Google switched off FAQ rich results in May 2026, and a large part of the industry concluded that FAQ content was finished. That conclusion is wrong, but for a more interesting reason than most defences of it suggest.
The expandable dropdown under a search listing is gone. The FAQ as a unit of retrieval is, on the best evidence available, the single strongest on-page lever for getting cited in an AI answer.
Those are two different things that happened to share a name. One was a display feature that got abused into oblivion. The other is a structural pattern: a page organised around the exact question a person asked, with the answer stated immediately underneath it.
This guide covers why the format works when the rich result does not, what the retrieval data says about questions as headings, how to choose which questions to answer, how to write answers that survive the filtering, where the content should live, and how to measure whether any of it worked.
Why The Format Works When The Rich Result Does Not
Google added a deprecation notice to its FAQ structured data documentation on 7 May 2026, and FAQ rich results stopped appearing in Search from that date, with reporting and API support withdrawn through the following months.
FAQPage remains a valid schema.org type, and Google has said unused structured data does not cause problems for Search, so there is no reason to strip existing markup out in a hurry.
Google’s separate guidance on AI features states there is no special schema.org structured data required to appear in AI Overviews or AI Mode. Its AI optimization guide goes further and lists overfocusing on structured data as a mistake.
Put those together and the honest position is uncomfortable for anyone selling FAQ schema as an AI tactic. The markup is not the mechanism.
The mechanism is the shape of the content. When someone asks an assistant a question, the system searches, retrieves candidate pages, and selects which to cite based substantially on how closely each page’s headings match the query it was searching for.
A question used as a heading is a literal match for a query phrased as a question. That is not a clever optimisation. It is the plainest possible way to signal what a section of a page answers, which is the argument developed in this guide to structuring content for citation.
What The Retrieval Data Says About Questions As Headings
The most useful study in this area comes from AirOps in partnership with Kevin Indig, which analysed 16,851 queries and 353,799 pages across ChatGPT’s retrieval pipeline, sending each query three times and capturing every fan-out search, every retrieved URL and every citation.
1. Heading match is the strongest content signal
The study measured cosine similarity between the query and the best-matching heading on each page. Citation rate rose from approximately 30.2 percent at similarity below 0.50 to approximately 41.0 percent at 0.90 and above.
The relationship is monotonic across every bucket, which is unusual in this kind of observational work and makes it harder to explain away.
The authors describe heading structure as the primary on-page lever for AI citation, more impactful than word count, topical breadth or body copy. An FAQ is the most direct available expression of that lever.
2. Query match still adds nineteen points at the top
Controlling for retrieval position and looking only at pages ranked in the top three, citation rate rose from approximately 55.9 percent at low heading similarity to approximately 75.3 percent at 0.90 and above.
That is the finding that justifies the work. The effect is not an artefact of well-ranked pages happening to have better headings, because it holds when rank is held constant.
It also sets a ceiling on what heading work alone can do, which the next point makes explicit.
3. Retrievability comes before everything else
Retrieval rank was the dominant signal in the study. A page at the first position in ChatGPT’s web search results was cited approximately 58.4 percent of the time, falling to approximately 14.2 percent at position ten.
The authors note that a mediocre page at the top position outperformed a strong page at position six or lower. A page with excellent heading matches at position eleven or beyond was cited around 21.5 percent of the time.
The order of operations is therefore fixed. Be crawlable, be indexed, be findable, then worry about how your headings are phrased. Reversing that order is the most expensive mistake available in this channel.
4. Focused pages beat comprehensive guides
This is the finding that should change most content plans. Holding heading match constant, pages covering approximately 26 to 50 percent of the fan-out subtopics were cited around 38.2 percent of the time, against approximately 34.0 percent for pages covering all of them.
The pattern held at every similarity threshold the researchers tested. Their reading is that exhaustive coverage signals generalist content, while moderate coverage paired with strong primary relevance signals focused expertise.
The ultimate guide playbook that dominated traditional SEO appears to work against citation rates here. That is a direct argument for many narrow FAQ answers rather than one enormous resource page.
5. Breadth of subheading matches dilutes the signal
The same study measured how many distinct subheadings on a page matched fan-out queries. Matching one performed essentially identically to matching none, and matching three or four dropped citation rate by roughly six percentage points.
The practical reading is that the fan-out subtopic list is not a content checklist. Stuffing a page with headings that gesture at every adjacent question makes it a worse candidate, not a better one.
One page, one primary question, a handful of genuinely supporting subheadings. That is the shape the data supports.
6. FAQPage markup correlates with citation but has not been shown to cause it
Pages carrying JSON-LD in the study were cited approximately 38.5 percent of the time against approximately 32.0 percent for pages without it, and FAQPage was among the highest-performing types at roughly 45.6 percent.
The researchers checked whether JSON-LD pages differed on other measurable signals and reported they did not, which makes the association more interesting than usual.
I would still not present it as causal, because a separate controlled test of 1,885 pages that added schema found AI citation changes barely moving, with two of three platform results statistically indistinguishable from zero. Add the markup because it is cheap and accurate. Do not budget on it.
How To Choose The Questions Worth Answering
The selection step decides most of the outcome, and it is where FAQ programmes usually go wrong, because the questions get written by the people who already know the answers.
Start from how buyers actually phrase things rather than how your marketing team writes headlines. “Does it integrate with our ERP” is a real query. “Seamless enterprise interoperability” is not.
Then account for the fact that the system is not searching your exact phrasing. AirOps found that approximately 88.6 percent of ChatGPT queries generated exactly two fan-out sub-queries, with only around 8.8 percent generating none.
Secondary reporting on related datasets puts approximately 95 percent of fan-out phrases at zero recorded monthly search volume, which means keyword tools cannot surface them. I have not verified that figure at source, so treat it as directional.
The tempting conclusion is to extract a fan-out list and answer everything on it. The controlled evidence above says not to. Use the fan-out set to understand intent categories, then pick the two or three that matter and answer those properly.
Five categories are worth covering deliberately, because each pulls from a different part of the source pool. Identity questions ask what you are. Comparison questions ask who else does this. Trust questions ask whether you are legitimate. Pricing questions ask what it costs. Objection questions ask what goes wrong.
Query type genuinely changes what gets cited. Muck Rack’s May 2026 analysis of more than 25 million cited links found industry trend questions driving journalism citations at more than double the rate of how-to questions.
That matters for expectations. A trust question is likely to be answered from third-party material regardless of what your FAQ says, which is why owned content alone cannot carry a reputation programme, a point developed in this breakdown of how AI chatbots source information about brands.
How To Write An Answer That Survives Retrieval
Once the question is chosen, the writing decisions are narrower than they look, and several of them are contrary to standard content advice.
1. Put the question in the heading, close to verbatim
Given that heading similarity to the query is the strongest content signal measured, the heading should be the question as a buyer would type it, not a cleverer version of it.
Resist the instinct to compress it into a noun phrase. “Pricing” is a worse heading than “How much does it cost per user per month” if the second is what people ask.
Use H2 or H3 for each question so the structure is machine-readable, and keep one question per heading rather than bundling several.
2. Answer in the first two sentences
Citations in the AirOps dataset were front-loaded, with roughly 40.7 percent landing in the first third of the answer and around 24.5 percent in the last third.
Whatever the retrieval mechanics, a section that states its answer immediately is easier to extract than one that builds to it. Put the direct answer first, then the qualification, then the detail.
Test each answer by asking whether it would still make sense if it were lifted out of the page and shown alone. If it needs the paragraph above it, it is not ready.
3. Keep each answer self-contained
Avoid pronouns and connectives that reference earlier sections. “This approach” and “as mentioned above” break an extracted passage in ways that are invisible when you read the page top to bottom.
Restate the subject in each answer rather than relying on the reader having arrived in order. A person asking an assistant one question is not reading your page in order, and neither is the system quoting it.
Name specific figures, dates and terms inside the answer rather than deferring them to a table elsewhere on the page.
4. Target the length range the data actually supports
The citation sweet spot in the AirOps data was approximately 500 to 2,000 words per page, with pages over 5,000 words underperforming pages under 500.
For subheading count, four to ten performed best for articles at around 33.2 percent, while one to three performed worst at approximately 28.0 percent, which was worse than having none at all.
The instruction that follows is unusual and worth stating plainly. Either structure a page properly with four to ten question headings, or do not add headings at all. A page with two token subheadings is the worst configuration measured.
5. Write at a professional reading level
Readability peaked at Flesch-Kincaid grade 16 to 17 at approximately 35.9 percent citation, above both simpler and more academic writing, consistent with prior research on how these systems weight attention.
That cuts against the standard advice to write FAQ answers as simply as possible. Clear is not the same as simplified, and the data suggests expert register performs better than plain-language register.
Write the way a knowledgeable practitioner would explain it to a peer. Keep the sentences ordinary and let the vocabulary be precise.
6. Refresh on a schedule rather than on instinct
Page age showed a clear relationship with citation in the same dataset. The sweet spot was approximately 30 to 89 days old at around 32.8 percent, with pages under 30 days underperforming at approximately 25.3 percent and pages older than two years falling to roughly 27.5 percent.
The dip for very new pages is worth knowing before you panic about a launch. The researchers suggest new pages may not have established retrieval signals yet.
Separately, Profound reportedly found a median of around 6.81 days to first citation for newly published pages, with roughly 90 percent of eventually cited pages cited within 37 days. I have not confirmed that at source, but it gives you a reasonable window before treating a page as a failure.
Where FAQ Content Should Live
Placement decides whether an answer is retrievable at all, and the default choice is usually the wrong one.
- Put questions on the page that already ranks for the topic, since retrieval rank dominates every content signal and a new page starts from nothing. A pricing question belongs on the pricing page.
- Keep a central FAQ hub only for genuinely cross-cutting questions, such as company identity, security posture or contract terms, where no single service page is the natural home.
- Give a standalone page to any question with real independent demand, because a focused page that answers one question well is exactly the profile the data associates with consistent citation.
- Render every answer in server-side HTML, since content that appears only after JavaScript executes is content several retrieval systems will not see, and one experiment reportedly found assistants unable to use data present only in JSON-LD during a direct fetch.
- Do not duplicate the same question across multiple pages, which splits the signal and forces the retrieval system to choose between near-identical candidates.
Placement also interacts with identity. A question about who your company is only helps if the system knows which company it is describing, which is the work covered in this explanation of entity SEO and reputation.
What Not To Do With FAQ Content
Several widely recommended tactics are now either dead or actively counterproductive.
- Do not bolt a generic FAQ block onto every template. This is the behaviour that got the rich result withdrawn, and the heading dilution finding suggests it also damages the pages it is added to.
- Do not add FAQPage markup expecting search visibility. The rich result is gone for all sites, including the government and health domains that retained eligibility after 2023.
- Do not write answers that exist only inside structured data. If the answer is not in the visible page, assume it does not exist for retrieval purposes.
- Do not pad the question count to look thorough. Moderate subtopic coverage outperformed exhaustive coverage in the controlled test, and three or four matching subheadings reduced citation rate.
- Do not assume domain authority will carry you. In the AirOps dataset, always-cited pages had slightly lower domain authority and roughly a third of the backlinks of never-cited pages, and authority showed no positive correlation with citation at any relevance level.
How To Measure Whether The FAQ Programme Worked
Measurement here is unusually unforgiving, because the outcome is invisible in analytics and the generation is non-deterministic.
Build a fixed set of the questions you chose to answer, phrased as a buyer would type them. Freeze the wording, because changing the prompt changes the answer and destroys your ability to compare.
Run each one several times in a fresh conversation and record the proportion of runs that named you. That proportion, not a single result, is the measurement.
Log four outcomes separately: whether an answer appeared, whether you were named, whether your domain was cited, and whether the description was accurate. Conflating those is the most common reporting error on this surface.
Expect the results to be lumpy rather than gradual. The AirOps citation distribution was bimodal, with approximately 58 percent of pages never cited for any query they appeared in and around 24.7 percent cited every time, leaving only about 17 percent in between.
Also track mention against citation, since Semrush’s index of 126 million prompts found the overlap between mentioned brands and cited domains falling as low as 30 percent on Gemini. Being named without being cited is a different problem from not appearing.
Finally, check your server logs for the retrieval crawlers before concluding anything about content. Zero hits from a retrieval agent on a site that ranks well is a blocking problem, and no amount of rewriting fixes it. The wider audit method sits in these GEO statistics and these AEO statistics.
How FAQ Work Fits A Wider Reputation Programme
FAQ content is the part of this you control outright, which makes it the obvious place to start and a poor place to stop.
Owned content is a minority input in every study of AI citation. The same Muck Rack analysis put earned media at roughly 84 percent of citations against approximately 0.3 percent for paid and advertorial content.
Your FAQ answers do two things inside that reality. They give the system an authoritative first-party source for facts only you can state, and they make your pages retrievable candidates for the specific questions buyers ask.
What they cannot do is displace a hostile third-party source or manufacture credibility you have not earned. Those need coverage, review generation and source-level correction, which is the argument in this piece on SaaS reputation management and this explanation of what AEO means for reputation work.
Nadernejad Media Inc. treats visibility and reputation as one connected programme, pairing first-party content built to be quoted with the monitoring that catches an inaccurate description early, an approach set out further in this case for a professional partner and in these ORM statistics.
Handled that way, an FAQ stops being a page nobody reads and becomes the material a machine reaches for when somebody asks about you.
Frequently Asked Questions
1. Is FAQ schema dead after the May 2026 deprecation?
The rich result is dead. The schema type is not. FAQPage remains valid schema.org markup and Google has said unused structured data does not cause problems for Search, so existing implementations can stay. What has ended is the case for adding FAQ markup in order to win search real estate.
2. Does FAQ content help with AI citation?
The format does, on the best available evidence, though the markup has not been shown to cause it. AirOps found heading-to-query match to be the strongest on-page citation signal, scaling from roughly 30 to 41 percent, and a question used as a heading is the most direct expression of that. A separate controlled test found adding schema moved citations by amounts indistinguishable from zero.
3. How many questions should a page answer?
Fewer than most guides suggest. Four to ten subheadings performed best for articles in the AirOps data, and pages covering 26 to 50 percent of fan-out subtopics outperformed pages covering all of them. One to three subheadings was the worst configuration measured, so either structure the page properly or leave it unstructured.
4. Should FAQ answers be short and simple?
Short, yes, in the sense that each answer should lead with the direct response. Simple, apparently not. Readability in the AirOps dataset peaked at college level, around Flesch-Kincaid grade 16 to 17, outperforming both simpler and more academic writing. That contradicts a lot of standard FAQ advice.
5. How long before an FAQ page shows up in AI answers?
Reportedly days rather than months for pages that will ever be cited, with a median around seven days and roughly 90 percent of eventually cited pages cited within 37 days per Profound. Pages under 30 days old also showed lower citation rates than pages aged 30 to 89 days, so a slow start is not necessarily a failure. Both figures reach me through secondary reporting and should be verified before you set internal targets against them.
