Structured Data for AI: Which Schema Markup Answer Engines Actually Use

Structured data is JSON-LD in the head of a page that states what the page is about in terms a machine does not have to guess at. This is an organization, this is a software application, this is its price.

For a software vendor hoping to appear in an answer engine’s shortlist, the useful question is which of those statements any engine documents reading. The answer is narrower than the schema.org vocabulary. This page lists what is documented, what the AI Shortlist Index publishes on its own pages, and what is wishful.

Which engines document reading structured data?

Google is the only one of the four engines the Index protocol started with whose documentation names schema.org types at all. Google’s structured data documentation exists to describe Search features, and the AI surfaces sit on top of Search.

Google’s Search Central page on AI features says there are no special requirements for AI Overviews or AI Mode beyond being indexed and eligible for Search. It says structured data should match the visible content on the page.

The Gemini API’s Google Search grounding, per Google’s documentation, draws its sources from Google Search results. That puts Gemini on the same footing: whatever helps a page rank and be understood in Search is what reaches the model.

OpenAI and Perplexity

OpenAI’s crawler documentation describes GPTBot, OAI-SearchBot and ChatGPT-User and their robots.txt controls. As of this writing it names no schema.org type and does not document structured data as an input to ChatGPT’s web search.

Perplexity’s documentation is the same in kind: PerplexityBot and Perplexity-User, robots.txt, and no schema type. Nothing in it says structured data affects which pages Perplexity uses. That does not mean markup is invisible to those engines.

Our working hypothesis is that JSON-LD reaches ChatGPT and Perplexity indirectly. It is text in the HTML they fetch, so a consistent price in the markup is one more place the fact appears. It also helps the page rank in the search index each engine reads.

That is a hypothesis about an undocumented path. It is also an argument for keeping markup consistent with the visible page, since a model reading both will notice when they disagree.

Which schema types does Google read?

Within Search, Google’s structured data documentation lists the types it reads for specific features. The six a software vendor should know are in the table.

Type	What Google’s documentation says	Fit for a property-ops vendor
Organization	Read for organization details in Search, including logo, name, and sameAs profiles	The entity anchor; every vendor site should carry one
SoftwareApplication	Read for the Software App feature, with name, operating system, application category, offers, and rating fields	The product itself, on the product page
Product	Read for product snippets and merchant listings, with offers, price, and availability	Hardware bundles such as access-control readers; otherwise SoftwareApplication fits better
FAQPage	Read, but since August 2023 the FAQ rich result is shown only for well-known authoritative government and health sites	Valid markup, no visible result expected
Dataset	Read for Dataset Search, with name, description, license, and distribution	Only if you publish data, as the Index does
Article	Read for article features, with headline, author, and dates	Guides, documentation, and research pages

Six schema.org types with a documented reader at Google, what each is read for, and where each fits on a property-ops software site. Google’s supported-features list changes; check it, not this table, before shipping.

Can markup change which pages an answer uses?

Only on your own domain, and your own domain is a small share. The Index’s archive holds one run whose consulted list is visible.

On prompt mvn-best-01, in a run captured through the OpenAI API on 3 September 2026, ChatGPT named Property Meld first and Latchel second. Of the 124 URLs it consulted, 38 were on reddit.com, 16 on appfolio.com, 12 on buildium.com and 4 on propertymeld.com.

The vendor that came out first had its own domain on 4 of 124 consulted URLs. The rest were forum threads, property management system vendor pages and listicles. That run is retired from the edition and kept in the archive.

In the 25 app captures the September edition rests on, the links in ChatGPT’s maintenance answers went the other way. The top three domains were latchel.com, propertymeld.com and lula.life, and reddit.com had 6 of 184 citations. The app shows only the links in the answer, so the two lists are different measurements.

Structured data on propertymeld.com can only affect the 4. It cannot make a Reddit thread say anything different, or make an AppFolio marketplace listing more accurate.

We expect that pattern, a small own-site share and a large third-party share, to hold across engines and categories once more runs exist. That would make schema a hygiene step on the vendor’s own pages, not a lever on the sources that carry most of the answer. The measurement framework explains source dependency, and the maintenance source map shows the domains.

What should a property-ops vendor mark up?

The list is short and the order matters.

Organization on every page

Spell the name the way it appears on LinkedIn, G2 and Crunchbase. Add url, logo, sameAs to each of those profiles, and foundingDate and headquarters if they are on the About page.

This is the node that lets an engine connect the company on your site to the company on the third-party pages that carry most citations. The guide to ranking in ChatGPT and the Gemini guide both come back to that connection.

SoftwareApplication on the product page

Give it name, applicationCategory of BusinessApplication, operatingSystem as Web, and an offers block with the price model you are willing to defend on the visible page. Add a featureList that says what the product does in the words a property manager uses: work-order dispatch, vendor network, resident portal.

If you integrate with AppFolio, Buildium, Rent Manager and Yardi, say so in the visible text and in the description. The example below uses the per-unit model Property Meld publishes, per our vendor record checked 2 September 2026.

{
  "@context": "https://schema.org",
  "@type": "SoftwareApplication",
  "name": "Example Maintenance Co",
  "applicationCategory": "BusinessApplication",
  "operatingSystem": "Web",
  "url": "https://example.com/product/",
  "publisher": { "@id": "https://example.com/#organization" },
  "offers": {
    "@type": "Offer",
    "price": "1.60",
    "priceCurrency": "USD",
    "url": "https://example.com/pricing/",
    "description": "Per unit per month, $160 monthly minimum"
  },
  "featureList": "Work-order intake, vendor dispatch, resident scheduling, AppFolio and Buildium sync"
}

Everything else

Article on documentation and guides, with a real author. Dataset if you publish benchmark data, and Product only for hardware.

FAQPage if you have an FAQ, and BreadcrumbList everywhere. The schema templates carry copy-and-edit versions of each for a software company, with the required fields marked.

What markup is wishful?

Several things get sold as “structured data for AI” that no engine documents.

1. Review stars on your own product. Google’s review snippet guidelines say ratings must come from users, and self-serving reviews on an organization about itself are not eligible. A vendor marking up an AggregateRating built from its own testimonials publishes markup Google says it will not show.
2. FAQPage as an answer-engine lever. The markup is valid and harmless, and its rich result is limited to government and health sites. Nothing in any engine’s documentation says a FAQPage node raises the odds of a page being used.
3. SpeakableSpecification. Documented for news publishers and the Google Assistant, not for software pages, and absent from answer-engine documentation.
4. Schema as a ranking signal inside ChatGPT or Perplexity. Neither company documents one. A vendor told that a specific JSON-LD block will get them into an answer should ask which document says so.
5. Marking up what the page does not say. Google’s guidance is explicit that structured data must match visible content. A price in the JSON-LD that is not on the page is the case to avoid.

What does this site publish?

A site that says what markup engines read should show its own. Every page on webpossible.com is generated from front matter, and the schema comes from the same source, so it cannot disagree with the page.

On every page

An Organization node for WebPossible with an @id, the site URL, and sameAs links to the LinkedIn company page and the X account. A WebSite node, and a page node typed WebPage, Article or TechArticle with headline, author, publisher and dates on the article types. A BreadcrumbList matching the visible breadcrumb.

This page is a TechArticle. Its author is a Person node for Ryan York, with jobTitle, worksFor, knowsAbout and sameAs links to LinkedIn, X and Upwork.

On the practice and Index pages

Practice pages add a Service node per product with an Offer carrying the price in USD. The Organization carries an OfferCatalog listing the three services.

The Index hub, each category’s edition page, each source map and the prompt library carry a Dataset node. It has creator and publisher pointing at the Organization, a CC BY 4.0 license, temporalCoverage set to the edition, and a distribution entry pointing at the page’s Markdown twin.

Each vendor row carries an Organization node for the vendor, with its name and website and a subjectOf link back to the row, and nothing more. The Index publishes only facts it has verified against a source URL. Typing a vendor’s product on its behalf would be an editorial claim, so vendor rows carry no SoftwareApplication.

Validation

Pages with an FAQ carry FAQPage, with no expectation of a rich result. The build validates every node: JSON-LD must parse, each type must carry its required fields, and Organization must be present on every page, or the build fails. The reference implementation scorecard reports the result.

Why does consistency beat coverage?

A company whose name, description, founding year and integration list differ between its site, LinkedIn, G2 and Crunchbase has a bigger problem than missing markup. The Index will not record a vendor fact without a source URL and a verification date. An engine reading four pages that disagree has the same problem with less patience.

Pick one legal name and one short description, and put them in the Organization node and on every profile. When the product is renamed or the company is acquired, update every profile in the same week. Property Meld’s acquisition of Mezo in January 2025 shows it happens in this category.

Add the old name as an alternateName in the markup rather than deleting it. The Index’s own extraction matches aliases for that reason, and the entity consistency guide covers the full audit.

Limits, and what changes month to month

Google’s supported-features list is a moving target. Features were removed in 2023 and again in 2025, and the AI features documentation is newer than either. Re-read the current list before any structured data work, and treat the table on this page as a snapshot dated in the changelog.

The Index cannot measure structured data’s effect directly. It records who appears in each answer and which pages it links; it does not see markup. Any statement here about markup causing a mention is a hypothesis and is labeled as one.

What the Index can show over editions is whether a vendor’s own domain share of citations moves. That is the number a markup change could plausibly touch.

The Index methodology is the source of truth for how citations are captured and classified, and the AI search optimization overview sits above this page. If you want your markup, your profiles and your category’s third-party sources worked on in order, the 90-day sprint is the scoped version of that.

Changelog

• 2026-10-28: Reformatted with a short answer, question headings, the wishful-markup list as a numbered list, and an FAQ. The vendor-markup section now comes before the inventory of this site’s own schema. Documentation readings unchanged.
• 2026-09-26: The September figures now cite the 25 ChatGPT app captures per category; the API run of 3 September 2026 is described as retired from the edition and archived.
• 2026-09-03: Rebuilt for the AI Shortlist Index scope. Replaced the earlier type table with the six types Google documents reading. Added the engine-by-engine documentation check, the reading of the 2026-09 run, the inventory of the Index’s own schema, and the wishful-markup section.
• 2026-06-10: First published as a general-purpose page on structured data for AI, without engine documentation or Index data.