Schema Markup for AI Search Engines: What ChatGPT, Perplexity, and Gemini Read

In the age of AI search, schema markup is treated as both a miracle cure and a myth. One camp sells structured data as a ticket into ChatGPT and Perplexity, the other considers it irrelevant once language models read the content. Both camps are wrong. In 2026, the evidence is far more nuanced and partly contradictory. This article sorts out what schema markup demonstrably does for AI search systems, where it is overrated, and what role it plays for B2B visibility in the DACH region. The focus is on verifiable data instead of promises.

Why schema markup matters for AI visibility

AI search is no longer a fringe phenomenon. According to Semrush, the AI Overviews trigger rate in Google Search rose from 6.49 percent of queries in January 2025 to a peak of 24.61 percent in July and settled at 15.69 percent in November, measured across more than 10 million keywords. On identical keywords, the zero-click rate fell from 33.75 percent to 31.53 percent, according to the same analysis. A growing share of searches therefore ends in a generated answer instead of a classic results list.

Traffic from these systems is high quality. According to an international analysis by Visibility Labs covering 94 e-commerce brands, ChatGPT traffic converted at 1.81 percent, 31 percent higher than non-branded organic traffic at 1.39 percent. The comparison covered 9.46 million organic sessions and 135,000 ChatGPT referral sessions over twelve months. If you show up in AI answers, you gain reach and purchase-ready users.

In this environment, schema markup is the machine-readable layer of a website. It translates content into an unambiguous vocabulary that machines understand without room for interpretation. The central question is: do AI search systems read this vocabulary at all, and does it make a measurable difference?

How AI systems process content and where schema comes in

Modern AI search mostly relies on RAG pipelines (retrieval-augmented generation). In simplified terms, three steps take place: the system retrieves relevant sources, extracts their content, and generates an answer from it. Schema markup, usually embedded in the HTML as JSON-LD, can help in the retrieval and extraction phases by naming entities clearly: who the author is, what the product is, which question this section answers.

Two functions need to be kept apart. Schema is primarily a tool for entity understanding, not a direct ranking or citation factor. It helps machines classify content correctly and establish connections between entities. It does not, however, force a language model to cite a source.

JSON-LD is the dominant format. According to the Web Almanac 2024, its adoption grew from 34 percent of crawled pages in 2022 to 41 percent in 2024, which makes it the fastest-growing structured data format and the one Google prefers. According to the same source, the most common types are WebSite at 12.73 percent, Organization at 7.16 percent, BreadcrumbList at 5.66 percent, and LocalBusiness at 3.97 percent of mobile pages. This broad, crawler-readable adoption is a key reason why schema remains the more dependable lever compared with newer proposals such as llms.txt: structured data is already part of the existing crawling ecosystem, while most AI crawlers have so far ignored llms.txt.

Conflicting evidence

This is where it gets uncomfortable. Reputable studies contradict each other, and that is the most important insight for practitioners.

The critical position (Ahrefs): Ahrefs tracked 1,885 pages that added JSON-LD schema between August 2025 and March 2026 and compared them with 4,000 control pages in a difference-in-differences test. The result: Google AI Overviews minus 4.6 percent (small but statistically significant), AI Mode plus 2.4 percent, and ChatGPT plus 2.2 percent (neither statistically distinguishable from zero). The authors' conclusion is clear: adding schema produced no meaningful increase in citations on any platform.

The optimistic position (Otterly): After a schema implementation across 319 prompts in the US market, Otterly.AI reports a very different picture: AI Overviews plus 611 percent, AI Mode plus 42 percent, but at the same time ChatGPT minus 71 percent, Copilot minus 64 percent, Gemini minus 35 percent, and Perplexity unchanged. The accompanying fetch test is decisive: only 1 of 7 AI platforms, namely Gemini, could correctly read the raw JSON-LD. Otterly, too, concludes that schema is an SEO lever and not a GEO growth lever.

The official position (Google): In its AI Optimization Guide, Google itself states plainly that structured data is not required for generative AI search and no special schema.org markup needs to be added, but it still makes sense to use it as part of the overall SEO strategy because it supports eligibility for rich results. Google explicitly frames optimization for AI search as a continuation of SEO.

What practitioners should take from this: schema is not a reliable citation booster for AI systems. The opposite directions of the Ahrefs and Otterly results and the weak statistical significance show that the effect is small and platform-dependent. Schema does remain an established SEO foundation, though, and through AI Overviews, Google's classic search is closely tied to the AI answer.

Which platforms actually read schema

The assumption that AI systems process structured data across the board does not hold up. Otterly's fetch test is clear here: 6 of 7 tested platforms could not retrieve or correctly interpret schema markup when asked directly; only Gemini returned the correct JSON-LD. The platforms tested were ChatGPT, Google AI Overviews, Google AI Mode, Perplexity, Microsoft Copilot, Gemini, and Claude.

That gives a realistic picture of the platform landscape:

Google AI Overviews, AI Mode, and Gemini: These systems are closest to schema because they build on Google's index, which processes structured data anyway. Gemini was the only system that could read JSON-LD directly.

ChatGPT, Perplexity, Claude, and Copilot: These systems mostly rely on the visible, rendered page content rather than the embedded markup. For them, what counts is what appears in the readable text.

Microsoft Bing and Copilot, with a caveat: There is an important official counterpoint here. Fabrice Canel of Microsoft Bing confirmed at SMX Munich in March 2025 that schema markup helps Microsoft's LLMs understand content. Canel also recommended pushing fresh content via the IndexNow API, because generative AI particularly values current content, including as a check against LLM training data. Microsoft has therefore officially confirmed that it uses schema, while the Otterly test found no positive effect for Copilot. This tension between official statements and measured impact is typical of the field.

The practical consequence: schema discipline pays off most in the Google and Microsoft ecosystems. For the other systems, clean, clearly worded plain text matters more than the markup.

NLWeb, MCP, and the agentic web

Beyond today's citation question, a structural role for Schema.org is emerging. Microsoft's NLWeb initiative uses Schema.org as the foundation for turning websites into conversational interfaces and combines this with the Model Context Protocol (MCP), the emerging standard for communication between AI agents and data sources. The idea: a website structured with schema can act as an endpoint that an AI agent queries directly.

This becomes especially relevant for agentic commerce. As AI agents begin to research and compare products and initiate purchases, they will need machine-readable product, price, and availability data. That is exactly what Product schema provides. Here the rationale for schema shifts from citation to transaction capability: structured product data becomes the prerequisite for an agent to capture an offer correctly in the first place and move it into an AI-driven checkout. For B2B providers with complex services, this is a strategic argument for structuring product and organization data cleanly, even if the immediate citation effect is small today.

Which schema types matter for AI visibility

Not every schema type is equally valuable. For AI visibility and for the connection to classic SEO, these types and properties are particularly relevant:

Organization: Defines the company as an entity, including name, logo, contact details, and profiles. With 7.16 percent adoption according to the Web Almanac, it is one of the most common types and the basis of entity understanding.

Person: Describes the creator of a piece of content. An Article points to this person via the author property. Schema.org has no “Author” type. This link between content and creator supports E-E-A-T signals (experience, expertise, authoritativeness, trustworthiness). In B2B in particular, verifiable authorship is a differentiator.

sameAs property for entity linking: Not a type of its own, but a property you set in Organization or Person, for example. It connects your own entity with authoritative references such as Wikidata, LinkedIn, or industry directories. These links help machines resolve the identity unambiguously and anchor it in the entity graph.

FAQPage: Structures questions and answers in a machine-readable way. Its benefit is disputed because many AI systems read the visible text anyway, but the format forces clearly answered questions, which makes AI extraction easier.

Article and Product: Mark up editorial content and offers, respectively. Product schema is the basis for agentic commerce; Article schema supports the classification of expert content.

What matters is their correctness and consistency with the visible content, not the number of types.

Schema versus content: what actually moves the needle

If you take AI visibility seriously, set your priorities right. The Princeton GEO study provides the most important guidance here. It shows that GEO can increase a source's visibility in generative answers by up to 40 percent, with the strongest levers being content-based: citing sources, adding statistics, including quotations, plus authority and linguistic clarity, not technical markup.

This matches the schema studies. While schema showed at best marginal and contradictory effects in the Ahrefs and Otterly tests, verifiable, well-structured content measurably increases the likelihood of being cited. The practical consequence is clear: content first, markup as an amplifier.

Both work together, though. Front-loading, meaning placing the core statement right at the start of a section, makes extraction easier for RAG systems. Schema can additionally mark up this well-structured content unambiguously. So if you first build a clearly argued page backed by data and sources and then annotate it cleanly with schema, you use both levers in the right order.

Common mistakes with schema for AI

Poor or misused markup does more harm than good. These mistakes are especially common:

Markup not backed by visible content: Schema that claims information not visible on the page violates Google's guidelines and can lead to manual actions. Markup and visible text must match.

Mistaking schema for a substitute for content: The research is clear that markup does not rescue weak content. If you rely on FAQPage schema but provide no substantive answers, you gain nothing.

Missing entity linking: Organization schema without sameAs references leaves the entity isolated. Without a connection to authoritative references, machine identity resolution remains weak.

Outdated or incorrect data: Wrong prices, old opening hours, or syntax errors in the JSON-LD undermine trust and can prevent display in rich results and AI answers. Schema needs maintenance.

Confusing llms.txt with structured data: Some teams invest in llms.txt and neglect schema. Since AI crawlers largely ignore llms.txt but process structured data widely, this prioritization is backwards.

Measurement and tooling for AI visibility

Without measurement, every schema strategy remains speculation. Three levels are relevant:

Schema validation: Google's Rich Results Test and the Schema.org Validator check whether the markup is syntactically correct and can be processed by Google. This is the basic technical check before every deployment.

AI referral traffic in GA4: Referrals from ChatGPT, Perplexity, or Copilot can be isolated in GA4 via referral sources and custom channel groups. This shows how much traffic and which conversions actually come from AI systems. The Visibility Labs finding of a 1.81 percent conversion rate for ChatGPT traffic shows that this separation is worth it.

Brand mention and citation monitoring: Tools such as Otterly, Ahrefs Brand Radar, and Semrush track when and how a brand is cited in AI answers. In a GEO context, brand mentions often count more than classic backlinks because AI systems recognize and recommend brands as entities, even without a linking source.

Choosing the right metric matters: schema effects show up primarily in Google rich results and AI Overviews, not in ChatGPT citations. If you measure the success of markup by ChatGPT mentions, you are measuring in the wrong place.

DACH context and next steps

For the Austrian B2B market, AI maturity is decisive. According to Statistics Austria, 20 percent of Austrian companies with 10 or more employees used AI technologies in 2024, up from just under 11 percent the year before, which means the share nearly doubled in a single year. According to the same survey, usage is highest in information and communication at 61 percent. Among companies using AI, 65 percent use it for text recognition and processing and 41 percent for language generation. The target audience in Austrian B2B is therefore increasingly working with language models itself, which raises the importance of AI-readable content.

This leads to the central recommendation for B2B providers: first build verifiable, clearly structured content with statistics, sources, and front-loaded key statements. Then annotate it consistently with JSON-LD in the head, with a focus on Organization, Person (linked to the article via the author property), and the sameAs property, so that your entity is anchored unambiguously in the graph. Validate the markup with the Rich Results Test, keep it up to date, and measure AI referral traffic separately in GA4. So schema is not a promise of growth in AI citations. It is solid, long-term infrastructure for the Google and Microsoft ecosystems and preparation for an agentic web in which machine-readable data becomes a foundation for doing business.