Structured data has a reputation problem. It got sold as a way to win rich results, Google then withdrew most of those rich results, and a lot of people concluded the whole exercise was over. That conclusion reads the situation backwards. The rich result was always the visible by-product; the durable thing was making a page machine-legible — and that has become considerably more valuable now that the machine reading your page is writing prose about it.
Does FAQ schema still work in 2026?
Its role changed rather than ended. Google narrowed FAQ rich results to a small set of site types, so the blue-link payoff largely went away. What remained is the machine-readable question-and-answer pairing itself, which is a clean extraction target for AI assistants regardless of whether a rich result ever renders.
The practical read: stop treating FAQ markup as a route to a bigger search listing, and start treating it as the way you hand a model pre-segmented answers.
What each schema type is for
FAQPage
Independent question-and-answer pairs where order carries no meaning. Any answer must make sense read on its own — that independence is the test for whether FAQPage is the right type.
HowTo
An ordered sequence of steps toward a single outcome, where order is the whole point. Using FAQPage for what is really a procedure is the most common structured-data mistake, and it destroys exactly the information the reader needed: the sequence.
The distinction is easy to test. If a reader could do step three before step one, they are FAQs. If they could not, it is a HowTo.
Article, Organization, BreadcrumbList
These do the unglamorous work that citation depends on. Article establishes what the page is and when it was updated. BreadcrumbList places it in a hierarchy. Organization binds the claims to a named entity — and that binding matters more than it appears, because a citation needs something to attribute to. A page full of excellent quotable facts with no entity attached is a page a model can use and forget to name.
Writing answers that survive extraction
The markup is the easy half. The answers are where posts win or lose, because they get quoted with no context whatsoever.
Length: 40 to 60 words
Long enough to answer completely without prompting a follow-up, short enough to be quoted whole. Under about thirty words you usually left out the qualifier that makes the answer true; past eighty, something else decides where to cut you off.
Direct answer first
Open with the answer and put the qualification second. "It depends on your setup — but generally, yes" is the wrong order: truncated at the first clause it says nothing at all. "Yes, in most setups. The exception is X" survives the same truncation intact.
No orphan pronouns
"It processes files locally" has no antecedent once the sentence is travelling on its own. Name the subject in every answer, even where doing so reads as slightly repetitive on the page itself. That repetition is the cost of being quotable, and it is a small one.
Answer the asked question
A question-shaped heading followed by a paragraph about something adjacent is worse than no markup, because you have explicitly promised a machine that the text beneath answers the question above.
The visibility rule
Structured data must describe content the visitor can actually see. Markup describing answers that exist only in JSON-LD is a policy violation, and it is also self-defeating, because the visible text is what retrieval systems quote when the markup is ignored.
There is a straightforward way to stay on the right side of this: write the visible FAQ section first, then generate the markup from it. If the two are produced from a single source, they cannot disagree — the same principle that keeps an llms.txt from going stale.
Getting the JSON-LD right
The failure modes are mundane and all of them silently void the block: a missing @context, mainEntity given an object where an array belongs, an acceptedAnswer without @type: "Answer", or unescaped quotation marks inside an answer string.
The Schema Markup Generator emits valid JSON-LD for FAQPage, HowTo, Article, Organization, BreadcrumbList and LocalBusiness from a form, which removes the hand-editing where those errors originate. If you are checking markup you already have, the JSON Formatter will validate the block and point at the exact character where a stray quote broke it.
Where to put it
A single <script type="application/ld+json"> block, in <head> or <body> — both are read. Do not split one schema type across several blocks, and do not describe a page with markup meant for a different page. Multiple types on one page are fine and normal: an Article plus a FAQPage plus a BreadcrumbList is a perfectly ordinary combination.
Frequently Asked Questions
How many FAQs should a page have?
Three to seven for most pages. The constraint is genuine questions rather than a target count — padding with questions nobody asks dilutes the useful ones and gives an extraction system worse material to choose from.
Can the same FAQ appear on more than one page?
Technically yes, but it splits the signal: two pages now claim the same answer, and neither is clearly the source. Put the canonical answer on one page and link to it from the others.
Do I need schema for AI assistants to understand my page?
Not strictly — models parse plain HTML perfectly well. Schema removes ambiguity rather than enabling comprehension. It matters most where a page's structure is not obvious from the markup alone, which describes most real-world pages.
Generate valid markup with the Schema Markup Generator, then check how the page reads with the Readability Checker. Both free, both entirely in your browser.