Which schema.org types actually drive AI citations?

What's changing right now

Structured data has stopped being merely fuel for rich results in Google. For models generating answers it is what a title page is for a human reader: it does not replace the content, but it lets you establish quickly whose site this is, what it covers and how it relates to other entities. Be careful with promises — neither Google nor OpenAI has confirmed that schema decides citation. What is known is that it makes facts easier to read unambiguously, and where signals conflict that can be the difference.

What you can do better

Concretely, starting today

• Implement FAQPage for models, not for rich snippets. Since August 2023 Google has shown FAQ rich results almost exclusively to authoritative government and health sites — everyone else will not see them. The value that remains is real but different: ready question-answer pairs a model can quote without guessing where in a paragraph the answer ends.
• Describe what you sell with types Google actually documents — Service for services, SoftwareApplication for a product, with featureList filled with specifics rather than slogans. A model comparing you to competitors reads those fields literally.
• Use Article with strong author (Person) linking to build E-E-A-T authority, which AI considers when selecting sources.
• Do not drop fields merely because they yield no rich snippet. These are two different goals: a rich result catches the eye in Google, while entity identifiers — legal name, VAT number, sameAs, relations between companies — tell a model who is being talked about. The latter appear nowhere and do all the work when brand identity has to be resolved.

Proof from our work

Rather than numbers we have not measured, here is what we did on our own site. We gave the Organization entity a set of identifiers: legal name, VAT number in VIES format, tax ID and a D-U-N-S number. Not one of them will produce a single rich snippet — and that is the point. A model deciding “is this the same company” gets a hard handle instead of a name that is not unique. The same goes for subOrganization: dtl.pl points to FlashAI and flashai.pl points back through parentOrganization, so the relation is declared from both ends instead of hoping a model infers it from dozens of mentions in prose. Along the way we found a bug worth remembering: a code example inside an FAQ answer contained a closing script tag, which ended the application/ld+json block early. The result — the entire schema on that page stopped parsing, not just one field. And nothing signals it: an invalid block simply stays silent. Validate your own pages rather than assuming that markup present in the source is markup that works. Schema is only one of three layers and does not suffice alone: markup tells a model who you are, llms.txt hints at what is on the site, and robots.txt decides whether it may read any of it. The best-described entity counts for nothing if the crawler bounces off a disallow.