1/11/2026
Structured data is not just a Google matter. schema.org is an open vocabulary read by search engines, social networks, voice assistants and the crawlers of generative engines: ChatGPT, Perplexity, Gemini, AI Overviews. This guide explains how AI engines read your markup, what it gives them, what is still unknown about its effect, and what to mark up first on a B2B website. For the fundamentals (definition, formats, types), start with our structured data guide.
How Generative Engines Read a Page
A generative engine does not "see" a page the way a visitor does. It fetches it with a crawler, extracts the text and the code, then decides which passages deserve a place in its answer. This is far from a marginal issue: over 50% of Google searches show an AI Overview (Squid Impact, 2025).
The Crawlers Behind Generative Engines
At the time of writing, each player crawls the web with its own bots: GPTBot and OAI-SearchBot for OpenAI, ClaudeBot for Anthropic, PerplexityBot for Perplexity, and Googlebot for AI Overviews and AI Mode, available in France since summer 2026. All of them read the page's HTML, and therefore the markup it contains, whether it is a JSON-LD block or Microdata attributes.
From Crawl to Cited Answer
Schematically, a page goes through five stages before it reaches a generated answer:
- crawling: the bot fetches the page, provided it is reachable and not blocked;
- extraction: the system isolates the visible text, the structure (headings, lists, tables) and the markup;
- entity identification: it works out which organisation, product, person or place the page is about;
- passage selection: it keeps the fragments that best answer the question asked;
- output: it reformulates and, depending on the engine, cites the source or not.
Markup comes into play mainly at stages 2 and 3. It is useless if the first stage fails: check that your robots.txt file does not accidentally block the AI crawlers you want to let in.
What Markup Gives an AI System
Structured data helps a system understand a page. It turns implicit cues (a layout, a byline, a mention in the footer) into explicit, standardised information. The table below maps each need of an AI system to what markup provides.
Disambiguation: Saying Which Entity You Mean
A brand that shares its name with another, software named after a city, a person quoted without a job title: these ambiguities expose you to confusion in a generated answer. An Organization connected to its website, its logo and its official profiles via sameAs makes it clear who is meant. Markup thus supports an accurate output: the right name, attached to the right organisation.
Accuracy: Standardised Facts
A "from" price, excluding VAT and monthly, shown in a box lends itself to several readings. In the markup, the same information becomes an explicit value, currency and period. The same goes for an address, opening hours, availability or an update date. The system no longer has to guess the format: it reads the fact exactly as you declared it.
What We Know, and What We Don't
Discussion of structured data and AI readily mixes findings with promises. Here is what is established at the time of writing, and what is not:
- established: AI crawlers read the HTML, and therefore the markup, and structured data describes a page's entities and facts in a shared vocabulary;
- established: schema.org is read by many systems, and display in Google is only one of its uses;
- not established: no study shows that adding markup gets a page cited by an AI;
- not established: no published statistic measures the effect of markup itself on generated answers.
No Proven Direct Effect on Citations
The available figures concern editorial structure, not markup. Pages with an H1-H2-H3 hierarchy are 2.8× more likely to be cited by AI (State of AI Search, 2025). This is a correlation about headings, not proof of an effect from structured data (see our GEO statistics).
The same study notes another trait shared by the content that gets picked up: 80% of pages cited by AI use lists (State of AI Search, 2025). Here again, it is the visible structure of the text that matters, not the JSON-LD.
What Matters Most for Accurate Output
Markup is therefore only one piece of a whole. Four factors weigh more heavily on how an AI understands and renders a page:
- explicitly named entities in the text: the brand, the offer, the author, the place;
- structured content: a single H1, readable H2 and H3 headings, lists and tables;
- consistent information between the page, its markup and the other sources that talk about you;
- reachable pages that AI crawlers can explore.
What to Mark Up First on a B2B Website
A B2B website does not need fifty types. It needs the right primary type on each template, connected to a single organisation. The table below shows which schemas to add by page type and what an AI system can take from them.
Offer Pages: Service or SoftwareApplication
On a B2B offer, the goal is to be understood correctly rather than simply displayed. Service suits a professional service (consulting, support, maintenance); SoftwareApplication suits software, with its category and platform. Both connect to the organisation via provider or publisher. Only add a rating or a price if they are visible on the page and kept up to date.
Editorial Pages and FAQs
On articles, Article or BlogPosting connected to a Person makes the author, publisher and dates explicit. For FAQs, remember that FAQ rich results have been restricted since August 2023 to authoritative government and health websites: a B2B website gets no display in Google. FAQPage still offers explicit structure for other systems, with no known measured effect. The rules and a complete example are in our guide to FAQPage markup in JSON-LD.
The Entity Graph: One Pivot Organization, Stable @id Values
Page-by-page markup, with no links between objects, leaves the system to guess that the article's publisher, the service's provider and the company on the home page are one and the same organisation. An entity graph tells it so. It rests on three principles:
- a single Organization, declared once with all its information, acting as the pivot node;
- stable
@idvalues (for example the site URL followed by#organization), reused on every page instead of copying the object; sameAslinks to official profiles: LinkedIn, Wikidata, Crunchbase, an institutional listing.
The example below connects, within a single @graph, the organisation, its website and a service it provides:
{"@context": "https://schema.org","@graph": [{"@type": "Organization","@id": "https://www.exemple.fr/#organization","name": "Exemple SAS","url": "https://www.exemple.fr/","sameAs": ["https://www.linkedin.com/company/exemple", "https://www.wikidata.org/wiki/Q000000"]}, {"@type": "WebSite","@id": "https://www.exemple.fr/#website","url": "https://www.exemple.fr/","publisher": { "@id": "https://www.exemple.fr/#organization" }}, {"@type": "Service","name": "Audit de données structurées","provider": { "@id": "https://www.exemple.fr/#organization" }}]}
The service and the website do not copy the organisation: they point to it by its @id. For other types (Article, Product, BreadcrumbList, LocalBusiness), see our JSON-LD examples by type.
Ghost Entities: When One Brand Becomes Several
A ghost entity appears when the same organisation is described differently from page to page: "Exemple SAS" on the home page, "Exemple" on the blog, "Groupe Exemple" in the footer, each with its own URL or logo. For a system that assembles information, these are potentially three distinct entities, and understanding becomes fragmented. The fix comes down to one rule: a single Organization object, a single @id, referenced everywhere.
Consistency Between the Page, the Markup and Other Sources
An AI system cross-checks what it reads. If your markup, your text and the other sources that talk about you diverge, it has to choose, and it may choose badly. Consistency applies at three levels:
- the page and its markup: the JSON-LD describes exactly what the visitor sees, no more, no less;
- pages with each other: the same name, address and key figures on the home page, the offer pages and the blog;
- the website and external sources: the profiles listed in
sameAsshow the same information as the website.
Marking up what is not visible is not just a source of confusion. Google can apply a manual action for spammy structured data and remove the rich results of the site concerned.
Errors That Distort the Output
Three frequent errors on B2B websites degrade understanding, for Google and AI alike:
- a marked-up price that differs from the displayed price, because the template reads an old field or the page was edited by hand;
- a missing author, or one reduced to an account name, while the page shows a byline and a job title;
- a frozen date: a
dateModifiedthat is never updated, or updated on every publication with no real change.
A single source of truth per page limits these gaps: no double markup from both a CMS plugin and the template, and values read from the same fields as the displayed content.
Markup, robots.txt and llms.txt: Who Does What
Three technical mechanisms and one editorial principle are often confused when discussing visibility in generative engines. Each has a precise role, and none replaces the others:
Markup makes a page understandable; the llms.txt file makes a corpus easier for models to navigate. It is a complement, not a substitute: a well-crafted llms.txt fixes neither a duplicated Organization, nor an inconsistent price, nor a page blocked in robots.txt.
Measuring Presence and Accuracy in AI Answers
Search Console reports measure eligibility for Google rich results, not what AI engines retain from your pages. For generative engines, measurement covers three questions:
- presence: for a list of target queries, do your brand or your pages appear in the answers?
- accuracy: are the name, offer, prices, address and dates rendered correctly?
- attribution: does the answer cite your website, or attribute your information to a third party?
This measurement is done query by query, over a period long enough to smooth out the variability of answers. Our GEO audit method for AI visibility explains how to build the query panel and score the answers.
Comparing Before and After by Page Group
To link a change to a markup project, compare a group of modified pages (for example the service pages) with a group left unchanged, on the same queries and over the same period. Even then, stay cautious: an engine update, new content or an external mention can explain the gap. Markup is judged first on the accuracy of the facts rendered, which is more measurable than citation frequency.
Depending on Your Situation
Depending on where you are, these resources take this guide further:
- you have just deployed or changed a template: test your structured data with the Rich Results Test;
- you are building a complete approach to AI visibility: our guide to Generative Engine Optimization (GEO);
- you are planning content that must be understood and rendered accurately: the GEO content strategy;
- you are moving a team and its SEO indicators towards generative engines: moving from SEO to GEO.
On a website with several hundred pages, the Incremys 360 GEO/SEO Audit module detects markup errors across all pages and tracks your brand's presence in AI answers, keyword by keyword.
FAQ: Structured Data and Generative Engines
Do AI engines read structured data?
Yes. Their crawlers explore the HTML of pages, and JSON-LD or Microdata markup is part of it. Structured data therefore helps the system identify a page's entities, author, dates and facts. However, no published study quantifies the effect of this markup on generated answers or on how often a page is cited.
Does markup guarantee being cited by an AI?
No. No direct effect of markup on citations has been demonstrated. Structured data supports an accurate output when a page is picked up: the right name, the right figure, the right attribution. Whether a page is selected depends first on the quality of its content, its visible structure, the consistency of its information and its accessibility to crawlers.
Which schemas should a B2B website add first?
Start with Organization and WebSite on the home page, with a stable @id and sameAs links. Then add Service or SoftwareApplication on offer pages, Article connected to a Person on the blog, then AboutPage and ContactPage. Each object points to the organisation by its @id instead of copying it, which avoids duplicates.
What is a ghost entity?
It is the same organisation declared in several forms depending on the page: a different name, URL or logo. A system assembling this information may see several distinct entities. The fix is to declare the Organization once, with a stable identifier, then reference it everywhere, including in markup generated by a CMS plugin.
Does the llms.txt file replace structured data?
No. The llms.txt file guides models towards a site's reference pages, at corpus level. Structured data describes the entities and facts of each page. The two complement each other: an llms.txt fixes neither a markup error, nor inconsistent information between two pages, nor a crawler blocked by mistake.
Should you block AI crawlers?
It is a business decision, to be made bot by bot in the robots.txt file. A blocked crawler reads neither the text nor the markup: the page then cannot serve as a source for the corresponding answer. Above all, check that no overly broad rule accidentally blocks crawlers you want to let explore your pages.
How do you measure your presence in AI answers?
Define a panel of target queries, query the generative engines regularly and record three points: brand presence, the accuracy of the facts rendered and the source cited. Then compare a group of modified pages with a control group. Do not automatically attribute a change to markup: updates and new content also play a part.
.png)
.jpeg)

.jpeg)
%2520-%2520blue.jpeg)
.avif)