Schema Markup for AI Citability: What ChatGPT and Perplexity Actually Parse

Ask ten GEO consultants whether schema markup helps you get cited by ChatGPT and you'll get ten confident answers — and about half of them are guessing. The honest position is more useful than either the hype or the dismissal: some of this is genuinely confirmed by the platforms themselves, and some of it is still speculation dressed up as certainty. This post on schema markup for AI citability separates the two.

What Google has actually confirmed

Start with the one platform that's explicit about it. Google's own AI features documentation states that AI Overviews and AI Mode draw from the same underlying index as organic Search — there's no separate schema or special markup exclusively for AI features, according to Google Search Central's structured data guidelines.

That doesn't mean schema is irrelevant to AI Overviews. It means the relationship is indirect: structured data helps a page earn rich-result eligibility and clearer entity understanding within the classic index, and AI Overviews pull from that same pool. If your page is well-marked-up enough to be rich-result eligible, it's more likely to be a clean, well-understood candidate when Google's systems synthesize an AI Overview answer — not because there's an "AI Overview schema type," but because the whole pipeline still runs through the same crawling, indexing, and understanding system it always did.

What's genuinely murkier: ChatGPT and Perplexity

This is where the confident claims outrun the evidence. There isn't a public, authoritative statement from OpenAI or Perplexity saying "yes, we parse JSON-LD as a distinct signal, weighted higher than prose." What exists instead is a mix of vendor testing and inconsistent results.

One notable test found that when researchers fed ChatGPT and Perplexity fake, invalid schema, both systems still extracted information from it — behaving as though they were reading the schema block as regular page text rather than validating it as structured data, per Search Engine Roundtable's reporting on the test. That's a meaningfully different claim than "AI systems parse and trust JSON-LD as machine-readable data" — it suggests, at least in that test, the models weren't distinguishing valid structured data from any other text on the page.

Other testing (much of it from GEO vendors with a product to sell, worth reading skeptically) claims the opposite: that FAQPage markup gets treated as a distinct, higher-confidence source for direct answer extraction, and that JSON-LD properties get tagged with more trust than the same fact stated in prose.

The honest summary: it's plausible, not proven. Schema almost certainly doesn't hurt, and it likely helps indirectly by making a page's facts unambiguous wherever an AI system's crawler does read it as text — a clearly labeled review count or price is easier to extract correctly than the same number buried in a sentence, whether or not the JSON-LD tag itself is "understood." But treat any claim of an exact citation-rate lift from schema alone with real skepticism until a platform confirms the mechanism directly.

Schema markup for AI citability: the types actually worth prioritizing

Given that uncertainty, prioritize the types that help regardless of whether AI systems parse JSON-LD specially — because they're either confirmed useful for classic Search (which still feeds AI Overviews) or they make your page's facts unambiguous to any system reading it as text.

Organization — establishes who you are as an entity: name, logo, sameAs links to your official profiles. This is foundational for entity disambiguation, which matters whether the reader is Google's Knowledge Graph or an LLM trying to figure out if "Acme" the software company is the same "Acme" as the anvil manufacturer.

Article — headline, author, datePublished, and dateModified. The datePublished/dateModified pair matters more than people think: freshness is one of the more consistently cited factors in AI citation research, and explicit, machine-readable dates remove any ambiguity about how current a page actually is.

FAQPage — structures question/answer pairs as discrete units. Whether or not the JSON-LD itself gets parsed as structured data, the underlying page content this schema type encourages — a clear question followed immediately by a direct, self-contained answer — is exactly the passage shape that gets extracted and cited, independent of the markup.

HowTo — step-by-step structure with numbered, discrete steps. Same logic as FAQPage: the content shape this encourages (clear, ordered, standalone steps) is citable on its own merits.

sameAs — links your entity to Wikipedia, Wikidata, LinkedIn, Crunchbase, and other authoritative profiles. This is one of the more concretely useful properties for entity resolution — it's how a system (AI or classic search) confirms which real-world entity your page is actually about.

| Schema type | Confirmed benefit | Speculative AI-specific benefit | |---|---|---| | Organization + sameAs | Entity clarity in Google's Knowledge Graph | Possible entity disambiguation for LLM retrieval | | Article (with dates) | Rich results eligibility, freshness signal | Possible freshness weighting in AI citation selection | | FAQPage | Rich results (where still shown) | Encourages citable Q&A passage structure | | HowTo | Rich results (where still shown) | Encourages citable step-by-step passage structure |

What's speculative and should be treated that way

Be skeptical of any claim that a specific, uncommon schema type (Speakable, certain Review sub-properties, custom extensions) produces a measurable AI citation lift, unless the claim comes with a methodology you can actually inspect — sample size, how "citation" was defined, whether it controlled for the page's existing ranking or authority. Most of the confident-sounding stats circulating in the GEO content right now don't survive that scrutiny.

The schema.org vocabulary itself — the source of truth for every type and property mentioned here — is maintained independently of any search engine at schema.org, which is worth bookmarking directly rather than relying on secondhand summaries of what a property does.

Google has also been explicit, in its 2026 guidance covering AI search, that AEO and GEO are best understood as extensions of existing SEO fundamentals rather than a separate discipline requiring new markup, according to Search Engine Journal's coverage of Google's AI search guide. That's consistent with the "same index" point above — there isn't a parallel AI ranking system with its own schema requirements, at least not one Google has confirmed.

How to validate your schema

Whichever types you implement, validation matters more than coverage. A page with broken JSON-LD — a missing required property, invalid nesting, a stray comma — isn't eligible for rich results at all, and there's no reason to believe a broken schema block helps an LLM either. Run your pages through our free schema checker to see exactly which schema types are present on a URL and inspect the raw JSON-LD, rather than guessing from the rendered page. If you're also checking whether AI crawlers can reach the page at all, pair it with our free robots.txt tester — validated schema on a blocked page never gets read by anyone.

Common mistakes

• Adding schema without validating it. Invalid JSON-LD is worse than none — it signals sloppiness to any system parsing it, human or machine.
• Assuming more schema is always better. Marking up content that isn't actually true (fake review counts, inflated ratings) is a policy violation for Google and, if AI systems do weight structured data, a fast way to get flagged as unreliable.
• Skipping dates. datePublished and dateModified are among the cheapest, most confirmed-useful properties to add, and they're the ones sites skip most often.
• Treating schema as a substitute for content quality. No schema type fixes a page that doesn't actually answer the question clearly in prose. Structure amplifies good content; it doesn't replace it.

FAQ

Does schema markup help AI citability? Confirmed for Google: AI Overviews draw from the same index as organic Search, so schema that helps rich-result eligibility indirectly helps AI Overview eligibility too. For ChatGPT and Perplexity, it's plausible but not confirmed by either company — treat specific citation-lift claims skeptically.

Do ChatGPT and Perplexity actually read JSON-LD structured data? Unclear. At least one test found both treated invalid schema as regular page text rather than validating it as structured data, which suggests they may not be parsing it as a distinct signal in all cases. No official statement from either company confirms or denies special JSON-LD handling.

Which schema type should I add first? Organization with sameAs links, for entity clarity, followed by Article with accurate dates. Both are confirmed-useful for classic Search regardless of any AI-specific effect.

Is FAQPage schema still worth using if Google removed some FAQ rich results? Yes, for a different reason than rich results: the Q&A structure it encourages — a direct question followed by a self-contained answer — is exactly the passage shape that gets extracted and cited, independent of whether the markup itself earns a visual rich result.

Run your own pages through the free schema checker to see what's actually implemented before assuming a citation problem is a schema problem — often it's the underlying content structure, not the markup, that needs the fix.