Which Schema.org Types Actually Matter for AI Answer Engines

What markup no longer does for you

For years, Schema.org markup was sold as a way to earn star ratings, a question accordion, or a step carousel in Google results. In August 2023, Google restricted FAQPage rich results to a narrow set of authoritative government and health sites, and shortly after removed HowTo rich results from standard search entirely. Many marketing leads drew a quick conclusion from that: markup stopped mattering. That conclusion mixes up two separate things.

A rich result is a visual treatment inside a Google results page. Structured data is a typed, machine-readable fact describing who is speaking, about what, and since when. Answer engines such as ChatGPT, Perplexity, Claude, or Google AI Overviews do not produce rich results at all: they write an answer from passages pulled out of several pages. What matters to them is not how a page looks in a results list, but how easily a passage can be extracted, attributed, and checked. That distinction is what decides, type by type, what still counts.

What an answer engine actually does with your structured data

An answer engine works, roughly, in two stages. A retrieval stage picks a small number of passages judged relevant to the question. A generation stage writes an answer from those passages and, often, cites or links a source. Schema.org markup mostly plays into the first stage, and into the attribution that follows the second: it helps identify who wrote a piece of content, which organization stands behind it, and how current the information is. Search Engine Journal has reported that, at a Search Central conference, a Google representative confirmed Gemini draws on structured data present on a page, without detailing the exact list of types involved.

Here is the passage worth quoting on its own, since it keeps its meaning outside this article: structured data does not make a piece of content more true or more interesting, it makes explicit what a human reader would otherwise infer from context, namely who wrote this, for which organization, and when. Faced with two passages that are equally good on substance, an answer engine favors the one whose origin is easiest to verify.

This is also why structured data behaves differently from ordinary on-page SEO signals. A meta description or a title tag is written for a human scanning a results page. A Schema.org property is written for a parser that never sees the page rendered at all. The two audiences do not overlap, which is exactly why a page can rank well in classic search while still being a poor source for an answer engine to cite: it may read clearly to a person while leaving its author, its organization, and its publication date entirely implicit.

The decision table: what matters, what changes nothing

The table below sorts the most common Schema.org types by their historical role in classic SEO against their real effect on citation by an answer engine. It reflects where existing guides agree, and settles the points where they contradict each other, particularly on FAQPage and HowTo.

A quick read of this table corrects a common assumption: FAQPage keeps real value for answer engines even after losing its visual treatment in Google, while HowTo, often mentioned in the same breath, adds nothing once the underlying content is not already written as a clear sequence of steps.

The Claude exception

Nearly every guide on markup and answer engines focuses on ChatGPT, Perplexity, and Google AI Overviews, and skips Claude entirely. Anthropic does not publish documentation describing how its retrieval crawlers use Schema.org, unlike Google, which has publicly confirmed that Gemini draws on structured data. Absent official documentation, the safest approach for a small company is to mark up for general content clarity rather than for one engine specifically: a well-built standalone passage, with a clearly declared author and organization identity, benefits every answer engine at once, including the ones that do not document their own retrieval mechanics. Treat every engine-specific claim in this space with some caution: none of the major AI answer providers publish a full technical account of what their retrieval pipeline reads, weighs, or discards.

A worked example: Organization and FAQPage on a service page

Here is a minimal JSON-LD setup for a service page, combining the organization's identity with a citable question-and-answer pair:

{
  "@context": "https://schema.org",
  "@type": "Organization",
  "name": "Company name",
  "url": "https://example.ch",
  "sameAs": [
    "https://www.linkedin.com/company/example",
    "https://www.instagram.com/example"
  ]
}

{
  "@context": "https://schema.org",
  "@type": "FAQPage",
  "mainEntity": [
    {
      "@type": "Question",
      "name": "How long does a GEO audit take?",
      "acceptedAnswer": {
        "@type": "Answer",
        "text": "An initial GEO audit typically takes one to two weeks: reviewing existing pages, testing a representative set of questions across several answer engines, then a prioritized list of fixes."
      }
    }
  ]
}

The second block illustrates this article's central point: the text inside acceptedAnswer.text keeps its meaning outside its original context, which is exactly the kind of passage a retrieval stage can lift and cite without rewriting it.

A five-step method for prioritizing markup

1. Mark up the entity first: Organization or LocalBusiness, with a sameAs field pointing to verified profiles (LinkedIn, trade register).
2. Add Person on author pages, with a sameAs link to a real professional profile, never a generic one.
3. Structure FAQs as JSON-LD even though the rich result no longer shows up in Google: this is the part that benefits citation by an answer engine the most.
4. Date every article consistently with datePublished and dateModified, and only update the second field after a substantive revision.
5. Validate each type with Google's Rich Results Test before publishing, to catch syntax errors that make the whole block unusable.

When Schema.org markup is not the priority

Structured markup does not fix vague content. A page with no verifiable fact, no precise definition, and no dated number does not become more citable just because it is wrapped in JSON-LD: answer engines extract a passage, they do not invent one out of markup applied to empty content. For a site that has just launched, with no track record yet, the first job is writing standalone, verifiable passages, not markup completeness.

Markup is also not the priority for a very small site, such as a single-page brochure site: the time spent adding five different types far outweighs the expected benefit, while fixing crawler access or clarifying that one page pays off more. Finally, for a sector where no local or factual answer is ever searched for through answer engines, such as a purely internal service, the return on markup stays marginal next to the underlying editorial work.

The same logic applies to timing. A company two weeks into a new domain, with a handful of published pages and no citation history anywhere, gains little from spending a day fine-tuning nine different Schema.org types. The pages themselves need to exist, and to say something specific, before the question of how they are described to a parser becomes worth solving carefully.

One practical rule of thumb helps decide: if a page still contains no precise definition, no dated number, and no section directly answering a frequent question, starting with markup amounts to decorating an empty room. The order that works best for a company just starting out is nearly always the same: content with standalone, verifiable passages first, then a clearly declared author and organization identity, and only then the rest of the markup covered in the table above.

Schema.org markup is one of six levers covered in our overview of visibility in AI answer engines: it replaces neither crawler access, covered in our article on GPTBot, ClaudeBot, and PerplexityBot, nor the machine-readable summary offered by llms.txt. For a company that wants to know where to start, the GEO Sprint begins with a concrete audit rather than a generic list of types to mark up.