Does Schema.org improve AI visibility?
By Eduardo Mendonça, founder of destaque.ai. Reviewed on 13 November 2026.
In the public studies, adding schema to a page showed no gain in citations in AI answers. It helps the page be read and the brand be identified, and it makes pages eligible for Google's rich results, but it is not a way to be cited more.
What Ahrefs measured
Ahrefs, which sells SEO tools, published the study We Tracked 1,885 Pages Adding Schema. AI Citations Barely Moved. on 11 May 2026, written by Louise Linehan and revised on 9 June. It followed 1,885 pages that added JSON-LD between August 2025 and March 2026 and compared them with 4,000 pages that did not. It counted each page's citations in the 30 days before and the 30 days after, in Google's AI Overviews and AI Mode and in ChatGPT.
The authors conclude that “adding schema produced no major uplift in citations on any platform”. Citations rose 2.4% in AI Mode and 2.2% in ChatGPT, both indistinguishable from zero. In AI Overviews they fell 4.6% against the comparison pages, a small but statistically significant difference that the authors do not attribute to schema: “we don't know why”.
The study has limits, and the authors state them. Every page already had more than 100 AI Overview citations before adding schema. For pages that AI systems are not seeing at all, they write that schema “might still play a role” and that the study “can't speak to that directly”. They also grouped all schema types together, measured only 30 days, and counted only schema written in the HTML, not schema injected by JavaScript.
What Google and Microsoft say
On the Search Central page AI features and your website, updated on 10 December 2025, Google says there are no additional requirements to appear in AI Overviews or AI Mode. On schema, it writes that “there's also no special schema.org structured data that you need to add”. Among its best practices, it lists “making sure your structured data matches the visible text on the page”.
The guide Optimizing your website for generative AI features on Google Search, updated on 10 July 2026, repeats that structured data is not required and gives the reason to keep it: “it helps with being eligible for rich results on Google Search”.
Microsoft goes further than Google, in two posts on the Bing Webmaster Blog. In the post of 20 November 2025, it writes that structured content “helps AI systems interpret and summarize your content more effectively”. Its examples are schema-marked product pages, FAQs and comparison tables, and it speaks of AI systems in general, not of Copilot by name. In the post of 19 May 2025, on IndexNow and shopping, it says product schema helps surface a product accurately. The words are “in search results, shopping experiences, and AI-driven assistants”. Interpreting and summarising a page is not the same as citing it.
Where schema still counts
- Identity. An
OrganizationwithsameAslinks the brand to its profiles elsewhere, so an engine knows those pages describe the same company. None of the sources above deals with Organization specifically. This is our practice, within Google's rule that structured data should say what the page shows. - Products. Product schema (in practice,
Productwith theOfferfor price and availability) is the case where Microsoft says it helps AI assistants, in the May 2025 post. - Google rich results. This is the reason Google itself gives for continuing to use structured data, in the July 2026 guide.
| Study or document | Sample | What it says | Date | Source |
|---|---|---|---|---|
| Ahrefs study on schema and citations | 1,885 pages with new JSON-LD and 4,000 comparison pages, 30 days before and after | No major uplift in citations. AI Mode +2.4% and ChatGPT +2.2%, indistinguishable from zero. AI Overviews -4.6%, cause not attributed | 11 May 2026 (revised 9 June) | Ahrefs |
| AI features and your website | Documentation, no sample | No special schema for AI Overviews and AI Mode. Structured data should match the visible text | 10 December 2025 | Google Search Central |
| Optimizing your website for generative AI features on Google Search | Documentation, no sample | Structured data is not required and helps with eligibility for rich results | 10 July 2026 | Google Search Central |
| How AI Search Is Changing the Way Conversions are Measured | Blog post, no sample | Structured content helps AI systems interpret and summarise | 20 November 2025 | Bing Webmaster Blog |
| Post on IndexNow and shopping | Blog post, no sample | Product schema helps surface a product in search results, shopping and AI assistants | 19 May 2025 | Bing Webmaster Blog |
The guide below, with the five schema types and the code for each, was published on 8 May 2026, before the Ahrefs study, and still holds for the three purposes above. Where it says schema makes a page citable by AI, read it in the light of what the studies measured.
For a B2B SaaS company, there are five schema types worth more than all the others combined: Organization, Service, FAQPage, BreadcrumbList and Article. Implemented in JSON-LD, with stable @ids and links between them, they are the minimum viable for the page to be read without ambiguity and the brand identified; on their own they are not a citation lever. This piece has the snippet for each one, ready to adapt.
Key takeaways
- JSON-LD is the right serialisation in 2026
- five types solve 80% of the upside: Organization, Service, FAQPage, BreadcrumbList, Article
- stable @ids turn loose entities into a graph
- validate at validator.schema.org and in the Rich Results Test
- irrelevant schema is worse than no schema
Why schema matters for GEO
Schema.org is the standard vocabulary for structured data. In HTML, describing something as an organisation is implicit (from tags like header, from CSS classes, from context). In schema, it is explicit: @type: Organization.
For traditional engines, schema unlocks rich results (stars, expanded FAQ, visible breadcrumbs). For LLMs, schema is the only reliable way to know, without ambiguity, what each thing is. When ChatGPT cites "Acme is a cybersecurity consultancy in Lisbon", it is inferring three things, what it is, what it does, where, and schema narrows the room for error.
The five types, one per tab
- 1. Organization
- The foundation. One Organization entity per company, with a stable @id, referenced by all the other entities as provider, author, publisher.
- { "@context": "https://schema.org", "@type": "Organization", "@id": "https://acme.com/#organization", "name": "Acme SaaS", "url": "https://acme.com", "logo": "https://acme.com/logo.png", "description": "Inventory management platform for EU retailers.", "foundingDate": "2021", "address": { "@type": "PostalAddress", "streetAddress": "Rua X, 42", "postalCode": "1234-567", "addressLocality": "Lisboa", "addressCountry": "PT" }, "contactPoint": { "@type": "ContactPoint", "contactType": "sales", "email": "contacto@acme.com", "availableLanguage": ["pt-PT", "en", "es"] }, "sameAs": [ "https://www.linkedin.com/company/acme-saas", "https://github.com/acme-saas" ], "knowsAbout": ["Inventory Management", "Retail Tech", "ERP"] }
- Three critical fields: sameAs (links to consistent external profiles, LinkedIn, GitHub, Wikipedia if it exists), knowsAbout (3 to 10 topics that assert the company's expertise), and contactPoint with availableLanguage.
- 2. Service
- One Service per service offered. It links to the Organization through provider:
- { "@context": "https://schema.org", "@type": "Service", "@id": "https://acme.com/#service-inventory", "serviceType": "Inventory Management Platform", "provider": { "@id": "https://acme.com/#organization" }, "areaServed": [ { "@type": "Country", "name": "Portugal" }, { "@type": "Country", "name": "Spain" } ], "audience": { "@type": "BusinessAudience", "audienceType": "Retail SMB" }, "description": "Real-time inventory management SaaS platform.", "offers": { "@type": "AggregateOffer", "priceCurrency": "EUR", "lowPrice": "99", "offerCount": 3 } }
- For B2B SaaS with variable pricing, AggregateOffer with lowPrice signals the entry point without committing to an exposed price table. It works well for "from EUR X" without revealing the ceiling.
- 3. FAQPage
- A block of questions and answers in structured format. AI models prefer to cite this format because it is trivial to extract:
- { "@context": "https://schema.org", "@type": "FAQPage", "@id": "https://acme.com/#faq", "isPartOf": { "@id": "https://acme.com/#website" }, "mainEntity": [ { "@type": "Question", "name": "Does Acme integrate with ERP?", "acceptedAnswer": { "@type": "Answer", "text": "Yes. Native integration with SAP B1, Primavera and PHC." } } ] }
- Rule of thumb: 5 to 10 questions per FAQPage. More than that is noise. The questions should match real queries customers make, not queries invented by the marketing team.
- 4. BreadcrumbList
- Small but important. It helps crawlers understand the hierarchy of your pages. It produces a visible result in Google SERPs and gives structural clues to LLMs:
- { "@context": "https://schema.org", "@type": "BreadcrumbList", "@id": "https://acme.com/product/inventory#breadcrumb", "itemListElement": [ { "@type": "ListItem", "position": 1, "name": "Home", "item": "https://acme.com" }, { "@type": "ListItem", "position": 2, "name": "Product", "item": "https://acme.com/product" }, { "@type": "ListItem", "position": 3, "name": "Inventory", "item": "https://acme.com/product/inventory" } ] }
- 5. Article (or BlogPosting)
- Every blog article should have BlogPosting (more specific than Article) with the critical fields:
- { "@context": "https://schema.org", "@type": "BlogPosting", "@id": "https://acme.com/blog/integrate-sap#article", "headline": "How to integrate Acme with SAP B1 in 30 minutes", "description": "Technical integration guide via REST API.", "image": "https://acme.com/og-image.png", "datePublished": "2026-05-01", "dateModified": "2026-05-01", "author": { "@id": "https://acme.com/#organization" }, "publisher": { "@id": "https://acme.com/#organization" }, "mainEntityOfPage": { "@id": "https://acme.com/blog/integrate-sap#webpage" }, "articleSection": "Integrations", "inLanguage": "en" }
- Three details many sites get wrong: mainEntityOfPage should point at a WebPage that actually gets emitted (not an @id that never exists). articleSection helps LLMs categorise. author and publisher reference the Organization through @id instead of duplicating data.
The @graph idea (and why it matters)
If you have several entities on the same page, you can wrap them in a @graph instead of scattering multiple application/ld+json script blocks:
{
"@context": "https://schema.org",
"@graph": [
{ "@type": "Organization", "@id": "https://acme.com/#organization", ... },
{ "@type": "WebSite", "@id": "https://acme.com/#website", ... },
{ "@type": "WebPage", "@id": "https://acme.com/#webpage", ... },
{ "@type": "Service", "@id": "https://acme.com/#service-inv", ... },
{ "@type": "FAQPage", "@id": "https://acme.com/#faq", ... }
]
}The advantage: less parsing, shared context, explicit links through @id. For LLMs, a coherent graph is significantly more legible than five loose scripts.
How the five link up through @id
| Type | @id | Links to |
|---|---|---|
| Organization | #organization | Referenced by the others as provider, author, publisher |
| Service | #service-inventory | provider → #organization |
| FAQPage | #faq | isPartOf → #website |
| BreadcrumbList | /product/inventory#breadcrumb | None in the example |
| BlogPosting | /blog/integrate-sap#article | author, publisher → #organization; mainEntityOfPage → /blog/integrate-sap#webpage |
Validation is not optional
Schema without validation is a trap. Before deploying, three tools:
- validator.schema.org, checks conformance with the specification.
- Rich Results Test, tells you which of your schema is eligible for rich results in Google.
- Manual inspection in DevTools: copy the JSON, parse it, confirm the @ids are linked.
Typical mistakes we see
- Duplicated entities. Organization redeclared on every page instead of referenced by @id.
- Missing @id. The schema works, but LLMs treat each block as an independent entity.
- FAQPage with fabricated questions. The team invents SEO-friendly questions instead of copying real ones. Result: it discredits everything.
- Article with no dateModified. Models prefer to cite recent content; with no dateModified, they have no update signal.
The minimum viable, before deploying
- One Organization with a stable @id, referenced by all the other entities as provider, author, publisher
- sameAs, knowsAbout (3 to 10 topics) and contactPoint with availableLanguage
- One Service per service offered, linked to the Organization through provider
- A FAQPage with 5 to 10 questions that match real queries customers make
- A BreadcrumbList with the hierarchy of your pages
- BlogPosting on every blog article, with mainEntityOfPage pointing at a WebPage that actually gets emitted
- author and publisher referencing the Organization through @id, and a dateModified
- Validated at validator.schema.org and in the Rich Results Test
- Manual inspection in DevTools: copy the JSON, parse it, confirm the @ids are linked
Frequently asked questions
Microdata, RDFa or JSON-LD?
JSON-LD. It is the only one Google explicitly recommends, it is the cleanest to maintain (it separates semantic markup from presentation HTML), and it is what LLMs extract with the highest fidelity. Microdata and RDFa are still supported for compatibility, but they are not worth starting with in 2026.
Where should the JSON-LD go: head or body?
Either works, the specification does not require one. In the Next.js App Router, the typical pattern is to inject it with a JsonLd component inside the page's return (in the body, before the content). In static sites, the head is cleaner. In both cases, one script tag per semantic block avoids duplication.
Do I have to have @id on every entity?
It is not mandatory, but it is the difference between having a set of loose entities and having a coherent graph. With stable @ids (like https://example.com/#organization) you can reference the same entity from several pages without duplicating it. Engines and LLMs strongly prefer linked graphs.
How much schema is too much?
Irrelevant or invented schema is worse than no schema. Each type should correspond to something real on the page. Marking a pricing page as Recipe is plainly wrong; marking it as AggregateOffer makes sense. Rule: if you cannot defend why you put it there, take it out.
Sources
- schema.org, the official vocabulary.
- Google Search Central, Structured Data
- validator.schema.org

