okaihub

Structured Data for AI Tools: Which Schema Types Get Used

SoftwareApplication, ItemList, FAQPage and BreadcrumbList — what each does for an AI product page, what to fill in, and the markup that gets ignored.

Published 2026-08-23 · 4 min read

Structured data does two jobs and they are usually conflated. It gives search engines explicit facts for rich results, and it gives any machine reading the page an unambiguous account of what the page is about. The second job is the one that matters for AI retrieval, and it works whether or not you ever get a rich result.

These are the types worth implementing for an AI product, in order of value.

SoftwareApplication — the product page

The primary type for any tool. It declares that the page describes a piece of software, with its category, its offer and its identity.

PropertyFill it withNote
nameYour exact product nameIdentical to the name used everywhere else — inconsistency breaks entity resolution
urlThe product's own siteNot your listing page
applicationCategoryA recognised category valueChoose the narrowest accurate one
descriptionDefinition-first, under 300 charactersThe same text as your meta description is fine and preferable
offersPrice and currency, if publishedOmit entirely rather than guessing — a wrong price is worse than none
sameAsEvery social profile and listingThe most important property for entity consistency, and the most skipped
operatingSystemWhere applicableOnly if true
aggregateRatingOnly real, sourced ratingsNever fabricate. Invented rating markup is a manual-action trigger

ItemList — the category and list page

For any page that presents a ranked or grouped set of products: category hubs, comparison pages, "top tools for X" lists. It declares that the page is a list and what the items are.

  • Each item is a ListItem with a position and a URL — and position should reflect the actual displayed order, not an arbitrary numbering.
  • The items should be the same entities as the individual product pages, linked by URL. This is how a machine learns that the list and the pages describe one catalogue.
  • A list with an honest ranking basis is more useful than one without. If the order follows a stated rule, say what the rule is on the page — an unexplained ordering is indistinguishable from an arbitrary one.

FAQPage — the highest-yield type for citation

Questions and answers in markup are extracted readily, because they are already in the shape an answer needs. This is the type most likely to result in your text appearing in a generated answer.

  • The visible text must match the markup exactly. A FAQPage schema containing questions that do not appear on the page is a guideline violation and a trust signal in the wrong direction.
  • Ask the questions buyers actually ask, not the ones you want to answer. Support conversations and listing comments are the source.
  • Answer in the first sentence. Then elaborate. The first sentence is what gets lifted.
  • Three to six questions is enough. A page with forty FAQ entries is padding and reads as such.

Declares the page's position in the site hierarchy. It is trivial to implement, it makes the hierarchy explicit rather than inferred from links, and it is what produces the path display in search results.

For a product page that is a few clicks from the homepage, this matters more than it sounds — crawl depth is a real constraint on how thoroughly a site gets read.

The supporting types

TypeUse it for
WebPage / CollectionPageDeclaring what kind of page this is, where it is not obvious from the primary type
PropertyValueIndividual specifications inside a product — pricing tier, platform support, language count
OrganizationYour company, once, with sameAs to every profile. The anchor for entity resolution
ArticleGuide and blog posts, with a real datePublished and dateModified
OfferPricing detail, where you publish it

What does not help

  • Fabricated aggregateRating. The clearest way to get a manual action, and trivially detectable.
  • Marking up content that is not visible. Hidden text with schema attached is treated as manipulation.
  • Every type at once. A page declaring itself to be six things declares itself to be nothing. One primary type plus supporting ones.
  • Schema instead of content. Markup describes a page; it does not substitute for one. A thin page with rich schema is still a thin page.
  • Review markup you did not collect. Third-party reviews require the source to be identifiable.

Implementation notes

  1. 1

    JSON-LD in the head, not microdata in the body

    Easier to generate, easier to maintain, and it does not require changing your markup structure. A script element per page, generated from the same data that renders the page.

  2. 2

    Generate it from your data layer

    Hand-written schema drifts from the page. If the product name changes in one place and not the other, you have created exactly the inconsistency the markup exists to prevent.

  3. 3

    Validate every page type once, then spot-check

    Validator tools catch syntax errors and missing required properties. They do not catch a truthful-but-wrong value, which is the more common failure.

  4. 4

    Keep sameAs maintained

    Add profiles when you create them, remove them when you abandon them. A sameAs pointing at a dead account is a negative signal.

  5. 5

    Re-check after framework upgrades

    Markup generation is easy to break silently. A build step that validates one page per type costs nothing and catches it.

Structured data is not a ranking trick. It is the difference between a machine guessing what your page is about and being told — and being told is what makes a page safe to quote.

More in SEO & GEO