Schema.org defines more than 800 types. Five of them do measurable work for AI citation: Article, Person, Organization, FAQPage, and BreadcrumbList. We mark every client site up with exactly these five and skip nearly everything else, because twelve months of monthly citation audits say the long tail is decoration.
We have said before that exotic schema is cargo cult. The fair follow-up question is which types are not.
Here are the five, what each one actually carries, and the specific fields that matter inside each.
Why schema works differently for AI engines
For classic Google, schema earns rich results — stars, prices, expandable panels. For AI answer engines the job is different: schema is machine-readable confirmation of what the page claims in prose. The visible byline says Ryan Walker, and the Person node confirms it. The page shows an updated date, and dateModified confirms it. Engines deciding whom to cite use that agreement as a trust tiebreaker between otherwise similar sources.
That reframing explains the whole priority list. The types that matter are the ones that confirm authorship, recency, identity, and question-answer structure.
It also explains the failure case. Schema that contradicts the visible page is not neutral. It reads as either sloppiness or deceit, and both cost you.
None of this requires new tooling. Every type here ships as JSON-LD in a script tag your template already knows how to render.
Consistency first, coverage second. Now the five.
Article: the freshness and authorship carrier
Article is the workhorse. Four fields do almost all of its work: headline, datePublished, dateModified, and author pointing at a real Person node. The one rule that matters most: dateModified must match the visible last-updated stamp on the page, to the day. In our audit set, pages where the two disagreed were cited less than pages with no date markup at all.
That result surprised us enough to re-check it. It held. An engine that catches your metadata lying discounts everything else you marked up.
Use BlogPosting for posts if you like — engines treat it as Article's child, and we have measured no difference between them.
On our sites the visible stamp and the schema field render from one CMS value at build time. Agreement is not a policy anyone has to enforce. It is a property of the build.
Skip the optional fields until the four core ones are generated from the CMS and never touched by hand.
Person: the credential carrier
Person is how a byline becomes an entity. Ours carries name, jobTitle, url pointing to the about page, and sameAs links to LinkedIn and GitHub. That gives every post a machine-readable answer to the question engines increasingly ask: who wrote this, and do they exist anywhere else? A byline that resolves to a consistent person across the site and across the web is an E-E-A-T signal engines can actually verify.
Build the node once and reference it everywhere by id. Twenty sites we have audited had three different spellings of the same author across their own schema.
The one-line credential belongs in the visible byline too. Schema confirms the human-readable claim. It does not replace it.
If the author has press mentions or a recorded talk, add those URLs to sameAs as well. Third-party corroboration is the strongest identity vote available.
One author, one node, referenced everywhere. That is the entire discipline.
Organization: the entity anchor
Organization is what lets an engine resolve your brand name to one specific thing. Ours carries name, url, logo, a one-sentence description, and sameAs links to LinkedIn and Crunchbase. When a model composes an answer and considers writing the word Avakata, this node is what connects the string to an entity with a site, a history, and verifiable profiles. Brand-name citations — the ones worth the most — hang on that resolution.
Put it on every page via the site template, not just the homepage. Engines land everywhere.
The sameAs links are the working part. Each one is a vote that the entity is real and consistent elsewhere.
A quick self-test: ask a fresh chat session what your company does. If the model hedges, or blends you with a similarly named firm, the anchor is not holding yet.
If you operate under two names, pick one for schema and stay with it. Split identity is diluted identity.
FAQPage: the verbatim lift
FAQPage hands engines the exact thing they want: clean question-answer pairs, pre-chunked for quoting. Every post on this site carries three questions phrased the way a person actually asks an AI, each with a 40-to-90-word self-contained answer. In our audits, FAQ answers are the single most commonly lifted passage type — they show up nearly verbatim in Perplexity answers more often than any body paragraph.
Two rules keep it honest. Mark up only questions visibly answered on the page, and keep each answer self-contained — no pronouns pointing at earlier text.
Write questions in the user's words, not yours. "How much does it cost" beats "Our pricing philosophy" in every audit we have run.
We also phrase at least one question per page the way spoken queries arrive — full sentence, first person. The lift is small but it has been consistent across our audit set.
Three good pairs beat ten thin ones. Padding dilutes the strong answers.
BreadcrumbList: the cheap context map
BreadcrumbList is the smallest of the five: it tells an engine where a page sits in your site's hierarchy — field notes, then category, then post. That context helps engines answer category-level questions with the right page instead of a random one, and it costs about 15 minutes to add sitewide from your routing data. The measured effect in our audits is real but modest, clearly the weakest of the five.
We include it because the cost is nearly zero and it also cleans up how your URLs render in classic search.
If you are triaging, do the other four first. This one is the rounding error that still pays for its 15 minutes.
Generate it from routes. Hand-written breadcrumbs drift the day you rename a category.
The types not worth your time, and the ten-minute habit
Almost everything else measured as noise in our data: HowTo lost its rich results and shows no citation effect, Speakable never went anywhere, VideoObject on pages without video is an integrity problem, and the long tail of niche types changed nothing we could detect across twelve months of audits. Spend the saved hours writing better lead answers instead — structure in the prose beats structure in the markup.
The habit that protects all five: JSON-LD lives in one template component, every value injected from the CMS, nothing hand-edited, ever.
Then validate one page per template on each deploy with a schema validator. Ten minutes, automated in our case, and it has caught four regressions this year.
Five types, one component, one check. Schema is a Tuesday afternoon, not a strategy — and if you want the audit that proves which pages it is helping, that is what our retainer runs monthly.