FAQ Schema for AI Citation: How to Structure It in 2026

Writing FAQ content to be cited by AI in 2026 stopped being a schema play and became a writing-structure decision, because Google’s FAQ rich result is gone: it stopped appearing in Search on 7 May 2026. What still holds is the format itself. A block of questions and answers is the most natural way to write atomic, self-contained, quotable claims, and that is exactly the kind of excerpt a generative system can lift without breaking it. This guide, part of the generative engine optimization guide, covers what changed and when, what the academic research found actually works (and what does not), the anatomy of an answer that survives being cut out of its page, and how to test the change with statistical rigor using the right metric, which is not citation.

What changed: the FAQ rich result is gone

The change happened in two steps, almost three years apart.

In August 2023, Google announced that FAQ rich results would only appear for well-known authoritative government and health websites (Google Search Central, Changes to HowTo and FAQ rich results). For the overwhelming majority of sites the feature was effectively over right there, even though plenty of teams kept shipping the markup out of habit.

On 7 May 2026, Google ended it outright: the FAQ rich result stopped appearing in Search, and the official FAQPage documentation was updated to record the deprecation (Google Search Central, FAQPage structured data). The surrounding infrastructure was withdrawn afterwards, including the feature documentation and its support in the reporting tools.

The most important consequence is about justification, not code. For years the argument for writing FAQ content was “we win more SERP real estate”. That argument is dead. What replaced it is a better and harder sell internally: the question and answer format produces the kind of excerpt that survives being cut out, and cutting is exactly what a generative system does with your page.

FAQ schema and AI citation: keep the markup or remove it?

Keep it. FAQPage is still a valid schema.org type, engines beyond Google still process structured data, and removing the markup burns engineering time without returning anything. The table below separates what changed from what did not:

Aspect	Before May 2026	After
FAQ rich result in Google Search	Existed (restricted since Aug 2023)	No longer appears
Rich result report and test coverage	Covered the feature	Support withdrawn
Validity of the FAQPage type on schema.org	Valid	Still valid
Parsing by other engines and crawlers	Possible	Still possible
Value of the FAQ block for human readers	High	High, and now the primary justification

If your pages already carry the markup, leave it alone. If you are building now, implement it because the FAQ block is useful to readers and because it disciplines the writing, not because a visual feature will show up.

The canonical shape has not changed. It is a JSON-LD document with the FAQPage type and a list of Question entries, each with an acceptedAnswer:

{
  "@context": "https://schema.org",
  "@type": "FAQPage",
  "mainEntity": [{
    "@type": "Question",
    "name": "How long should an A/B test run?",
    "acceptedAnswer": {
      "@type": "Answer",
      "text": "An A/B test runs for as long as it takes to accumulate the sample size calculated before launch, typically two to four weeks..."
    }
  }]
}

Notice that the value of text is the entire, self-sufficient answer. It is that text, not the structure wrapped around it, that has to make sense in isolation.

What the research shows actually works

The paper that coined the term GEO, “GEO: Generative Engine Optimization”, presented at ACM SIGKDD 2024, tested optimization tactics on a benchmark of 10,000 queries and reported that GEO methods can boost visibility in generative responses by up to 40% (arXiv 2311.09735). More useful than the headline number is the pattern of which tactics worked. The methods that added verifiable evidence to the text came out on top: quotation addition led, followed by statistics addition and citing sources, all comfortably above the unoptimized baseline. Keyword stuffing landed below the unoptimized baseline, so the classic SEO reflex did not merely fail to help, it cost visibility.

One result deserves to be stated precisely, because it is the easiest one to get backwards. The paper also tested a method it literally named “Authoritative”, which rewrites the text in a more persuasive and authoritative tone without adding any evidence, and found no significant improvement from it. The authors read that as generative engines already being somewhat robust to changes of tone. Adding a source moved the needle; sounding like an authority did not.

That completely reorients how to write an FAQ block:

Practice	Expected effect	Why
Answer carrying a number with an attributed source	High	Statistics addition and citing sources were among the top performers in the GEO benchmark
Quote from a credible source inside the answer	High	Quotation addition was the best performing method in that benchmark, and it gives the excerpt a verifiable anchor
Self-sufficient claim in the first sentence	High	It survives the cut, which is the real operation a generative system performs
Question phrased the way a person actually types it	Medium	Raises the chance of matching the query that triggers the answer
Twelve near-identical questions carrying the keyword	Low or negative	This is the cosmetic tactic the study found performs poorly
Answer that opens with “it depends on several factors”	Low	It asserts nothing; cut out, it says nothing at all

One caveat for rigor: that study measured visibility under controlled laboratory conditions, with a specific methodology, at a specific moment in the evolution of these models. It is not a promise of results on your site, and nobody should sell it as one.

Anatomy of an answer that survives the cut

The operation a generative system performs on your page is cutting. It does not read the article start to finish to compose its answer; it retrieves excerpts and synthesizes. So the practical question is not “is this text good”, it is “does this paragraph stay true and intelligible if it shows up alone, with nothing around it”.

Four rules fall straight out of that:

1. Name the subject in the first sentence. “It usually runs for two weeks” does not survive. “An A/B test usually runs for two weeks” does.
2. Assert before you qualify. Give the answer, then the caveats. An excerpt that opens with three paragraphs of context before the claim gets cut in the wrong place.
3. Put the number next to its source. “According to X, the value is Y” is a complete, verifiable unit. A bare number with no source is exactly what a cautious system avoids quoting.
4. One question, one idea. Answers that weld three subjects together have no natural cut point, so the cut lands in the middle of an argument.

The measurement mistake: citation is not a test metric

Here is the part that separates honest GEO work from metrics theater. After rewriting your FAQs, the natural question is “did it work?”. The temptation is to answer by looking at citations. That does not hold up, for three concrete reasons:

• Volume. Traffic arriving from generative assistants is typically a small fraction of the total, far too small to reach sample size in any reasonable window.
• Granularity. A citation is not a reliably countable event. A large share of generated answers produce no click at all, so they never show up in your analytics.
• Stability. The same query, run twice, can produce different answers. A metric that shifts between runs cannot serve as the verdict of an experiment.

What is left, and what is honest, is measuring the effect of the change on site-wide conversion. It is the same logic applied in does A/B testing affect how AI engines cite your site: run the test where the volume is, decide by it, and treat citation signals as a directional dashboard.

A worked example: testing the FAQ block with the right metric

A content site with 9,000 visits per week and a baseline conversion rate of 3.2% wants to test adding an answer-first FAQ block to the end of its main articles, aiming to detect a 15% relative improvement (taking the rate to roughly 3.68%), at 95% confidence and 80% power.

Current conversion rate%Minimum detectable effectrelative (%)absolute (pp)Confidence909599%Power8090%Visitors per week (total)Testtwo-sidedone-sided

-Visitors per variation

-Total (2 variations)

-Estimated duration

Two-proportion normal approximation, 2 variations (50/50). Tweak the inputs and watch it update live.

Set the calculator above to baseline rate 3.2, minimum detectable effect 15 (relative) and 9,000 weekly visitors, and the result is 22,631 visits per variation (45,262 in total), running for roughly 36 days. Note that 15% relative is already an ambitious effect for a content structure change: asking for 5% would demand several times that volume, which in practice would put the test out of reach for this site.

Now suppose the test ran the planned window and accumulated 14,000 visits in each arm on the participating pages:

• A (control, no FAQ block): 448 conversions out of 14,000 visits, a rate of 3.20%.
• B (variation, answer-first FAQ block): 518 conversions out of 14,000 visits, a rate of 3.70%.

Control (A)

VisitorsConversions

Variation (B)

VisitorsConversions

Confidence90%95%99%

Control (A) · Rate-

Variation (B) · Rate-

Relative lift-

p-value-

95% CI of the difference-

Two-sided two-proportion z-test. "Not significant" almost always means not enough sample, not that the versions are equal.

Paste 14000/448 into A and 14000/518 into B in the calculator above to check: the absolute lift is 0.50 percentage points (a relative lift of +15.63%), the z score comes out around 2.29, and the two-sided p-value is roughly 0.0219, below the 0.05 threshold. The 95% confidence interval for the difference runs from 0.07 to 0.93 percentage points, does not cross zero, and variation B wins.

Notice the honest detail the calculator hands you and a headline would hide: the interval is wide and its lower bound sits very close to zero. The result clears the significance bar, but the true effect may be much smaller than the 0.50 percentage points observed. That is normal when the measured effect lands near the minimum detectable effect the test was sized for, and it is exactly the nuance that a “p below 0.05, we won” reading erases. If you want a tighter interval, the path is more sample, not more confidence in the conclusion.

The most common mistakes

Mistake	Warning sign	Fix
Justifying FAQ work with the rich result	“We will win more SERP space”	The feature ended in May 2026; the justification now is content structure and usefulness
Padding the block with keyword questions	Twelve near-identical questions	GEO research shows cosmetic tactics perform poorly; prefer few and self-sufficient
Answers that depend on the previous paragraph	They start with “it”, “this”, “as we saw”	Name the subject and state the fact in the first sentence
Numbers without a source	“Studies show that 70%…”	Attribute it (“according to X”) or drop it; a bare number is what a cautious system avoids
Using AI citation as the test metric	The test report lists “AI mentions” as a KPI	Decide by site-wide conversion; citation is a directional dashboard
Ripping out the markup in a panic	“The rich result is gone, remove the schema”	FAQPage is still valid; removing it costs work and returns nothing

Make this automatic with Donnu

Rewriting FAQs to be extractable is content work, and it only pays if you can prove it moved something that matters. That is where most teams slip: they measure what is easy (question count, presence of the markup, scattered mentions) instead of what decides (conversion by real people). Donnu handles the proof: you define the change and the primary metric, Donnu sizes the sample needed before you launch, and returns the verdict with the entire confidence interval, including when that interval is too wide to support the headline.

Start a 14 day free trial and test your next content structure change with that rigor. For the full picture, see the generative engine optimization guide.

References

• Google Search Central. FAQPage (FAQ) structured data. Documentation carrying the record of the feature deprecation. developers.google.com/search/docs/appearance/structured-data/faqpage.
• Google Search Central Blog. Changes to HowTo and FAQ rich results. August 2023. developers.google.com/search/blog/2023/08/howto-faq-changes.
• Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., Deshpande, A. GEO: Generative Engine Optimization. ACM SIGKDD 2024. arxiv.org/abs/2311.09735.
• Schema.org. FAQPage. Type definition. schema.org/FAQPage.

Read next:

• Generative engine optimization (GEO): the 2026 guide
• Does A/B testing affect how AI engines cite your site?
• A/B testing statistical significance: the complete guide
• Conversion rate optimization (CRO): the complete guide