Schema Markup for AI Search

What schema does and does not do

Structured data has been sold twice: first as a way to win rich results in search, and now as an AI ranking lever. The first was true and is now less true. The second is not quite right either, and the difference matters if you are deciding where to spend a week.

Schema does one thing well. It removes ambiguity. It states, in a form no parser has to guess at, what a page is, who published it, when, and which real-world entity it belongs to. For a system deciding whether your page answers a question and whether your brand is the brand it means, that is genuinely useful.

The correction most guides have not made

Google retired FAQ rich results on 7 May 2026. The expanding accordion under a search result is gone. A great deal of published advice still recommends FAQ schema on the strength of a visual feature that no longer exists, and some reporting still counts rich results as a KPI.

Add FAQPage anyway, for a different and better reason: a question paired with a self-contained answer is precisely the shape a retrieval system lifts. You are no longer buying an accordion. You are formatting your content the way the thing reading it wants it.

The types that actually earn their place

Type	What it does for a retrieval system
Organization	Entity resolution. Tells a system which company you are, with sameAs links to the profiles that corroborate it. This is the highest-value single block for most sites, and the most often missing.
Article or WebPage	Establishes authorship and a real datePublished and dateModified. Freshness is an input to selection, and a page with no date is harder to trust.
FAQPage	Question and answer pairs in the exact shape answer engines extract. No longer a rich result, still the right format.
HowTo	Ordered steps, machine-readable. Strong where the query is procedural.
BreadcrumbList	Places a page in a hierarchy, which helps a system understand scope and topical context.
Product and Service	Only where they genuinely apply. Marking a blog post as a Product is the sort of thing that gets a site distrusted.

Organization markup, which is the one to fix first

Entity resolution is the quiet foundation of AI visibility. Before a system can decide whether to cite you, it has to decide which "Acme" you are among the several that exist. sameAs is how you tell it, by pointing at the independent profiles that already describe you.

This is also where structured data meets the finding that off-site mentions correlate roughly three times more strongly with being recommended than backlinks do. Your markup asserts the identity. Third-party sources corroborate it. Neither works nearly as well alone.

The technical trap that silently voids all of it

Most AI crawlers do not execute JavaScript. They fetch the HTML your server sends and read that. Schema injected client-side, by a tag manager or a framework hydrating on load, is invisible to them however correct it looks in a browser inspector.

The check takes ten seconds and is worth doing before anything else: fetch your own page the way a crawler does, without a browser, and look for your JSON-LD in what comes back. If it is not there, none of your markup exists as far as most AI systems are concerned.

Four rules that keep markup from doing harm

1. Only mark up what is visible on the page. Schema describing content a reader cannot see is the definition of misleading markup and it is the fastest route to being distrusted.
2. Never invent aggregateRating or review counts. Fabricated proof is a liability, not a lever, and it is the single most common cause of a manual penalty in this area.
3. Keep dates honest. A dateModified that updates on every deploy without the content changing trains systems to ignore your dates.
4. Validate after every template change. Schema breaks silently. Nothing on the page looks different when a required property disappears.

The layer schema does not reach

Schema.org describes a page for something that reads pages. It has no vocabulary for the thing an autonomous agent needs, which is a machine-readable statement of what your product can do and how to call it.

That lives in a different artefact entirely: a capability manifest at a well-known path (what that involves), currently specified by Agentic Resource Discovery and served at /.well-known/ard.json. Perfect schema markup contributes nothing to it, and it contributes nothing to your schema.

It is also close to uncontested. A crawl of 375,998 domains in September 2026 found 207 serving such a manifest, with 85 percent of those still on a superseded path. If you have already done the structured data work, this is the natural next hour, and there is almost nobody there.

A sensible order

1. Confirm your JSON-LD is in the server-sent HTML, not injected by script.
2. Add or fix Organization with honest sameAs links.
3. Add Article or WebPage with real dates to every content page.
4. Add FAQPage where genuine questions and self-contained answers exist on the page.
5. Validate, then set a reminder to validate again after the next template change.
6. Publish a capability manifest, which is the layer schema cannot reach.