Schema Markup for AI Search: What Actually Moves the Needle

In short: Pages that AI search engines cite are almost 3 times more likely to carry JSON-LD schema markup than pages they don't cite (Ahrefs, 2026, analysis of 6 million URLs). But when Ahrefs later added schema to 1,885 pages, the citation count didn't budge. Schema, then, is a ticket, not a lever: without it you're unreadable to the machines, yet on its own it won't push you into the answer. What follows: which types carry the signal, four ready-made JSON-LD blocks to copy, and a list of mistakes that turn markup into junk.

Schema markup (structured data, machine-readable tags in a page's code) is the language your page uses to tell machines who's who: which company stands behind the content, who the author is, when the text was published, and which public profiles it all ties back to. For classic Google search it was, for years, the ticket to rich results. For AI search it has become something more important, the machine language of trust. An AI system assembling an answer doesn't "read" your brand; it resolves entities. Schema hands it a signed ID card instead of anonymous text.

Now. A lot of marketing and few hard numbers have piled up around schema. So this article does three things: it shows what the data actually says, gives you recipes you can copy straight into your own code, and states plainly what not to do.

How schema actually reaches an AI answer

A short but important technical detail: most AI answers are not produced by an AI crawler parsing your JSON-LD at the moment of the query. Google AI Overviews and AI Mode inherit Google's index, everything Google has learned from your structured data over the years feeds into that system. ChatGPT search pulls content through Bing's index and its own. Neither OpenAI nor Perplexity states, in the official documentation for their crawlers (GPTBot, OAI-SearchBot, PerplexityBot), whether they parse JSON-LD.

Google, for its part, is officially clear: to appear in AI Overviews and AI Mode there are "no additional technical requirements" beyond the standard conditions for Search, with one explicit rule: structured data must match the page's visible text (Google Search Central, "AI features and your website").

What does that mean in practice? Schema doesn't act as a direct command to an AI system. It works indirectly, through the indexes and knowledge graphs those systems stand on. That's why the effect is slow, cumulative, and tied to consistency, not to a trick.

What the data shows

Numbers first, interpretation after:

The first four rows are correlations: pages AI cites tend to have schema. The last row is a controlled experiment, and it's a cold shower. Between August 2025 and March 2026 Ahrefs tracked 1,885 pages that had JSON-LD added, compared them against 4,000 controls, and measured: no significant lift in citations on any platform.

How do the two reconcile? Like this: schema is a hallmark of well-built pages, not a magic switch. Pages with tidy markup usually also have clear structure, sources, an author, and authority, and that's the package AI cites. Adding schema to weak content and waiting for citations is like putting on a tie and waiting for a raise. The reverse holds too: doing everything else while skipping schema means deliberately leaving a gap where 61% of cited pages have it filled.

"Schema doesn't write your answer, it signs it. AI has to know who is speaking before it decides whom to cite. That's why we treat schema as one of 15 signals we check, never as the whole strategy.", Dario Drmač, Chief Growth Officer, Risely Digital

Which types carry the signal, and why

The order is not random. This is the order in which the entity signal grows:

1. Organization + sameAs. The foundation of everything. It links your domain to public profiles (LinkedIn, social networks, directories like Clutch) and thereby gives machines proof that the "Risely Digital" on your page is the same entity as the "Risely Digital" elsewhere on the web. Without it, every other piece of markup is an orphan.

2. Person (author). AI systems check who stands behind the text, an author with a name, a job title, and a sameAs link to a real professional profile is a verifiable signal; an anonymous "admin" is not. This is the shortest bridge between schema and E-E-A-T.

3. Article with dates. datePublished and dateModified are machine-readable freshness. Recency is a citation factor in its own right, and without these fields AI has no reliable way to read it.

4. Service. For commercial pages: a machine-readable description of what you do, for whom, and where. When AI answers "who in Croatia does GEO optimization," Service markup is the closest thing your page has to a direct answer to that question. We described the whole approach on the GEO optimization service page.

5. FAQPage, with a footnote. Back in August 2023 Google restricted FAQ rich results to authoritative government and health sites. The restriction concerns display, not processing, but the data on impact is mixed: in SE Ranking's dataset, pages with FAQ schema averaged fewer citations in ChatGPT than those without it (3.6 vs 4.2), and only a hair more in AI Mode (4.9 vs 4.4). The takeaway: an FAQ section as visible content is worthwhile because it answers real customer questions; FAQPage markup is "nice to have," not a priority.

Alongside this comes a structural finding that technically isn't schema but belongs to the same discipline of machine readability: a correct heading hierarchy (H1→H2→H3, with no skipped levels) correlates with a 2.8 times higher citation rate (AirOps, 2026). We laid out the complete list of conditions a page must meet in GEO checklist.

Recipes: four JSON-LD blocks you can copy

All blocks are valid JSON-LD, drop them into <script type="application/ld+json"> in the page's <head> and swap our data for yours. The examples use real data from risely.digital, not "Acme Corp".

Recipe 1, Organization + sameAs (on every page of the site)

{
  "@context": "https://schema.org",
  "@type": "Organization",
  "@id": "https://risely.digital/#organization",
  "name": "Risely Digital",
  "url": "https://risely.digital",
  "logo": "https://risely.digital/wp-content/uploads/2024/05/RISELY-logo.svg",
  "email": "hello@risely.digital",
  "address": {
    "@type": "PostalAddress",
    "streetAddress": "Križna cesta 18",
    "addressLocality": "Zagreb",
    "addressCountry": "HR"
  },
  "sameAs": [
    "https://www.linkedin.com/company/riselydigital",
    "https://www.facebook.com/Risely.Digital.studio/",
    "https://www.instagram.com/risely.digital"
  ]
}

Into sameAs go only profiles you actually maintain, a dead profile is a weaker signal than none. Add your profile on the directories AI search engines already cite for your industry.

Recipe 2, Service (on the service page)

{
  "@context": "https://schema.org",
  "@type": "Service",
  "@id": "https://risely.digital/services/geo-optimizacija#service",
  "name": "GEO optimizacija",
  "serviceType": "Generative Engine Optimization",
  "description": "Optimizacija vidljivosti u AI tražilicama, ChatGPT, Perplexity, Gemini i Google AI odgovori, s mjesečnim mjerenjem udjela u odgovorima.",
  "provider": { "@id": "https://risely.digital/#organization" },
  "areaServed": { "@type": "Country", "name": "Hrvatska" },
  "availableLanguage": ["hr", "en"],
  "url": "https://risely.digital/services/geo-optimizacija"
}

Note the provider: instead of repeating the whole organization, it references the Organization block via @id. That's how a graph is formed, and it's precisely multiple linked types (3 or more) that carry that +13% (AirOps, 2026).

Recipe 3, Article + Person author (on every blog post)

{
  "@context": "https://schema.org",
  "@type": "Article",
  "headline": "Schema markup za AI pretraživanje: što stvarno pomiče iglu",
  "inLanguage": "hr",
  "datePublished": "2026-07-07",
  "dateModified": "2026-07-07",
  "author": {
    "@type": "Person",
    "name": "Dario Drmač",
    "jobTitle": "Chief Growth Officer",
    "worksFor": { "@id": "https://risely.digital/#organization" },
    "sameAs": ["https://www.linkedin.com/in/vas-linkedin-profil"]
  },
  "publisher": { "@id": "https://risely.digital/#organization" },
  "mainEntityOfPage": "https://risely.digital/blog/ai-search-optimization/schema-markup-za-ai-pretrazivanje"
}

In the author's sameAs put a real link to your LinkedIn profile, replace the placeholder above before publishing. And keep dateModified honest: change it when you change the content, not every week to fake freshness.

Recipe 4, FAQPage (only where an FAQ actually exists on the page)

{
  "@context": "https://schema.org",
  "@type": "FAQPage",
  "mainEntity": [
    {
      "@type": "Question",
      "name": "Pomaže li schema markup da vas AI tražilice citiraju?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "Korelacija je čvrsta: 71% stranica koje ChatGPT citira ima strukturirane podatke (SE Ranking, 2026). Kauzalnost nije dokazana, kad je Ahrefs dodao schemu na 1.885 stranica, citati se nisu pomaknuli. Schema je ulaznica, ne poluga."
      }
    },
    {
      "@type": "Question",
      "name": "Koje schema tipove dodati prve?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "Organization sa sameAs poveznicama na cijeli web, zatim Article s Person autorom na svaki tekst, pa Service na stranice usluga. 61% stranica koje AI citira koristi tri ili više schema tipova (AirOps, 2026)."
      }
    }
  ]
}

The questions and answers in the markup must exist word for word in the page's visible content, that's Google's explicit rule and the most common mistake in practice.

What NOT to do

Schema spam exists and it costs. The four mistakes we see most often:

• Markup that doesn't match the visible content. Reviews that aren't on the page, FAQ answers that exist only in the code, an "AggregateRating" without a single real rating. Google treats this as structured-data spam and can manually penalize the entire site. And AI systems, which compute trust from consistency, get exactly the opposite of the intended signal.
• FAQPage on every page. Slapping FAQ markup on pages that have no FAQ content brings nothing, the figures from SE Ranking's dataset show that FAQ schema on its own doesn't raise citation rates, and it dilutes the graph.
• Ten types for no reason. The goal isn't quantity but a connected graph: an Organization, Person, and Article that reference one another via @id are worth more than ten random blocks that don't know each other.
• Set it and forget it. Markup without validation is Russian roulette. Before publishing, run the code through a validator (validator.schema.org or Google's Rich Results Test), broken JSON is invisible, and it fails without any error message.

And one structural reminder: if your site renders content exclusively client-side (JavaScript), there's a chance AI crawlers never see your JSON-LD. Server-side rendering isn't just SEO hygiene, in the AI era it's a precondition for being readable.

How readable is your page to AI search engines?

Schema is one of 15 signals checked by the free GEO Audit, you enter a URL and get a report: what AI sees, what it's missing, and in what order to fix it.

Sources

1. Ahrefs (2026): We Tracked 1,885 Pages Adding Schema. AI Citations Barely Moved.: https://ahrefs.com/blog/schema-ai-citations/ (6M URLs: cited pages ~3× more likely to have JSON-LD; intervention on 1,885 pages: AIO −4.6%, AI Mode +2.4% n.s., ChatGPT +2.2% n.s.)
2. AirOps (2026): Structuring Content for LLMs: https://www.airops.com/report/structuring-content-for-llms (2.8× sequential headings; "rich schema" +13%)
3. AirOps (2026): The 2026 State of AI Search: https://www.airops.com/report/the-2026-state-of-ai-search (+13% for 3+ schema types; 61% of cited pages; 2.8× repeated)
4. SE Ranking (2026): Structured Data for SEO and LLMs: https://seranking.com/blog/structured-data/ (71% ChatGPT / 65% AI Mode; FAQ schema: ChatGPT 3.6 vs 4.2, AI Mode 4.9 vs 4.4; explicitly "does not imply causation")
5. Google Search Central: AI features and your website: https://developers.google.com/search/docs/appearance/ai-features ("no additional technical requirements"; structured data must match the visible text)
6. Google Search Central blog (2023-08): Changes to HowTo and FAQ rich results: https://developers.google.com/search/blog/2023/08/howto-faq-changes
7. OpenAI crawler documentation: https://developers.openai.com/api/docs/bots · Perplexity crawler documentation: https://docs.perplexity.ai/docs/resources/perplexity-crawlers (neither states JSON-LD parsing)

Keep reading: How AI engines choose who to cite · We asked 4 AI engines about Croatian brands