Schema.org markup for AI citations: what actually matters in 2026

Schema markup is one of the most oversold pieces of the AI visibility stack. Half the agencies selling GEO packages treat structured data as the silver bullet that will get your brand cited by ChatGPT and Claude. Google's own documentation says the opposite. A December 2024 study found no correlation between schema coverage and AI citation rates. Bing's team, on the other hand, publicly confirmed that schema helps their Copilot product. The reality is messier than any of the marketing copy. This post is the honest version: what the evidence actually shows, which schema types to implement anyway, and where to stop spending hours on markup that does not move the metric.

The honest state of schema for AI in 2026

Two facts set the frame. First, Google says no special schema is needed for AI Overviews — its public documentation states "there's no special schema.org structured data that you need to add" for AI Overviews or AI Mode. Second, Bing and Google Search both say schema still helps generally. Bing confirmed in 2025 that structured data helps its LLMs understand content for Copilot, and the Google Search team separately acknowledged that structured data "gives an advantage in search results."

Soar is a community marketing agency that has run 4,200+ community campaigns across 280+ brands since 2017, and the schema question comes up on roughly every AI visibility scoping call. The answer we give clients has not changed: schema still matters for classical Search ranking, schema still matters for Bing-powered surfaces (which includes ChatGPT Search), and schema does not magically unlock AI Overview citations. There is no schema type labeled "AI visibility" you can add and expect results. The interventions that move the metric are upstream — the community signals and on-page content that AI retrieval pipelines actually evaluate at runtime.

What the evidence actually shows

The most-cited piece of evidence is a December 2024 study by Quoleady and Search Atlas that analyzed schema markup coverage against AI citation rates. The study found no correlation between how much schema a site had and how often it was cited by AI search engines. Sites with heavy schema coverage were not cited more. Sites with minimal schema coverage were not cited less. The variable that did predict citation rate was content authority and relevance to the query, not the JSON-LD payload.

A related stat is floating around from a different agency's case study, claiming that GPT-5 accuracy improves from 16 percent to 54 percent when structured data is present, a 300 percent improvement. We mention this because you will see it quoted, but we cannot verify the methodology and the number is disputed. Do not build a schema strategy around it.

The practical takeaway is that schema is not load-bearing for AI citations the way some agencies pitch it. It is load-bearing for classical ranking, which is upstream of ChatGPT Search's Bing integration, which is upstream of AI citation rates. The path exists, but schema is a second-order contributor, not a direct lever.

How LLMs actually see structured data

The technical detail most schema debates miss: LLMs do not read JSON-LD payloads directly the way a search crawler does. They receive structured data indirectly, via Data-to-Text conversion during training. When a model is trained on a page, the pipeline that prepares the training data reads the HTML, extracts the visible text, and also parses the JSON-LD and converts it into natural-language statements that get folded into the training corpus. That is the mechanism by which structured data influences what the model "knows" about a site.

This matters for two reasons. Schema contributes to training, which can influence the brand facts a model retains across sessions. But schema does not contribute at retrieval time for most of the current crop of AI systems, because retrieval pipelines look at the rendered page, not the JSON-LD. If you want your brand facts embedded in the training corpus, schema helps. If you want your page cited in an answer generated right now, schema is not the reason it will happen.

The four schema types that actually matter

Practitioners across the GEO category consistently recommend the same handful of schema types: Article, Organization, Product, Review, FAQPage, HowTo, QAPage, Author, Person, and Dataset. Of that list, four do almost all the useful work for a typical brand site, and the rest are situational or redundant.

Article goes on every blog post and editorial page. It defines the title, author, publish date, and canonical URL in a structured form that both search engines and Data-to-Text pipelines understand. This is the single most important schema type for content-heavy sites.

Organization goes on every page or once in the sitewide header. It defines the brand name, logo, social profiles, founding date, and contact info. This is the schema that teaches models "when someone asks about Acme Software, here is the canonical entity."

Product goes on every product page. It defines name, description, price, availability, and SKU. For SaaS sites, the product equivalent is SoftwareApplication. For brands with a physical product catalog, Product is non-optional and the highest-leverage schema type you can ship.

FAQPage goes on any page with a genuine question-and-answer structure. The important caveat is that Google deprecated rich-result support for FAQPage in 2023, so this schema no longer produces the expanded snippet in classical Search. It still provides useful structured context for Data-to-Text pipelines, which means it still has GEO value even though the classical-SEO rationale is gone — provided you populate it with real attributes per the Frase finding above.

What to skip

Review schema and HowTo schema are still valid but rarely move the needle for a marketing site. QAPage is usually redundant with FAQPage. Author and Person schema are worth adding if your editorial voice has named humans with credentials worth surfacing, but the lift is marginal for most brands. Dataset schema is genuinely useful if you publish datasets, which most marketing sites do not.

The schema types we explicitly tell clients to stop implementing are the obscure ones that get added by default by WordPress plugins and Shopify apps: BreadcrumbList on pages that do not need it, SiteNavigationElement, WPSideBar, and the long tail of markup that adds weight to the page without adding information the model can use. Delete them. The pattern across the 280 brands we have audited is that more than half of the JSON-LD shipped on a typical marketing site is plugin-default markup that no retrieval or training pipeline values. Cleanup yields faster pages and a cleaner entity graph.

A minimal working example

Here is the Article plus Organization payload we ship on a typical blog post. JSON-LD, embedded in the page head as a script block.

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@graph": [
    {
      "@type": "Organization",
      "@id": "https://example.com/#organization",
      "name": "Acme Software",
      "url": "https://example.com",
      "logo": "https://example.com/logo.png",
      "sameAs": [
        "https://twitter.com/acme",
        "https://www.linkedin.com/company/acme"
      ]
    },
    {
      "@type": "Article",
      "headline": "The 2026 guide to generative engine optimization",
      "author": {
        "@type": "Person",
        "name": "Jane Doe"
      },
      "publisher": {
        "@id": "https://example.com/#organization"
      },
      "datePublished": "2026-04-13",
      "dateModified": "2026-04-13",
      "mainEntityOfPage": "https://example.com/blog/geo-guide-2026"
    }
  ]
}
</script>

That is the whole thing. One script block, two entities, a shared Organization reference. Everything else on a standard blog post is optional.

The practical recommendation

Implement Article, Organization, Product, and FAQPage on the relevant page types. Delete the schema types that WordPress or Shopify added by default that do not belong to that list. Do not treat structured data as the thing that will move AI citation rates. Use the saved engineering hours to ship content that LLMs will actually want to cite, which is the subject of our post on content that AI tools are more likely to cite, and to understand how LLMs decide what to cite in the first place.

For a marketing leader, the priority order is: ship a clean classical-SEO baseline (which includes the four schema types above), then invest the rest of the budget in the levers that actually move citations — community signal on Reddit and Quora, brand search volume, and earned third-party mentions on G2 and Capterra. Schema gets a one-week project. Citation share takes a quarter.

Frequently asked questions

Conclusion

Schema markup is a classical-SEO hygiene item that matters for AI citations as a second-order effect, not a first-order lever. Google says you do not need special schema for AI Overviews. Bing says schema helps Copilot. The Quoleady study says coverage does not correlate with citation rate. The practical answer is to ship the four schema types that help across the board, skip the ones that do not, and spend the saved time on the content and community work that actually moves metrics.

Related reading

• The 2026 guide to generative engine optimization

• How to appear in Google AI Overviews

• What is an llms.txt file and should your brand have one?

• How to create content that AI tools are more likely to cite

• How LLMs decide what to cite