, ChatGPT, Perplexity, Google AI Overviews, You.com — have quietly turned it into one of the strongest signals they use to extract and cite source content.
This is the 2026 guide to FAQ schema: what is and isn't eligible, the exact JSON-LD that works, how each major AI crawler treats it, and the four mistakes that get your perfectly valid markup silently ignored.
The August 2023 update did two things. First, Google restricted FAQ-rich-result eligibility to "well-known authoritative" domains. In practice this means major government domains, recognized health authorities, and a small set of household-name brands. If you run a 50k-monthly-visitor SaaS blog, you do not qualify and almost certainly never will. Second, How-To rich results were dropped from desktop entirely and restricted to mobile on the same authoritative sites.
What did not change: Google still parses FAQPage markup, still uses it for entity understanding, still uses it as input to AI Overviews, and still uses it as a candidate signal for People Also Ask boxes (where FAQ markup is one of several inputs but not a guarantee). The Rich Results Test still validates FAQPage. Search Console still reports FAQ structured-data errors. Nothing about the markup itself was deprecated.
| Surface | Uses FAQPage? | How it shows up |
|---|---|---|
| Google SERP rich result | Authoritative only | Expanded accordion (rare) |
| Google People Also Ask | Candidate signal | Question listing with link to your page |
| Google AI Overviews | Strong input | Answer text + source citation |
| ChatGPT Search | Strong input | Cited source with answer extracted from acceptedAnswer |
| Claude with web search | Strong input | Cited source, often quotes Answer text directly |
| Perplexity | Strong input | Sourced answer card; FAQ pages cite more reliably than prose |
| Voice assistants (Google, Alexa) | Yes for direct answer | Spoken Q&A response |
| Bing chat / Copilot | Yes | Answer extraction with source citation |
The pattern is consistent: anywhere a machine has to pick one chunk of your page to lift as "the answer," pre-segmented Q&A markup wins against unmarked prose. The model does not have to guess where the answer starts and ends — you told it. From an extraction-cost perspective your page is cheaper to cite, and cheaper-to-cite pages get cited more.
<script type="application/ld+json">
{
"@context": "https://schema.org",
"@type": "FAQPage",
"mainEntity": [
{
"@type": "Question",
"name": "How long should a meta description be in 2026?",
"acceptedAnswer": {
"@type": "Answer",
"text": "Google truncates desktop meta descriptions around 158 characters and mobile around 120. Aim for 150 characters with the primary keyword in the first 60 so it survives both truncation points."
}
},
{
"@type": "Question",
"name": "Does Google still use meta descriptions for ranking?",
"acceptedAnswer": {
"@type": "Answer",
"text": "No, meta descriptions are not a ranking factor and have not been one since 2009. They are used as the SERP snippet about 30% of the time; the other 70% Google generates a snippet from the page body."
}
}
]
}
</script>Three rules govern this markup and they are non-negotiable:
display:none accordion, not in JavaScript that fires after Googlebot stops waiting. Crawlable, server-rendered, or hydrated within the standard render budget.name and text. Schema.org allows HTML inside Answer.text, but in practice many AI extractors choke on inline HTML and silently fall back to unmarked prose extraction. Plain text or, if you must, the minimum <p>, <a>, <ul>.QAPage.1. The text is in the markup but not in the DOM. The single most common pattern: a team ships a fresh JSON-LD block via a CMS plugin while the visible page still has the old answers. Or the answers are inside a React tab panel that only mounts on click. Either way Googlebot fetches the page, parses the JSON-LD, scans the DOM for the same answer text, fails to find it, and drops the structured data as inconsistent. Search Console reports zero errors because the markup itself is valid — it just isn't credited.
2. Promotional content inside the answer. Answers cannot be sales pitches. "What is the best CRM for small teams? — Acme CRM is the best because…" is policy-violating. The answer must objectively address the question. You can mention your product as one option among several or in a closing line, but the answer body must be informative first. Promotional FAQPage markup is the most common reason a page gets quietly demoted from PAA candidacy.
3. Mixing FAQ with QA on the same page. A help-center article where you wrote the FAQ at the top and then opened a user-comment thread at the bottom is no longer a pure FAQPage. Either keep the comments separate or split the markup — FAQPage for your authored Q&A and a separate QAPage entity for the user thread — with the right entity for each section.
4. Duplicate FAQ across many pages. Copying the same 8-question FAQ onto every product page on a 2,000-SKU catalog triggers Google's duplicate-content treatment of structured data. The fix is to vary the questions per product (sizing for a t-shirt, charging time for a battery), and to keep cross-site catalog-wide questions in one place — the parent category page or a single help-center article that all product pages link to.
If your goal is to be cited by ChatGPT, Claude, Perplexity, and AI Overviews — rather than to chase the disappearing rich result — the optimization rules are different from classical SEO.
Answer length: 40–90 words. Long enough to be a complete answer the model can lift verbatim, short enough that the model does not need to summarize and lose your phrasing. Sub-30-word answers get rewritten; over-150-word answers get truncated.
Lead with the conclusion. AI extractors quote the first sentence disproportionately. Don't bury the answer behind preamble. "Yes, in 2026 Google still indexes…" wins over "There has been a lot of discussion about whether…".
Include a year, a number, or a named entity. Citations skew toward answers with concrete factual anchors. "Google reduced FAQ rich results in August 2023" is cited more often than "Google reduced FAQ rich results recently."
One unambiguous answer per question. If your acceptedAnswer says "it depends, but typically X unless Y, although Z," the model will rewrite it. State the typical case, then add a one-line nuance.
Google's Rich Results Test validates the markup against schema.org and Google's eligibility rules. It does not validate whether the answer text appears in the rendered DOM — that check happens server-side after crawl. So a green Rich Results Test does not mean the page is credited for the markup. To confirm credit, fetch the URL with the URL Inspection tool in Search Console and check the "Enhancements → FAQ" section after the next crawl.
For AI-search citation testing, the only real check is the manual one: ask Claude, ChatGPT, and Perplexity the exact question from your FAQ and see whether your page is cited. If the answer text in the AI response matches your acceptedAnswer word-for-word, the markup is being read. If it matches the surrounding prose but not the JSON-LD, the model fell back to extraction.
For deeper reading on structured data and AI search behavior, the Google Search Central FAQPage reference is the source of truth for eligibility rules, the schema.org FAQPage definition covers the underlying type system, and the Semrush schema markup guide has up-to-date case studies on FAQ schema in AI Overviews.
The FAQ-rich-result era is over for everyone who isn't a top-100 authoritative domain — and probably will never come back. But the markup itself has quietly become more useful, not less, because every AI search engine in the citation game reads it as a primary signal. Three to twelve real questions, plain-text answers between 40 and 90 words, the same text visible on the page, the conclusion first — and you'll get cited in places SEOs weren't even tracking three years ago.
Free, no signup. Paste your questions and answers, get validator-clean FAQPage JSON-LD with the right schema.org context, plus a paired Article block and Open Graph tags for the page that hosts it.
Open SEO Meta Generator →