Structured Data as AI Comprehension Layer: Confirmed and Measurable

Structured data took on a new role in 2025 and 2026. For a long time, it was seen mainly as a tool for rich results: star ratings, FAQ dropdowns, or recipe cards in Google Search. Now Microsoft officially confirms that the same markup also helps its generative systems understand content. Google confirms an advantage in search. This article explains what structured data is technically, why it has become a comprehension layer for AI systems, which of these claims are proven and which are not, and how B2B companies in the DACH region can deliver it cleanly and measurably in a modern Next.js architecture. The focus is on reliable statements rather than speculation.

Why structured data matters as an AI comprehension layer

Structured data is machine-readable additional information that a website provides to state what its content means. Instead of showing a phone number only as text, the markup tells a machine explicitly: this is a phone number, this is a product price, this is an organization with this address. This layer of meaning has always been useful for search engines. What is new is that generative AI systems can use the same layer to extract entities and facts more reliably.

This matters because search behavior is shifting. In Germany, according to a representative Bitkom survey, 67% of people aged 16 and over use generative AI at least occasionally, up from 40% in summer 2024 (data for Germany, Bitkom 2025). The most used services are ChatGPT at 43%, Microsoft Copilot at 39%, and Google Gemini at 28%.

Internationally, OpenAI reported 800 million weekly active users for ChatGPT in October 2025 (international figure, OpenAI). In Austria, digital reach is high enough for this shift to take full effect: 8.69 million internet users at 95.3% penetration at the start of 2025 (data for Austria, DataReportal). If you want to appear as a source in these systems, you have to provide your content in a way that machines can interpret reliably. That is exactly what structured data does.

An honest assessment matters from the start. Microsoft has officially confirmed that structured data helps generative systems. Google confirms an advantage in search but does not say explicitly that markup directly feeds generative answers. A direct, causal effect on visibility in ChatGPT or Perplexity, by contrast, is not supported by independent, peer-reviewed studies. This article keeps these levels strictly apart.

What structured data is technically and why JSON-LD is the recommended format

Structured data is based on a shared vocabulary and an output format. In most cases, the vocabulary is Schema.org, a community-maintained collection of types and properties. Google recommends JSON-LD as the format.

• Schema.org vocabulary: Schema.org defines types such as Organization, Product, or Article and their properties such as name, price, or author. The vocabulary determines which terms exist and how they relate to each other. That makes it the dictionary that search engines, and increasingly AI systems, rely on.
• JSON-LD as syntax: JSON-LD (JavaScript Object Notation for Linked Data) is a script block in the HTML that delivers the markup as a JSON object. The key fields are @context, which points to the vocabulary (usually schema.org), and @type, which sets the entity type. The advantage: the markup sits bundled in one block, separate from the visible HTML, which makes maintenance and validation easier.
• Microdata and RDFa: Both older formats weave the markup directly into the HTML tags of the visible content. That is more error-prone and harder to maintain.

JSON-LD is technically more convenient and the format Google recommends. According to the HTTP Archive's Web Almanac 2024, JSON-LD was present on 41% of pages, up from 34% in 2022, making it the fastest-growing of the three formats (international crawl data, HTTP Archive). The most common format, however, is still RDFa, on over 60% of pages, mostly on older CMS platforms. Within JSON-LD, some types are widespread and others are rare.

In the same dataset, the WebSite schema leads with 12.73% of mobile pages, followed by Organization at 7.16%, BreadcrumbList at 5.66%, and LocalBusiness at 3.97%, while the Product schema appears on only 0.77% of pages (international crawl data, mobile, HTTP Archive 2024). For B2B sites, the lesson is clear: the basic types are standard, but more specific markup is rare enough to give you an edge.

The AI comprehension layer: what is officially confirmed and what is not

The core of the topic is whether and how AI systems use structured data. Precision is essential here, because you have to distinguish between official confirmation, qualified statements, and unsupported assumptions.

Microsoft is the clearest source. At SMX Munich in March 2025, Fabrice Canel, Principal Product Manager at Microsoft Bing, confirmed that schema markup helps Microsoft's LLMs understand content (Search Engine Land, quoting Fabrice Canel, Microsoft Bing). Microsoft added that generative systems particularly value fresh content, among other things as a check against their training data. This statement directly concerns Copilot and the Bing index. Whether it also applies to ChatGPT Search is an open question: OpenAI names third-party search providers as a source but does not disclose how large Bing's share is.

Google is more nuanced. In April 2025, the Google Search team stated that structured data provides an advantage in search results (Search Engine Land). In its guide to AI Overviews and AI Mode from May 21, 2025, Google writes that structured data provides information in machine-readable form that its systems consider. Google does not say explicitly, however, that markup directly feeds generative answers. For purely generative systems such as ChatGPT or Perplexity, there is so far no comparably clear official confirmation that schema markup is processed as a separate signal.

This results in three levels of evidence:

• Officially confirmed: Microsoft Bing uses schema for its LLMs (Copilot), as stated at SMX Munich in March 2025.
• Confirmed with qualifications: Google says structured data provides an advantage in search (April 2025) and is considered by its systems (May 2025). Google does not say that markup directly feeds generative answers.
• Unconfirmed: A standalone, causal schema effect on ChatGPT or Perplexity is not officially confirmed and not supported by peer-reviewed studies.

This distinction is the most important message of the article. Structured data is a well-founded measure, partly backed officially by two major providers, not a guaranteed lever with proven causal impact in every AI interface.

NLWeb and the agentic web: Schema.org becomes the agent interface

One concrete development shows how closely structured data and AI are converging. On May 19, 2025, Microsoft introduced the NLWeb project. It combines existing markup with language models into a natural-language interface. According to Microsoft, NLWeb uses semi-structured formats such as Schema.org and RSS that websites already publish and combines them with LLM-powered tools to create natural-language interfaces for humans and AI agents (Microsoft Source, May 2025).

The decisive point is the bridge to the agentic web. According to Microsoft, every NLWeb instance is also a Model Context Protocol (MCP) server, which lets websites make their content discoverable and accessible to agents and other participants in the MCP ecosystem (Microsoft Source, May 2025). MCP is the standard through which AI agents talk to external data sources and tools. NLWeb was developed by R.V. Guha, who joined Microsoft and is regarded as the creator of the widely used web standards RSS, RDF, and Schema.org. That the same person is behind both Schema.org and NLWeb underlines the direction: the vocabulary that has made websites machine-readable for years is becoming the data foundation for agents.

For how websites model their data, this means a strategic shift. If you annotate your content cleanly and completely with Schema.org, you also lay the groundwork for AI agents to query that content in a structured way. Markup thus moves from a mere SEO detail to part of the data architecture.

ChatGPT Shopping and agentic commerce: feed specification meets schema

Agentic commerce is another area where structured data is gaining importance. With ChatGPT moving into product-related queries, new requirements for product data are emerging. A clear distinction of terms matters here, because two different mechanisms are at play.

• Schema.org/Product on the page: This is the classic product markup embedded in the HTML. It describes name, price, availability, and other properties in machine-readable form directly on the product page. It remains the basis for organic visibility and for crawlers that visit the page.
• Product feed specification: In addition, there are structured product feeds through which merchants provide their data directly and in bundled form, instead of only placing it on individual pages. Such feeds use their own required and recommended fields and allow frequent updates, for example of prices and inventory levels.

The two mechanisms complement each other. On-page schema ensures consistency and crawlability, while the feed ensures freshness and bundled data delivery. For B2B providers with a product catalog, this means product data should be consistent and complete in both places. Specific required fields and update frequencies depend on each provider's specification and change quickly, so always check them against the official documentation. Generally valid figures on this do not currently exist in a reliable, citable form and should not be taken from secondhand sources.

Which schema types matter and which ones Google has retired

Not every schema type is equally valuable, and the field is in motion. In recent years, Google has restricted or discontinued several rich result types, most visibly FAQ rich results, which are now shown only for a few authoritative sources. The distinction matters: when a rich result is retired, that does not necessarily mean the underlying markup becomes worthless for machine understanding. It only loses its guaranteed visual display in classic search.

Types that clearly describe entities and content structure are particularly relevant for AI visibility:

• Organization: Describes the company as an entity with name, logo, address, and connections. It is the foundation for being recognized as a distinct entity and belongs on every site.
• Article: Marks up editorial content with author, date, and headline. It helps systems attribute content, source, and freshness.
• Product: Describes products with price, availability, and properties. It is the basis for product-related answers and commerce scenarios.
• BreadcrumbList: Maps the navigation hierarchy and helps place a page within the overall context of the website.
• FAQPage and HowTo: Structure questions and step-by-step instructions. Even though their rich results have been restricted in classic search, the structured form remains a clear extraction signal for content that directly answers a question.

The lesson is to choose schema types based on machine understanding, not solely on the current rich result display. Google can change the latter at any time, while the value of correctly described content remains.

Implementation in the Next.js SEO architecture

For modern B2B sites built with Next.js, structured data can be delivered cleanly and maintainably. What matters is that the markup is rendered server-side and matches the visible content exactly.

• Render JSON-LD server-side: The markup belongs in the initial server-rendered HTML, not in JavaScript that loads later on the client. That way it is reliably visible to all crawlers, including those that execute no JavaScript or only limited JavaScript. In Next.js, you can output JSON-LD directly in page or layout rendering.
• Generate schema dynamically per route: Instead of static blocks, the markup should be generated from the same data sources as the visible content, for example the CMS or the product database. That keeps markup and content in sync automatically, and every route gets the right schema.
• Keep markup and visible content consistent: Every detail marked up must also be visible on the page. Markup that claims content the user cannot see violates search engines' spam policies.
• Build validation into CI: Schema checks belong in the continuous integration pipeline, so faulty or incomplete markup is caught before deployment rather than in production. That keeps quality stable across many routes and over time.

This architecture ensures that structured data is not a bolted-on afterthought but an integral, tested part of every page.

Data delivery and AI crawler management

Structured data only has an effect if the relevant systems actually crawl the pages and keep them up to date. This is where crawl control and fast updates come in.

• Configure robots.txt deliberately: With robots.txt, you control which crawlers get access, including AI-specific bots. If you want to be visible in AI systems, you have to allow the corresponding crawlers instead of blocking them across the board. This is a strategic decision between reach and control.
• IndexNow for fast updates: IndexNow lets you actively notify participating search engines, including Bing, about changes instead of waiting for the next crawl. That matters directly for Copilot, because Microsoft stresses that its generative systems value fresh content.
• Don't underestimate the Bing index: According to Bitkom, 39% of people in Germany use Microsoft Copilot, barely fewer than use ChatGPT at 43%. For Bing and Copilot specifically, Microsoft has confirmed that schema markup helps.

The economics of AI crawling also need to be considered. According to a Cloudflare analysis, the ratio of crawls to referred visitors in July 2025 was 38,065.7 to 1 for Anthropic, 1,091.4 to 1 for OpenAI, and 5.4 to 1 for Google (international data, Cloudflare). The same analysis shows that around 80% of AI crawling serves training, compared with 18% for search and 2% for user actions (international data, Cloudflare). This asymmetry between intensive crawling and few referrals is one reason to manage crawl efficiency and access rights deliberately instead of letting every bot work without limits.

GDPR and consent when delivering structured data

In the DACH region, data protection is a mandatory part of the picture. Structured data is generally unproblematic as long as it only makes information machine-readable that is public and factual anyway, such as company name, address, product data, or article metadata. This information is published deliberately and usually contains no third-party personal data that needs protection.

Caution is needed with personal data. If, for example, a natural person is marked up with name, photo, or contact details, the usual GDPR requirements for legal basis and purpose limitation apply. Markup also spreads such data in machine-readable form, which makes it easier to reuse and should factor into the assessment.

The separation between markup and consent mechanics is important. JSON-LD itself sets no cookies and is not tracking. It therefore does not belong behind a consent gate and may be delivered regardless of consent, because it is an integral part of the content. When dealing with AI crawlers, control remains a matter of access rights via robots.txt, not of user consent. Personal fields should still be marked up sparingly and only with a clear legal basis.

Common mistakes and quality assurance

Most problems with structured data come not from missing markup but from faulty or incomplete markup. The following mistakes are especially common.

• Missing required properties: If required fields of a type are missing, the markup can lose its effect and a potential rich result is lost. Check for each type which properties are required and which are recommended, and fill them in completely.
• Markup without a visible counterpart: Marking up content that is not visible on the page violates search engines' spam policies and can lead to demotion. Markup and visible content must match.
• Maintaining outdated or retired types: Creating markup for rich results that Google no longer shows ties up effort with no return in classic search. Check regularly which types are still supported and set your priorities accordingly.
• Inconsistency between sources: If on-page markup, feed, and visible content provide different information, such as different prices, trust problems arise. All sources must be fed from the same data foundation.
• Setting markup once and forgetting it: Schema.org, provider requirements, and supported rich results change. Without ongoing maintenance, the markup becomes outdated and loses its effect. Regular review is part of operations.

The most effective safeguard is automated, recurring validation, supplemented by manual spot checks of the most important page types.

Measurability and GEO evidence: how to check the impact

The central question is how the impact of structured data can be demonstrated at all. It helps to separate technical validation, classic search measurement, and AI visibility.

• Technical validation: With the Rich Results Test and similar validators, you can check whether the markup is syntactically correct, contains all required fields, and is eligible for a rich result. That is the basic prerequisite but says nothing yet about the actual impact.
• Classic search measurement: Google Search Console shows which structured data types were detected, whether there are errors, and how impressions and clicks of the affected pages develop. This is the most reliable data source for impact in classic search.
• AI citation tracking: For visibility in AI answers, you can monitor mentions and citations in systems such as ChatGPT or Perplexity. This kind of measurement is still young, the methods are less standardized, and a clean causal link to schema markup is hard to establish.

For optimizing content for AI visibility, there is early scientific evidence, but it concerns how text is written, not schema markup. A research team led by Princeton University presented the Generative Engine Optimization method at the KDD 2024 conference and showed that GEO can increase visibility in generative answers by up to 40% (international study, Princeton et al., KDD 2024). The most effective methods were adding sources, quotations, and statistics, which can increase a source's visibility by over 40%.

In a real-world evaluation on Perplexity.ai, strong methods achieved improvements of up to 37% on one of the two metrics, measured on a benchmark of 10,000 queries and nine optimization methods. These findings concern how content is written, meaning well-documented texts backed by sources and figures. They are a strong argument for fact-dense content, but not proof of a causal effect of schema markup. This gap is the honest counterpoint: there are no peer-reviewed studies to date on the direct effect of structured data on AI visibility.

Further reading and conclusion

Structured data has evolved from a mere rich-results tool into a comprehension layer that can help generative systems capture content and entities. Microsoft has officially confirmed its use for its LLMs, Google has confirmed the advantage of structured data in search with qualifications, and NLWeb shows how Schema.org is becoming the data foundation for AI agents. At the same time, the honest caveat applies: a direct, causal effect on purely generative systems such as ChatGPT or Perplexity is not supported by peer-reviewed studies.

For B2B companies in the DACH region, this points to a pragmatic course. Structured data is a well-founded, technically manageable measure with clear benefits in search and plausible benefits for AI systems, without inflated promises. Sensible next steps are an audit of your existing schema types with the Rich Results Test, server-side rendering of complete, consistent JSON-LD in the Next.js architecture, anchoring validation in CI, and deliberate AI crawler management via robots.txt and IndexNow. That turns structured data into a reliable, measurable building block of technical visibility instead of wishful thinking.