Schema.org's June 2026 usage-statistics dataset tells you how many sites actually use each schema type, which turns "what structured data should we add?" from an opinion into a prioritisation you can defend with data. This is a step-by-step method for auditing your own pages against real adoption: fix the near-universal basics first, add the high-value types for your content next, ignore the exotic long tail, and measure whether the markup moves your citations.
It assumes you already accept that structured data helps AI parse and cite your facts; the question here is which types, in what order. The companion reference explains the dataset itself; this is the playbook.
Step 1: Cover the near-universal basics first
Start with the types the whole web uses, because a page missing them is below baseline for machine-legibility. Schema.org's May 2026 data puts these in the top adoption bucket (roughly 10M+ domains):
| Priority | Type | What it declares |
|---|---|---|
| 1 | WebPage | What this page is |
| 2 | WebSite | The site as an entity (enables sitelinks/search box) |
| 3 | Organization | Who publishes it. Your brand identity |
| 4 | BreadcrumbList | Where the page sits in your hierarchy |
These four give an engine the who and where it needs to attribute you. If your CMS does not emit them, that is the first fix. It is the entity-clarity foundation every other signal builds on, and part of making your site agent-readable.
Step 2: Add the high-value types for your content
Next, add the types that make your specific facts extractable, but only where the content genuinely exists. The high-leverage types for AI visibility, matched to page type:
- Product / Offer: on product pages, so an engine can lift your price, availability, and attributes cleanly. This is what wins commerce answers; see writing conversational product pages and preparing your catalogue for AI agents.
- FAQPage: on genuine Q&A sections, giving engines pre-formatted question-answer pairs to quote directly.
- Article / BlogPosting: on editorial content, with author and date, feeding freshness and authorship signals.
- DefinedTerm: on glossary and definition pages, labelling each term as a defined concept.
The rule for this step is match schema to real content. Adoption data shows these types are widely supported and safe to implement; it does not license marking up content you do not have. Fake FAQs or unearned reviews get ignored at best and penalised at worst.
Step 3: Ignore the exotic long tail
Resist the temptation to add rare types for their own sake. The dataset's under-1K-domain bucket is full of specialised action types (AllocateAction, AgreeAction and the like) that almost no site uses. Usually for good reason.
A schema type is worth adding when it describes content you actually have and helps a machine lift a fact. Never because it is rare and you hope it is a secret edge. Rare usually means niche, not advantageous.
The exception is the genuinely niche type that is exactly right for your vertical (medical, legal, government). Those sit in low buckets because their audience is small, not because they are low-quality, so a specialised type that precisely fits your content is worth it even at low adoption. The test is always fit-to-content, not popularity.
Step 4: Validate and keep the markup crawlable
Structured data only counts if a machine can read it. After implementing:
- Validate every type with a schema/rich-results validator so malformed markup does not silently fail.
- Server-render the JSON-LD: if the schema only appears after JavaScript executes, some crawlers miss it, as covered in do AI assistants render JavaScript.
- Confirm crawlers are not blocked at the CDN or robots layer, or none of this is visible. See is your CDN blocking AI crawlers.
- Keep markup in sync with content: schema that contradicts the visible page is worse than none.
Machine-legibility is the same discipline as accessibility: clean, semantic, parseable structure serves both.
Step 5: Measure whether it actually moved citations
Structured data is an input, not a guarantee, so treat this as an experiment with a before and after. Note your baseline AI citations and AI-referred traffic for the relevant queries, ship the markup, then watch the trend over several weeks: not overnight. Because AI answers are volatile and crawl-to-cite lag is real, read a smoothed line rather than a single reading, the same discipline as getting cited by AI generally. If citations for the target queries rise after the markup lands and stays crawlable, the schema is earning its place.
Then keep it current: the dataset refreshes monthly and best-practice types drift, so revisit your priorities on a refresh cadence instead of treating structured data as a one-time project. Watching whether the markup you add actually changes how the engines surface, cite, and recommend your brand. Across every engine, over time. Is exactly what Buffy Intel is built to measure.