Structured Data and AI Search: What JSON-LD Actually Does for You in 2026

Structured data doesn't change how your page looks to a visitor. It changes how much a machine - a search engine or, increasingly, an AI assistant - actually understands about what's on the page instead of guessing from raw text. Here's what it is, what it's worth doing in 2026, and how to check whether yours is any good.

What JSON-LD Actually Is

JSON-LD is a block of JSON dropped into your page inside a <script type="application/ld+json">tag. It's invisible to a visitor - it doesn't render anything - and it describes the page using a shared vocabulary called schema.org: this is an Article, written by this Person, published on this date, part of this Organization's WebSite. It's a separate, parallel description of the same content that's already visible in your HTML, written in a format a machine can parse without inferring anything.

It isn't a ranking factor by itself. Google has said this directly for years. What it actually does is make your page eligiblefor things a plain HTML page isn't eligible for - rich results in search, and increasingly, reliable interpretation by systems that read your page programmatically rather than rendering it for a human.

Two Audiences, One Markup

The classic reason to add JSON-LD is rich results: FAQ accordions, breadcrumb trails, article cards with a byline and date, star ratings on a product listing. These are the visible payoff in a Google search result page, and they exist because your markup told Google's crawler exactly which text is the question, which is the answer, and which is the price.

The newer reason is AI crawlers and assistants. When an LLM-based crawler or agent reads a page, it's working from the same raw HTML a search engine sees - but parsing prose for entities and relationships is lossy and error-prone. Structured data hands it the answer directly: this organization's name is X, this product's price is Y, this article's author is Z. It's the difference between an agent inferring your business hours from a paragraph and reading them from a OpeningHoursSpecification property. Same information, much less room for the machine to get it wrong. Our SEO audit's Agent Readiness scoringbreaks this exact signal down in more detail if you want to see how it's weighted alongside other AI-discoverability factors.

The Schema Types Worth Using

Most sites only need a handful of types, and they cover the large majority of real use cases:

• Organization - name, logo, social profiles, contact info. Belongs site-wide, usually in a shared layout.
• WebSite - enables sitelinks search box eligibility and ties your pages back to one site entity.
• Article - headline, author, publish/update dates, image. What powers article rich cards.
• FAQPage - question/answer pairs that can render as an expandable accordion directly in search results.
• BreadcrumbList - your page hierarchy, shown as a breadcrumb trail instead of a raw URL.
• Product - price, availability, ratings, for anything you sell.

One worth flagging because it still gets recommended in outdated guides: HowToschema. Google retired HowTo rich results in search in 2023. Adding it today doesn't hurt anything, but it also won't earn you the step-by-step rich result it used to - don't prioritize it.

A Practical Audit Checklist

Three questions, in order, cover almost every real problem we see in structured data:

• Is it present at all? A surprising number of sites ship zero JSON-LD anywhere, including the homepage.
• Does it validate? Well-formed JSON, required properties present for the type you're declaring, no typos in property names that silently get ignored.
• Does it match the visible page?This is the one people skip, and it's the one that matters most. Schema that claims a 4.8-star rating when no rating is shown on the page, or an FAQ block with answers that don't appear in the visible content, isn't just wasted effort - Google explicitly treats structured data that misrepresents page content as grounds for a manual action, not just a missed rich-result opportunity.

llms.txt Is a Different Thing

It's easy to lump structured data and llms.txt together because they both target machine readers, but they're not the same tool. JSON-LD is per-page, machine-parseable markup describing that specific page's content in a formal vocabulary. llms.txt is a plain-text file at /llms.txton your domain - a human-readable summary of your site's structure and purpose, intended to give an AI agent quick orientation before it starts crawling, the same way robots.txt gives crawlers rules before they start requesting pages. One describes a page in detail; the other describes the site at a glance. Worth having both, but they solve different problems and neither substitutes for the other.

What We Actually Check

To be direct about what our own tooling does here: Scanverra's Agent Readiness checks look for whether JSON-LD is present on the page and whether llms.txt exists at your domain root, as part of a broader score covering AI-crawler access in robots.txt, semantic landmarks, and layout stability. That's a presence check, not a schema validator - it tells you whether search engines and AI assistants have anything structured to read at all, not whether every property on every type is correctly typed. For the visible-content-match problem above, that part is still on you to check by hand, since it requires understanding what your page actually says.