FAQ Schema for AI Visibility

Most teams pulled their FAQ schema off the site the week Google stopped showing FAQ rich results. We have shipped more than 250 articles on this site testing what AI engines actually pull into an answer, and the FAQ block is doing more work now than it did when it earned a dropdown. Below, what FAQPage markup really buys you in AI search, what it does not, and the 7 rules we use to write questions an engine will quote.

What Is FAQ Schema and What Does It Do Now?

That distinction is the whole article. For 6 years FAQ schema had two jobs bundled into one tag, and almost nobody separated them. Job one was earning the expandable dropdown under your listing in Google. Job two was telling any machine reading the page which sentences were answers and to which questions.

Job one is dead. Job two never depended on it, and job two is the one that matters now that a growing share of B2B research happens inside a generated answer instead of a results page.

FAQPage schema

A Schema.org type declaring that a page contains a list of frequently asked questions, each paired with a single accepted answer. It is implemented as JSON-LD in the document head, and Google's guidelines require every marked-up question and answer to also be visible to the reader on the page itself. The type remains valid and supported by Schema.org after Google retired the rich result it used to produce.

The reason the two jobs got conflated is that only one of them was ever measurable. Google Search Console reported on FAQ rich results, so teams could see impressions and clicks tied to the markup. Nothing reported on the parsing job, so it may as well not have existed. When the visible result went away, the only number anybody was watching went to zero, and the natural conclusion was that the markup had stopped doing anything.

We took the opposite read, and the rest of this piece is why. If you want the wider frame first, what generative engine optimization is and GEO vs SEO differences explained cover how the target moved from a ranked list to a generated paragraph.

Did Google Killing FAQ Rich Results Make the Markup Worthless?

The timeline is worth having straight, because a lot of the panic in May came from people who thought this was sudden. It was the last step of a phase-out that ran for nearly 3 years.

In August 2023, Google published a Search Central post titled changes to HowTo and FAQ rich results, announcing that FAQ rich results would only show for well-known, authoritative government and health sites. For every other site, the dropdown stopped appearing regularly. Google said site owners could drop the structured data if they wanted, and that there was no need to proactively remove it.

The reason was abuse, and it was not subtle. Sites had been bolting FAQ blocks onto pages purely to occupy more vertical pixels in the results, often with questions nobody typed and answers written to fill the box rather than to answer anything. Search Engine Journal's coverage of the final removal traces the same arc: a feature introduced to help readers, gamed until it stopped helping, then withdrawn.

On May 7, 2026, Google added a deprecation notice to the top of its FAQ structured data documentation. No blog post, no announcement, just a label on a developer doc. Search Console reporting and Rich Results Test support came out the following month, and API support was scheduled to follow in August.

Read those three dates as one story and the lesson is not about FAQ markup at all. It is that a structured data type that exists to win a visual feature is a rented asset, and the landlord can end the lease with a footnote. The types that survive are the ones doing real descriptive work: Organization, Person, Product, Article. We keep 7 schema blocks in the head of every article on this site for exactly that reason, and the reasoning is in entity consistency for AI search.

Do ChatGPT and Perplexity Actually Read Your JSON-LD?

This is the single most misunderstood point in the whole GEO conversation, and getting it wrong sends teams down an expensive dead end.

The mental model most people carry is that schema is a back channel. You whisper something to the machine in JSON, the machine acts on it, and the visible page is for humans. That was roughly true for Google's rich result pipeline, which genuinely did read the markup and render a feature from it. It is not how a language model works.

A model ingesting your page sees a sequence of tokens. The JSON-LD block is part of that sequence when it is included at all, and it reads as what it is: a chunk of braces and quoted strings that happens to repeat text found elsewhere on the page. As ZipTie's analysis of FAQ schema for AI answers puts it, ChatGPT and Perplexity do not parse your FAQPage markup as structured data. They extract the visible question and answer content that the schema mirrors.

So the causal chain runs the other way from what most teams assume. The schema does not make the content extractable. The content is extractable, and the schema describes it. Status Labs reaches the same conclusion across schema types, and Contently's guide to writing FAQs that LLMs cite spends almost all of its word count on the prose rather than the markup, which tells you where the leverage sits.

There is a hard number behind this too. Ahrefs identified 1,885 pages that added JSON-LD between August 2025 and March 2026, matched them against roughly 4,000 control pages with similar prior citation levels, and measured the 30 days before against the 30 days after. As Search Engine Roundtable reported, there was no meaningful citation uplift, and AI Overviews actually showed a small decline for the schema group. The Ahrefs research team has run several studies in this area and they point the same direction.

We covered what that study does and does not prove in answer capsules explained. The short version: schema still helps engines identify and attribute a page correctly, which is a real job. It does not decide which sentence gets quoted.

What Does FAQ Schema Actually Buy You in AI Search?

Take those one at a time, because they are worth different amounts.

Parsing clarity is the smallest of the three. Google, Bing, and several AI crawlers do parse structured data, and a page that declares its question and answer pairs is easier to represent than one where the same content sits in undifferentiated prose. This is real but modest, and it is the part the Ahrefs study shows does not by itself move citations.

Writing discipline is the largest, and it is entirely indirect. You cannot mark up an FAQ without first writing a question and a self-contained answer to it. That constraint drags the writer into exactly the shape that gets quoted: the question stated plainly, the answer starting in sentence one, no wind up, no backward pointing pronouns. Teams that add FAQ sections write more extractable prose than teams that do not, and it has almost nothing to do with the JSON.

Chunking behavior is the one nobody talks about. Retrieval systems split pages into fragments at index time and match a question against fragments rather than pages. A question heading followed by a short answer is a natural boundary. A wall of prose is not, and a fragment that starts mid-argument and ends mid-example is worth very little at retrieval time. We went deeper on that mechanic in why AI answers cite some companies and not others.

The scale of the opportunity is easy to underrate. Semrush's expanded 2026 AI Visibility Index analyzed 126 million US AI search prompts and found ChatGPT citing an average of 15 sources per response. That is 15 slots on a single answer, and most of your competitors are still writing pages with no cleanly liftable block anywhere in them.

The visitors those slots send are also not ordinary. Semrush's study of AI search traffic found AI search visitors converting at several times the rate of traditional organic visitors, which is the honest argument for doing this work while the volume is still small.

How Do You Write FAQ Questions an Engine Will Cite?

The rich result era trained an entire industry to write FAQs badly, and most teams have not unlearned it. When the goal was occupying pixels, the incentive was volume and keyword stuffing. When the goal is being quoted inside somebody else's answer, the incentive inverts completely.

Seven rules cover almost every FAQ we write. None is complicated, and all of them get broken constantly.

1. Pull questions from conversations, not keyword tools. The questions worth marking up are the ones prospects asked on calls, in replies, and in the chat widget. Volume data tells you what people search. Your inbox tells you what they are stuck on, and the second list converts better.
2. Phrase the question the way a person types it. "How much does an SDR cost per month" beats "SDR Cost Considerations". Models match a user's phrasing against your text, so the closer your heading sits to the natural question, the better the match.
3. Answer in the first sentence. Not a restatement of the question, not a note about why the topic matters. The answer. If sentence one could be deleted without losing information, delete it and start with sentence two.
4. Name the subject by name. A model lifting your answer has no access to the question above it. "It usually takes 6 weeks" is useless out of context. "A new sending domain usually needs 6 weeks of warmup" survives extraction intact.
5. Carry exactly one qualifier. A number, a range, a timeframe, or a condition. Unqualified claims read as marketing and get passed over for a source that committed to something specific.
6. Keep answers between 40 and 80 words. Under 40 and you usually dropped the qualifier. Over 80 and you are covering two ideas, which forces the engine to cut, and a cut quote is a quote you no longer control.
7. Mirror the markup to the visible page exactly. Google's guidelines require marked-up content to be visible to the reader. Marking up questions that do not appear on screen is a policy violation, and it breaks the only mechanism that helps with AI answers anyway.

Rule 4 is the one that breaks most often. Writers use pronouns to avoid repetition, because repetition reads clumsy to somebody reading top to bottom. Inside an FAQ answer that instinct is backwards. Repetition is what makes the block portable. Our longer treatment of that pattern is in how to structure content for ChatGPT.

Where Should FAQ Schema Go, and Where Does It Backfire?

The placement rule is simple and depends on page length.

• Long articles: put the FAQ near the bottom. The body has already answered the main questions, so the section works as a mop-up for the follow-ups that did not fit the argument.
• Short service and comparison pages: put it higher, often directly after the main pitch. On a short page the FAQ is frequently the only block with a clean question and answer structure, so burying it wastes the most extractable content you have.
• Product and pricing pages: put it wherever the objection actually lands. If people ask about contract length after they see what is included, the question goes there.

The failure modes are more useful than the placement rules, because they are what most sites get wrong.

• Invented questions. If nobody has asked it, do not mark it up. A page with 14 questions and 4 real ones reads as padded to a human and to a model.
• Markup that is not on the page. A direct spam policy violation, and pointless anyway, since the model reads the visible text.
• The same FAQ block on every page. Sitewide FAQ components are convenient and they train engines to treat your questions as boilerplate. Write per-page questions or skip the section.
• Answers that sell. An answer that names your product in the first sentence gets skipped as promotional. Earn the citation with the answer and let the attribution link do the selling.
• Contradicting your own site. If an FAQ answer conflicts with your llms.txt or your LLM info page, you have given the model two versions of the truth and no way to choose. That is a bigger problem than a missing FAQ.

The consistency point deserves more weight than it usually gets. Every claim we publish about the offer has to match across the blog, the schema, and the machine-readable files, or an engine picks whichever version it saw last. The mechanics of keeping those in line are in schema markup for podcast pages and the generative engine optimization checklist.

How Do You Measure Whether FAQ Schema Is Doing Anything?

Losing the Search Console report was the moment most teams stopped being able to see this work, and it is why so many concluded it had stopped mattering. The measurement did not need to die with the feature. It needed to move.

The loop we run is deliberately boring. Fix a list of 20 to 40 questions that a real buyer would type. Run them across the major assistants on a set cadence. Record 3 fields per run: whether the brand appeared, where in the answer, and the exact sentence that got quoted.

That third field is the one teams skip and the only one that turns the next revision into an edit rather than a guess. A yes or no tells you the score. Knowing which 50 words got lifted tells you what the engine considers quotable in your category, and after a few months that log is the most useful content asset you own.

Rank tracking cannot substitute for it. A page can sit at position 3 in organic results and never appear in a generated answer, and the reverse happens too. It is also worth knowing what the ranked list is now worth: Ahrefs measured position 1 click-through rate falling from 1.41% to 0.64% on pages where an AI Overview appears, across a 300,000 keyword study. Our full setup is in how to track AI search visibility, the brand audit version is in how to audit your brand in ChatGPT, and the reason we do not judge this on sessions is in LLM citation vs SEO traffic.

Set expectations on timing before you start. Crawl, index, and re-embed all lag, so a fair first read on an FAQ rewrite is 4 to 8 weeks out, not next Tuesday. Anyone promising a faster loop is selling something. The platform-specific patterns are in how to get cited by ChatGPT and Perplexity, how to get cited by Perplexity, and AI Overviews for B2B.

What Does This Look Like on a Podcast Library?

This is where the topic stops being abstract for us. Every client show produces recordings full of specific, first hand answers from operators who actually run the thing being discussed. That is precisely the material engines want to quote, and in raw transcript form almost none of it is usable.

The fix is mechanical. Pull the real questions the episode answered. Write each as a heading. Write a 60 word answer underneath that says what the guest said in a form that stands alone. Mark the set up as FAQPage, link to the full episode below it. A wall of transcript becomes 6 retrievable sections, each competing for a different question. The longer version is in turning episodes into answer content and podcast transcripts for AI search.

The compounding effect is real and slower than most GEO promises. A marked-up FAQ does not get you cited this week. It makes a page eligible for a selection step it was previously failing, and eligibility only pays once retrieval catches up. How to get your podcast cited by AI assistants covers what that looks like across a full library.

What makes this work for us is that the search footprint is the second return, not the first. The invite is the acquisition channel. We invite a client's ideal buyers onto their own show by email, the recording is the conversation, and our commitment is 30 recorded conversations with those buyers in 90 days or their money back. Editing is included, the client owns the show and every recording, and invites go out by email only. The citable library that builds up afterward is a second return on work already done, which is the honest way to frame it. More on the mechanics in what reverse outbound is and what a podcast does for your search footprint.

One last note for anyone running this alongside cold email. The two disciplines share a habit worth naming: both reward being specific and punish being generic, and both take weeks to read properly. If the sending side is new to you, start with what email deliverability is, SPF, DKIM, and DMARC explained, how to warm up a new email domain, and what domain reputation means. Getting your ideal customer profile right is the input to both.

The Part Nobody Wants to Hear

There is no version of this where the JSON does the work. Adding FAQPage markup to a page whose answers are vague, padded, or written to fill a box changes nothing, because a model reading that page still finds nothing worth quoting. The markup describes the content. It cannot rescue it.

Which means the actual task is the unglamorous one. Sit with the questions your buyers really ask, answer each in 60 honest words, delete the 6 questions you invented to make the section look thorough, and check that every marked-up answer appears on the page as written. There is no tool to buy and no dashboard that lights up when you finish.

The upside is that most of your competitors read the May deprecation notice, concluded the format was finished, and moved on. They are wrong in a way that takes 4 to 8 weeks to prove, which is exactly long enough for the gap to stay open. Write the questions people actually ask, answer them so the answer survives being lifted out of the page, and start logging which sentences come back to you in generated answers.