Schema markup does not directly cause AI citations. The studies that claim 2.5× lifts and the studies that find zero correlation are both real, and both are measuring something narrower than they admit. What schema actually does is ground your brand as an entity Google's Knowledge Graph and Bing's index can resolve, which is what ChatGPT and Google AI Overviews then read against. The honest 2026 playbook is four blocks (Organization with sameAs, Person, Article, and Product) implemented for entity grounding, not for rich snippets that no longer exist.
If you are reading another schema-and-AI guide that opens with "pages with schema are 2.7× more likely to be cited," close it. The number is real, the framing is misleading, and the deprecation calendar Google has been publishing for three years means half the markup those guides recommend stops rendering in June. The playbook below is what we ship for Parse customers as of May 2026: a short, opinionated stack that grounds the brand as an entity, plus the explicit list of schema types we tell teams to stop maintaining. Parse tracks AI visibility across ChatGPT, Google AI Overviews, and Perplexity, and the pattern in our citation data is consistent with the AccuraCast and BrightEdge studies referenced below: schema correlates with visibility because it correlates with brands that have done the rest of the work.
Why the schema-and-AI evidence looks contradictory
The two most-cited 2025–2026 studies disagree by an order of magnitude, and both are credible. BrightEdge reported sites running structured data plus FAQ blocks saw a 44% lift in AI search citations, with comprehensive-schema pages roughly 3× more likely to appear in Google AI Overviews (BrightEdge). AccuraCast's analysis of 9,000 citation sources from 2,000 prompts across ChatGPT, AI Overviews, and Perplexity reported that 81% of cited pages had some schema, but found Person was the only schema type with a clean positive correlation; FAQPage appeared on only 1.8% of cited pages and HowTo on under 1% (AccuraCast). A December 2024 Search/Atlas study went further, finding no correlation at all between schema coverage and AI citation rates (Search Engine Land). All three can be true at once. Schema lifts citations on the surfaces where it lifts indexing, and is invisible on the surfaces that retrieve from training data.
What schema actually does for AI
Schema does not feed AI directly. It feeds the search infrastructure AI runs on. Google's official position, restated in the AI Features documentation, is that no special schema is required for AI Overviews or AI Mode and that no new markup unlocks them (Google Search Central). Microsoft's Fabrice Canel has been less hedged, saying in March 2025 that schema markup helps Microsoft's LLMs understand content. ChatGPT Search runs on Bing's index. Google AI Overviews and AI Mode are constructed from the Google top-10 plus Knowledge Graph entities. The mechanism, then, is two-step. Schema makes a page easier for a search engine to parse and a Knowledge Graph to resolve as an entity. The AI surface inherits the result. If your content lives on a domain neither index trusts, schema will not rescue it. If your content already ranks but your brand is unresolved as an entity, schema is the cheapest fix on the board.
Start with one block: Organization plus sameAs
If you implement one piece of schema this quarter, make it Organization with a populated sameAs array pointing at your Wikipedia article (if you have one), your Wikidata Q-ID, your LinkedIn company page, and your verified social profiles (Stackmatix on Knowledge Graph). The reason this block earns its priority slot is mechanical: Google's Knowledge Graph uses Wikidata as a primary entity-resolution source, and sameAs is the explicit assertion that the entity at your domain is the same entity at that Wikidata Q-ID. ChatGPT reportedly weights Wikipedia at roughly 20% and Wikidata at 10% for entity understanding. When AI is asked "what is Acme HR Software" and you have not closed this loop, the AI is reading a fragmented entity. When you have, the same query resolves to a unified profile with your founders, funding, product category, and verified social presence. For the deeper Wikidata workflow, see our guide to creating a Wikidata entry for your brand.
Person schema is the only schema type that consistently correlates
AccuraCast's data on the 9,000-citation sample is the cleanest signal in the public record: Person schema appeared on 70.4% of ChatGPT-cited pages and 56% of Google AI Overview-cited pages, and it was the strongest single predictor across the platforms (AccuraCast). The mechanism is consistent with Google's E-E-A-T documentation and Microsoft's stated position: AI systems are evaluating author identity as a proxy for content trust. Implement Person on every author byline, with name, jobTitle, worksFor (linked to your Organization @id), sameAs (LinkedIn, ORCID for technical authors, X/Twitter, personal site), and knowsAbout (the topic clusters this author writes about). Bare bylines are wasted entity surface. If you publish under a single editorial brand voice, attach Person schema to the named editor or the most senior reviewer; do not attach it to a generic "Editorial Team" page that no Knowledge Graph will resolve.
Article and Product schema: useful, not transformational
Article (or BlogPosting) and Product round out the working stack. Article gives the search engine an unambiguous signal of headline, datePublished, dateModified, author (linked to the Person schema above), and publisher (linked to the Organization schema above). The lift is real but modest; the AccuraCast sample showed Article appearing on roughly 40% of cited pages, well below the threshold that would suggest it is causally driving citation. Product schema matters more on commerce surfaces because it feeds Google Shopping, the Merchant Center, and the agentic-commerce pipelines emerging on the same infrastructure. Implement it on every product detail page with name, image, description, brand (linked to your Organization), offers (price, availability, currency), and aggregateRating where you have verified review data. Service businesses use the equivalent Service or LocalBusiness types. None of these blocks unlock AI by themselves; they make the rest of the stack legible.
What to remove from your schema stack
The cleanup is the part most teams skip, and it is now non-trivial because Google has retired several feature types. FAQPage rich results no longer appear in Google Search as of May 7, 2026; rich-result reporting and the Rich Results Test stop supporting FAQPage in June 2026 and the Search Console API in August 2026 (Google Search Central). HowTo rich results on desktop ended in 2023 and were fully retired across surfaces shortly after (Google Search Central Blog). Google has also wound down Book Actions, Course Info, ClaimReview (for non-fact-checkers), Estimated Salary, Learning Video, Special Announcement, and Vehicle Listing rich results. Removing the markup is not required (Google has explicitly said unused structured data does not harm Search), but maintaining it as if it earned snippets misallocates engineering time. Keep the underlying FAQ section in your prose, because it still helps AI retrieval. Strip the FAQPage JSON-LD only if you are reorganizing the page anyway.
The deprecation kills the rich-result rendering, not the value of an FAQ section. AI models still chunk and retrieve FAQ blocks because of the question-and-answer structure in the HTML, regardless of the JSON-LD wrapper. Keep the section. Drop the schema only if you are touching the page for other reasons. Do not invest engineering time maintaining FAQPage markup expecting it to help AI directly.
How the studies stack up at a glance
The conflicting evidence base is easier to read as a chart than as a list of citations.
Treat this as a snapshot for the May 2026 version of the playbook. Every row should be re-checked annually, because schema-to-AI behavior is the most volatile part of the stack.
| Study (date) | Sample | Headline finding | What it actually proves |
|---|---|---|---|
| BrightEdge (2025) | Cross-vertical | Schema + FAQ blocks → 44% citation lift; comprehensive schema → 3× AIO presence | Schema correlates with brands doing the broader work |
| AccuraCast (2026) | 9,000 citations, 2,000 prompts | 81% of cited pages had schema; Person was the only consistent positive | Person schema is the cleanest signal; most other types are noise |
| Search/Atlas (Dec 2024) | Cross-domain audit | No correlation between schema coverage and AI citation rate | Comprehensive schema does not cause citation by itself |
| ALM Corp (Feb 2025–Feb 2026) | 9-industry AIO tracker | AI Overviews now trigger on roughly half of tracked queries; +58% YoY | The surface schema affects is growing fast; investment window is short |
The diagonal of these studies is consistent: schema lifts visibility through entity grounding and Bing-index parsing, not as a standalone ranking signal. The vendor pieces that report 2.5–3× lifts are mostly observing self-selection: the brands that ship comprehensive schema also ship the rest of the program. That does not make schema useless. It means it is a hygiene layer, not a lever.
If you want to see which sources shape AI answers about your brand, run a free brand check — it takes a minute.
A SaaS company JSON-LD template
This is the block we drop on the corporate footer of B2B SaaS sites we work with. It assumes you have a Wikidata Q-ID, even a small one; if you do not, get one before you ship this and use a placeholder URL only as a stop-gap. Replace bracketed values; keep the @id pattern so other schema blocks on the site can reference this Organization without redefining it.
{
"@context": "https://schema.org",
"@type": "Organization",
"@id": "https://yourdomain.com/#organization",
"name": "Acme HR Software",
"url": "https://yourdomain.com/",
"logo": {
"@type": "ImageObject",
"url": "https://yourdomain.com/logo.png",
"width": 600,
"height": 60
},
"description": "Acme HR Software is a B2B SaaS HR platform for mid-market companies (50–1,000 employees).",
"foundingDate": "2018-04-15",
"founder": [{ "@type": "Person", "name": "Jane Founder" }],
"sameAs": [
"https://www.wikidata.org/wiki/Q1234567890",
"https://en.wikipedia.org/wiki/Acme_HR_Software",
"https://www.linkedin.com/company/acme-hr/",
"https://www.crunchbase.com/organization/acme-hr",
"https://github.com/acme-hr",
"https://x.com/acmehr"
],
"knowsAbout": [
"Human resources software",
"Performance management",
"Employee onboarding"
]
}
The sameAs array is the part that does the work. Wikidata first because it is the canonical entity URL Google reads. Wikipedia next if you have an article. LinkedIn and Crunchbase round out the B2B identity layer. GitHub and X confirm the digital footprint. Long-tail social profiles add more noise than signal, so keep the array tight to the references AI infrastructure actually resolves.
An ecommerce JSON-LD template
For ecommerce, your Organization block on the homepage looks similar to the SaaS template (swap knowsAbout for product categories). The lift comes from Product schema on each PDP. Implement it once in the template, populate from your PIM, and validate before launch. The agentic-commerce pipelines OpenAI's Operator, Google's Universal Commerce Protocol, and Amazon Rufus run on are reading this exact field set.
{
"@context": "https://schema.org",
"@type": "Product",
"@id": "https://yourstore.com/products/acme-running-shoe-v3#product",
"name": "Acme Running Shoe v3",
"image": [
"https://yourstore.com/img/shoe-v3-1.jpg",
"https://yourstore.com/img/shoe-v3-2.jpg"
],
"description": "Lightweight, high-rebound running shoe designed for road racing.",
"sku": "ACME-RUN-V3-W9",
"gtin13": "0123456789012",
"brand": { "@id": "https://yourstore.com/#organization" },
"offers": {
"@type": "Offer",
"url": "https://yourstore.com/products/acme-running-shoe-v3",
"priceCurrency": "USD",
"price": "149.00",
"availability": "https://schema.org/InStock",
"itemCondition": "https://schema.org/NewCondition"
},
"aggregateRating": {
"@type": "AggregateRating",
"ratingValue": "4.6",
"reviewCount": "327"
}
}
gtin13 (or the ISBN/MPN equivalent) is what lets the agent reconcile your product against the broader catalog graph. Without it, the same shoe on five retailers reads as five different products and your offer competes against itself. aggregateRating should be populated from a verifiable on-page review widget; do not synthesize it.
An agency or services JSON-LD template
Agencies and service businesses get a different lift profile. The right primary type is Organization with a serviceType array, plus Service blocks for each named offering. Person schema on every consultant byline matters more than it does on a SaaS site, because the AccuraCast data shows Person is the cleanest predictor and consultancies sell the people.
{
"@context": "https://schema.org",
"@type": "ProfessionalService",
"@id": "https://youragency.com/#organization",
"name": "Northwind Growth",
"url": "https://youragency.com/",
"description": "Growth and AI visibility consultancy for B2B SaaS companies.",
"sameAs": [
"https://www.wikidata.org/wiki/QXXXXXXX",
"https://www.linkedin.com/company/northwind-growth",
"https://www.crunchbase.com/organization/northwind-growth"
],
"areaServed": "Worldwide",
"knowsAbout": [
"AI visibility",
"Generative engine optimization",
"B2B SaaS growth strategy"
],
"employee": [
{
"@type": "Person",
"@id": "https://youragency.com/team/jane-consultant#person",
"name": "Jane Consultant",
"jobTitle": "Principal, Growth",
"sameAs": [
"https://www.linkedin.com/in/janeconsultant",
"https://x.com/janeconsultant"
],
"knowsAbout": ["AI visibility", "B2B SaaS growth"]
}
]
}
Each consultant page should expand the Person block with worksFor pointing at the @id of the Organization above, plus an alumniOf array if their education is part of the credibility surface (it usually is for executive-coaching and management-consulting verticals).
A local business JSON-LD template
Local businesses inherit a separate path because Google's Knowledge Graph integrates LocalBusiness with the Maps and Business Profile entity layers. Schema App and the Search Engine Land guidance both flag this as the single highest-ROI schema work for local. The block belongs on the contact or location page, not the homepage.
{
"@context": "https://schema.org",
"@type": "LocalBusiness",
"@id": "https://yourbusiness.com/#localbusiness",
"name": "Acme Dentistry of Brooklyn",
"image": "https://yourbusiness.com/storefront.jpg",
"address": {
"@type": "PostalAddress",
"streetAddress": "123 Atlantic Ave",
"addressLocality": "Brooklyn",
"addressRegion": "NY",
"postalCode": "11201",
"addressCountry": "US"
},
"geo": {
"@type": "GeoCoordinates",
"latitude": 40.6904,
"longitude": -73.9926
},
"telephone": "+1-718-555-0142",
"openingHoursSpecification": [
{
"@type": "OpeningHoursSpecification",
"dayOfWeek": ["Monday", "Tuesday", "Wednesday", "Thursday", "Friday"],
"opens": "09:00",
"closes": "18:00"
}
],
"sameAs": [
"https://www.google.com/maps/place/?q=place_id:ChIJXXXXXXX",
"https://www.facebook.com/acmedentistrybrooklyn",
"https://www.yelp.com/biz/acme-dentistry-brooklyn"
]
}
The Google Maps sameAs link is the high-leverage one. It collapses the entity into Google's Maps graph, which the local pack and AI Overview "near me" answers both pull from.
Validate, deploy, and audit on a 90-day cycle
Schema rots. Deploys clobber it, CMS upgrades reset it, and Google deprecates rich-result types on a steady cadence. Treat the schema stack as code with an explicit validation and audit loop. On every deploy, run the affected URLs through Google's Rich Results Test and the Schema.org validator. After deploy, monitor Google Search Console's Enhancements section for warnings; the report is incomplete after the FAQPage and HowTo retirements but still surfaces breakage on the supported types. Quarterly, run a structured-data crawl across the full site (Screaming Frog with the Custom Extraction or Structured Data tabs) and reconcile against the canonical templates. Anything that drifted gets fixed in the next sprint. Anything that has been deprecated since the last audit gets removed or marked as legacy. The single most common failure mode we see is invisible schema: a CMS plugin updated, the JSON-LD now references the wrong @id, and the entity graph silently fragments.
What to do this quarter
Skip the full-site audit if you cannot ship every fix; pick the highest-leverage subset instead. Step one is the Organization block with a real sameAs array, including a Wikidata Q-ID. If you do not have a Q-ID, prioritize creating one this month over any other schema work. Step two is Person schema on the top ten author bylines on your site, weighted toward the editorial pages AI is most likely to cite for your category. Step three is Product or Service schema on the pages that drive pipeline. Step four is removing the FAQPage and HowTo JSON-LD only on pages you are touching for other reasons; do not run a separate cleanup project. Five is setting the 90-day audit cadence and running the first one. The combined effort for a mid-market site is usually one engineer-week and one strategist-week. The compounding payoff comes from the entity grounding, not from any single schema type. For where this fits in the broader citation program, see how to structure content so AI models cite it, and for the source-side picture of which domains AI cites in your category, see which domains do AI models cite most.
Frequently asked questions
Does schema markup directly cause AI to cite my page?
No. Both the BrightEdge and AccuraCast studies, which broadly disagree, agree on this much. Schema correlates with citation because it correlates with brands doing the rest of the work: entity definition, content quality, technical SEO. The mechanism schema acts through is the search infrastructure AI runs on: Bing for ChatGPT, Google index plus Knowledge Graph for AI Overviews. Implementing schema on a low-authority domain will not generate AI citations on its own.
Should I still use FAQPage schema after the May 2026 deprecation?
Keep the FAQ section in your prose; AI retrieval still chunks question-and-answer structures regardless of JSON-LD. Drop the FAQPage schema only when you are reorganizing the page for other reasons. Google has explicitly said unused schema does not harm Search, so there is no urgent cleanup need. Do not invest engineering time maintaining FAQPage markup as if it earned rich snippets; those snippets stopped rendering on May 7, 2026.
What is the single highest-leverage schema implementation for AI visibility?
Organization schema with a populated sameAs array linking to your Wikidata Q-ID and Wikipedia article (if you have one), plus LinkedIn, Crunchbase, and verified social profiles. This block grounds your brand as a resolved entity in Google's Knowledge Graph, which AI Overviews and AI Mode read against. If you do not have a Wikidata Q-ID, create one before doing any other schema work; it is the upstream dependency for everything else.
Is Person schema worth implementing on every author byline?
For the top ten to twenty bylines on a mid-market site, yes. AccuraCast's 9,000-citation sample found Person schema on 70.4% of ChatGPT-cited pages and 56% of Google AI Overview-cited pages, the only schema type with a clean positive correlation. Implement Person with name, jobTitle, worksFor linked to your Organization, sameAs to LinkedIn and other verified profiles, and knowsAbout for the topic clusters the author covers.
How often should I audit our schema stack?
Quarterly is the right cadence for most mid-market sites. Run a structured-data crawl, reconcile against your canonical templates, fix drift, and remove anything Google has deprecated since the last audit. After every deploy, validate the affected URLs through Google's Rich Results Test. The most common failure mode is silent breakage: a CMS or plugin update that points the JSON-LD at a wrong @id, fragmenting the entity graph without throwing any warning.
:::
The honest version of this playbook is shorter than the vendor articles and asks for less work. Three or four well-implemented schema blocks with real entity grounding will outperform a comprehensive 12-type rollout that ignores sameAs, every time we have measured it. Treat schema as a hygiene layer that helps AI infrastructure read your brand correctly, not as a lever that promises citations. Then put the engineering hours saved into the upstream work (Wikidata, Wikipedia, the earned-media surface AI is actually citing) that the data says drives the bigger lift.