Schema Markup for LLMs: 7 Elements That Matter, 12 That Don't

There is a generation of SEO advice, much of it written between 2019 and 2023, that treats schema.org like a universal good. More is better. Mark up everything. Add Review on every product page even if there are no reviews. Add FAQPage wherever you can shoehorn it. Add Organization with every optional field filled in.

That advice was calibrated for Google's rich snippets era, where structured data primarily fought for SERP real estate. It is not calibrated for how language models actually use structured data. When a crawler working on behalf of OpenAI, Anthropic, or Google's Gemini pipeline ingests your site, it cares about a much narrower set of properties — and it is actively skeptical of markup that does not match visible page content.

This post is the triage. We will look at what actually moves the AI Discoverability dimension on BrandGEO's 150-point rubric, and by extension what contributes to Knowledge Depth when LLMs describe your brand. Seven schema elements worth real investment. Twelve you can stop agonizing over. One that is more important than the rest combined.

The One That Matters More Than the Rest

Before the lists, the single highest-leverage piece of schema you can implement is a complete, well-formed Organization object at the root of your site, with sameAs links to every authoritative external profile of your brand.

That one object — done correctly — does more for AI visibility than thirty other markup implementations combined.

Why: Organization with full sameAs is how crawlers disambiguate your brand in the knowledge graph. "Acme" on your site gets linked to "Acme" on LinkedIn, "Acme" on Wikipedia, "Acme" on Crunchbase, "Acme" on GitHub. Without those linkages, the model treats the string "Acme" as ambiguous across many entities. With them, the model has a single canonical identity to attach facts to.

A minimal but effective Organization block:

{
  "@context": "https://schema.org",
  "@type": "Organization",
  "name": "Acme",
  "legalName": "Acme Holdings Inc.",
  "url": "https://acme.com",
  "logo": "https://acme.com/logo.png",
  "description": "A brief, factual description of what you do.",
  "foundingDate": "2017",
  "founder": [{ "@type": "Person", "name": "Jane Doe" }],
  "address": {
    "@type": "PostalAddress",
    "addressLocality": "Singapore",
    "addressCountry": "SG"
  },
  "sameAs": [
    "https://www.linkedin.com/company/acme",
    "https://en.wikipedia.org/wiki/Acme",
    "https://twitter.com/acme",
    "https://www.crunchbase.com/organization/acme",
    "https://github.com/acme"
  ]
}

If you implement nothing else from this article, implement this — and audit it quarterly to make sure every sameAs link resolves to a live profile you actually control or appear on.

The Seven That Matter for AI Visibility

1. Organization with sameAs

Covered above. Treat it as non-negotiable.

2. Person schema on leadership pages

LLMs want to answer "who founded X?" and "who is the CEO of X?" correctly. A Person block on each leadership bio page, linked back to the Organization via worksFor, is one of the cleanest signals you can provide. Include jobTitle, sameAs (LinkedIn, personal site, published author profiles), and a short factual description. This directly feeds Recognition and Knowledge Depth.

3. Product or SoftwareApplication with structured properties

For each product or core offering, a Product (or for software, SoftwareApplication) block with:

• name
• description (factual, not marketing)
• brand (linked to your Organization)
• category — critically important for Contextual Recall
• applicationCategory (for software)
• offers with price and priceCurrency if stable

The category string is what lets the model answer "what are the best X tools?" and include you in the answer. A fuzzy description of your product without category metadata means you get omitted from category-level queries even when the model knows you exist.

4. BreadcrumbList

Simple, cheap, underrated. Clear breadcrumbs tell crawlers the structural hierarchy of your site. This does not move Recognition, but it helps the model cluster your pages correctly, which improves how accurately it describes your product structure.

5. Article with author and datePublished

For every editorial piece, blog post, research note, or case study. Three properties do real work:

• author linked to a Person who has credentials elsewhere
• datePublished and dateModified (models penalize content that looks stale)
• about — linking the article to its topic entity

Sentiment & Authority is partly built from how LLMs judge the editorial quality of content associated with your brand. Author-attributed, dated articles read as higher-quality than anonymous undated posts.

6. FAQPage — but only on pages that genuinely contain Q&A

The honest version of FAQ schema. If you have a real, user-facing FAQ page, mark it up. The questions you ask there — if they match the phrasing users actually type into LLMs — directly improve Contextual Recall. If you do not have a real FAQ, do not fake one to chase schema. Google started demoting faked FAQ in 2023, and LLMs ignore it.

7. Review and AggregateRating — if and only if they are real

Structured reviews are a legitimate authority signal. But they must be real and the markup must match what a visitor sees on the page. Inflating AggregateRating values, marking up testimonials with no visible rating, or repeating the same review across pages will get your whole domain's structured data distrusted by crawlers. The downside is worse than the upside of fair markup.

The Twelve That Do Not Matter (in 2026, for LLMs)

These are properties that retain some Google rich-snippet value in narrow cases but do not move AI visibility. Unless you have a specific SERP reason, deprioritize them.

1. Event schema on every webinar. Rarely ingested into LLM training data; ephemeral.
2. VideoObject beyond the most basic fields. YouTube's own structured metadata is what models use. Duplicating it on your page does not compound.
3. HowTo with nested HowToStep. Google deprecated most rich-snippet support in 2023, and LLMs prefer prose walkthroughs over nested step markup.
4. LocalBusiness for pure-software brands. Unless you have a real physical location that matters commercially, Organization is sufficient.
5. Recipe — unless you are a food brand.
6. JobPosting — only matters if you actively want the hiring pages ingested for recruiting; orthogonal to brand description.
7. ImageObject with full creator and license blocks — low leverage for brand visibility.
8. Speakable — originally intended for voice assistants, largely abandoned.
9. SiteNavigationElement — noise.
10. WebPage with redundant metadata that duplicates your HTML <title> and <meta> tags. Pick one source of truth.
11. CollectionPage everywhere. Overused. Rarely parsed beyond Organization-level signals.
12. ProfilePage — redundant with Person.

Implementing these does not hurt you unless the markup contradicts your visible content. But the hours spent on them are hours not spent on the seven above.

The Principle Behind the Triage

What separates the seven from the twelve is one question: does this property teach the model something about your brand's identity, offering, or authority that it cannot easily infer from the prose?

Organization teaches canonical identity. Person teaches who the humans are. Product with category teaches what you sell and where you belong. Article with author teaches who wrote what.

Event, VideoObject, HowTo — these teach the model about one-off content artifacts, not about your brand. The content artifact will be ingested regardless through normal crawling. The schema does not add marginal signal.

This is why a well-crafted Organization block outperforms a sprawling schema implementation across fifty pages. The narrow, identity-defining markup compounds. The broad, artifact-describing markup mostly does not.

Implementation Checklist

If you want a thirty-minute triage of your current markup, here is the operational checklist:

1. Fetch your homepage and one deep page and extract all JSON-LD blocks.
2. Does your homepage have a single canonical Organization block? If yes, check that sameAs includes LinkedIn, Wikipedia (if you have one), Crunchbase, and your core social profile. If not, build one.
3. Does every leadership bio have a Person block linked to the Organization? If not, add them.
4. Does every product or service page have a Product or SoftwareApplication block with a category? If not, add them. This is frequently the biggest gap.
5. Do you have Article markup with author on your content? If your CMS is generating anonymous, undated article markup, fix it at the template level.
6. Are any of your existing markups lying? AggregateRating without real reviews, FAQPage faked, Review with inflated scores. Remove anything that does not match visible content.
7. Validate with the schema.org validator and Google's Rich Results Test. Then publish.

You can complete this in a single focused afternoon on a small site. For a large marketing site, scope it to the templates — homepage template, product template, bio template, article template — because every page is generated from templates anyway.

Diagnosing the Effect

Schema changes show up on the AI Discoverability tile first. It is the dimension most directly tied to how AI crawlers perceive your site. Expect movement in the range of 3–10 points on the 25-point sub-score within six to twelve weeks of implementation as crawlers re-ingest.

Knowledge Depth follows with a longer lag — typically one model training cycle, which for base models means three to nine months. Search-augmented providers (ChatGPT with browsing, Gemini 3 Pro, Grok 4) react faster because they re-fetch pages on demand; if a user asks about you tomorrow, they will retrieve your fresh markup and weight it.

The way to see this is through a Monitor. Run weekly or daily scans, tag the week you shipped the schema changes, and look at the trajectory of AI Discoverability and Knowledge Depth from that anchor point. Without a Monitor, the signal is too slow and too noisy to attribute.

The Anti-Checklist Takeaway

Schema markup is a case where an eighty-percent job on the seven elements above beats a hundred-percent job on the full schema.org tree. The discipline is saying no to the markup that does not move anything.

If you are writing a GEO audit for a client or your own brand and you are tempted to include "add schema.org markup site-wide" as a recommendation, refine it. Specify Organization with complete sameAs. Specify Product with category. Specify Person on bio pages. Those four specifications do more than the generic one.

Common Implementation Questions

A few questions that come up repeatedly in schema implementations for AI visibility.

"Should we use JSON-LD or Microdata?"

JSON-LD, essentially always. It is Google's preferred format, easier to maintain, and cleaner for crawlers and LLM ingestion pipelines to parse. Microdata and RDFa remain valid but offer no advantage in 2026 and have higher maintenance cost because they live intermixed with the HTML markup.

"Does schema help if my content itself is weak?"

Not much. Schema makes clear content clearer. It does not make thin content authoritative. The entity-first content approach described in the entity-first content playbook is the prerequisite. Schema is the structured expression of the entities your prose already names.

"What happens if our schema claims contradict the visible page?"

The page gets distrusted. Google has been explicit about this for rich snippets since 2019. LLM training pipelines and retrieval layers apply similar heuristics — if your AggregateRating says 4.8 and the visible page shows three reviews averaging 3.2, the mismatch is detected. Downstream, everything else you mark up on that domain is weighted lower.

"Should we worry about schema updates when we ship product changes?"

Yes. A Product block with outdated pricing, discontinued features, or an old category string is actively harmful because it entrenches stale facts in crawler memory. Add schema updates to your product release checklist. If a marketing page changes, the structured data on it should change too.

The Compounding Effect

One reason schema works better in practice than on paper is the compounding effect of consistency. A single page with perfect Organization markup has some value. A whole domain with consistent, cross-referenced markup — Organization links to Person bios, which link to Article authors, which link to Product reviews — has much more. The crawler's confidence in your entity graph rises as the internal consistency checks succeed.

This is why the template-level approach matters. Fixing schema on one high-value page is less leveraged than fixing it at the CMS template level for every page of that type. The investment is the same, the output is orders of magnitude larger.

Want to know where your AI Discoverability sits across five providers today? A BrandGEO audit covers that dimension with concrete per-provider recommendations.