Schema Markup Won't Save You: What Structured Data Does and Doesn't Do for LLMs

A specific claim keeps circulating in AI-visibility threads: add FAQ schema to your pages and you will start showing up in ChatGPT. It is stated with confidence, sold in audits, and baked into checklists. The claim is wrong in the direction that matters, and it is wrong for a reason worth understanding, because the same misunderstanding leads people to skip the things that actually change how AI systems find and quote them. Structured data is genuinely useful. The useful part is almost never the part being advertised.

Start with what structured data is. Schema.org is a shared vocabulary for describing things on a page in a way machines can read without guessing. JSON-LD is the format most people use to express it: a block of JSON, usually dropped in the page head, that says in plain key-value terms what the page is about. A recipe page can declare its cook time and ingredient list. A product page can declare its price, availability, and review count. An article can declare its author, publish date, and publisher. The markup does not change what a human sees. It hands a parser a clean, labeled version of facts that would otherwise have to be inferred from layout and prose.

{
  "@context": "https://schema.org",
  "@type": "Organization",
  "name": "Acme Analytics",
  "url": "https://acme.example",
  "sameAs": [
    "https://www.wikidata.org/wiki/Q000000",
    "https://www.linkedin.com/company/acme-analytics"
  ]
}

That block does one valuable thing. It tells any system that reads it exactly which entity this page is about and connects that entity to canonical identifiers elsewhere. Hold onto that, because it is where the real value lives.

Where structured data genuinely helps

The honest case for schema rests on three mechanisms, and none of them is a direct pipe into an LLM's answer.

The first is parsing. A crawler that reads your JSON-LD does not have to reverse-engineer your page structure to learn that the number 49 is a price rather than a model number or a page count. Clean structured data reduces ambiguity for any automated consumer, and reduced ambiguity means fewer wrong inferences about what your page asserts. This is unglamorous and real.

The second is entity disambiguation and knowledge graphs. When your Organization markup includes sameAs links to Wikidata, an official social profile, or an industry registry, you are telling indexing systems that the "Acme" on this page is the same Acme they already know from those other sources. Search engines build knowledge graphs out of exactly these connections, and those graphs are part of how any downstream system decides whether two mentions refer to one entity or two. If an AI system leans on a search backend that maintains such a graph, and most of them do, clean entity signals help you be understood as a distinct, known thing rather than an unresolved string of characters.

The third is eligibility in the classic search backends. This is the mechanism most AI-visibility advice fumbles. ChatGPT's browsing mode uses a provider-managed web retrieval system whose implementation can change over time. Some rich-result and knowledge-panel features in traditional search are gated behind valid structured data. Where an AI system draws candidate documents from a search backend that already uses schema to determine eligibility or to render enhanced results, your markup can affect whether you are in the pool that the backend considers well-formed and indexable. Schema does not talk to the model. It can influence the index that feeds the model.

Those three reasons are sufficient to do the work. Do it for them.

Why "FAQ schema gets you cited" breaks down

Now the part the pitch gets wrong. To see why, you have to look at what happens between a user's question and the answer that names some sources and not others.

A retrieval-augmented answer is assembled, not recalled. The system rewrites the user prompt into several sub-queries, a step called fan-out. For each sub-query it retrieves a candidate set of documents from a search backend. A cross-encoder reranker trims that pool down to a handful of passages that fit the context window. Those passages get assembled into the prompt, the model reads them, writes an answer, and attaches citations to some of what it used. Retrieval and attribution are separate stages, and the model's reading happens on text, not on markup.

That last point is the one that dissolves the FAQ-schema claim. By the time a passage reaches the model, it is rendered text pulled from the page body. The reranker scores passages on semantic relevance to a sub-query, comparing the meaning of your prose against the meaning of the query. Your JSON-LD block is not the passage that competes at that stage, and the answer the model composes is written from the words a reader would see. You can mark up an FAQ perfectly and still lose every reranking contest, because the contest is about whether a chunk of your visible text answers the sub-query better than a chunk of someone else's.

There is a narrower, true version of the FAQ story worth separating out. Marking up a question-and-answer block does not hurt, and it can help a search backend recognize the passage as a self-contained answer to a specific question, which is a parsing and eligibility benefit, the first and third mechanisms above. But that is a long way from "you will rank in ChatGPT." The slogan treats a rendering hint as if it were a ranking guarantee inside a system the markup never enters.

Necessary in places, sufficient nowhere

The accurate framing is that structured data is necessary-but-not-sufficient for a specific slice of the funnel, and irrelevant to the rest.

The AI Visibility Funnel has four stages: Retrievability, Candidacy, Citation, Attribution. Retrievability is whether a crawler can reach and parse your page at all. Candidacy is whether your page makes the candidate set for a given sub-query. Citation is whether the model, having read the candidates, quotes or references you. Attribution is whether the visible answer names you as the source.

Schema does its work at the edges of Retrievability and, indirectly, at Candidacy, by keeping your entity legible to the backend that assembles candidate sets. It contributes almost nothing at Citation and Attribution, because those stages run on the rendered passages the reranker selected and the model read. Treating schema as a Citation lever is optimizing a stage it does not touch. Skipping it entirely can leave you illegible to the index at the stages where it does matter. Both mistakes are common, and they point in opposite directions.

Here is the test that keeps you honest. Ask of any schema task: does this help a system correctly identify what my page and my entity are? If yes, it is worth doing, at the parsing and entity level. Ask a second question: does this change which of my rendered passages wins a semantic reranking contest against third-party pages? If the honest answer is no, then no amount of markup substitutes for the passage work that does change it.

The crawlers feeding AI still run on classic signals

The AI crawlers are real and named. OpenAI operates GPTBot, OAI-SearchBot, and ChatGPT-User. Anthropic operates ClaudeBot and Claude-User. Perplexity operates PerplexityBot and Perplexity-User. Google-Extended and Common Crawl's CCBot are training-related controls. Check them separately from search-specific agents such as OAI-SearchBot and PerplexityBot.

What those crawlers and the backends behind them reward looks a lot like durable, well-understood web presence rather than a markup trick. Cited sources include both durable references and newly refreshed pages. Source age, branded demand, link authority, and topic stability are confounded in observational datasets, so none should be treated as a causal ranking factor. The pattern they describe is that being a known, linked, retrievable entity is the substrate, and schema is one small input to being known rather than the mechanism itself.

The Princeton and IIT Delhi GEO study points the same way about content. Aggarwal and colleagues built GEO-Bench, roughly 10,000 queries across about 25 domains, and tested content interventions against a visibility metric. The methods that raised visibility were adding citations, adding statistics, adding direct quotations, and improving fluency, with the best moving their metric by up to around 40 percent and lower-ranked sources gaining the most. Keyword stuffing did nothing or slightly hurt. Notice what is absent from that list of winners. The interventions that worked all changed the readable substance of the passage. None of them was a markup annotation, because the benchmark measured how a model apportions visibility among text it actually reads.

What to prioritize over it

If schema is a small, worthwhile input and not the lever, here is where the leverage actually sits, roughly in order.

Get retrievable first. Confirm the AI user-agents can reach you, that your important pages render their key content without requiring client-side execution the crawler may skip, and that nothing critical is buried behind interactions a parser never triggers. A page the crawler cannot read cleanly cannot be helped by any tag inside it.

Then build passages that win reranking. Write self-contained answers that resolve a specific question without needing the rest of the page for context. No universal ideal chunk length has been established. A chunk that stands on its own survives the trip into a context window and reads as a complete answer to a sub-query. This is the single most direct thing you control at the Citation stage, and it is prose work, not markup work.

Then work your Citation Surface, the full set of third-party pages where an AI can find and cite you. Practitioner datasets often show third-party pages contributing heavily to brand mentions, but the observed share is not universal. If that holds even loosely for your market, most of your retrieval opportunities live on comparison roundups, community answers, documentation, and analyst notes you do not own. Being mentioned accurately, in a self-contained passage, on pages that already retrieve well for your prompts puts you into candidate sets your own site was never going to reach.

Then map Prompt-Space Coverage, the real space of prompts customers ask and your visibility across it, instead of a flat keyword list. Fan-out means one question becomes several sub-queries, so a page tuned for a single head phrase can miss the reformulations that actually get retrieved. Measuring across the prompt space tells you where your Retrieval Gap is widest, meaning the distance between the pages that could answer a query and the smaller set the pipeline actually pulls.

And do the schema. Add clean Organization and Article markup, wire up sameAs to your canonical identifiers, keep your entity legible. It belongs on the list. It belongs near the bottom, doing the quiet parsing and entity work it is good at, not carrying a promise it was never built to keep.

A note on measurement

The reason schema keeps getting oversold is that it is visible and easy to verify. You can open a validator, see green checkmarks, and feel finished, in a way you cannot with the messier question of whether your passages are actually being retrieved. The check that matters is upstream of citations: Share of Retrieval, the fraction of relevant candidate sets your content appears in before any citation decision is made. That is the leading indicator, and it is the one no markup validator reports. Tracking it across the four AI engines is what shareof.ai↗ automates, and you can approximate it by hand by running your real prompt space through the assistants and recording where you surface and where you do not. Either way, measure retrieval, not markup validity, because retrieval is the thing the answer is built from.

Structured data is worth doing for entity clarity and backend eligibility. Just do it for those reasons, and spend the hours you would have spent chasing FAQ-schema citations on the passages and third-party pages that decide whether you are in the room at all.

Primary sources and evidence

• GEO: Generative Engine Optimization (Aggarwal et al.)↗ — the original paper and benchmark.
• Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks↗ — the foundational RAG paper.
• OpenAI web crawlers↗ — GPTBot, OAI-SearchBot, and ChatGPT-User controls.
• Perplexity crawler documentation↗ — PerplexityBot and Perplexity-User guidance.
• Google common crawlers↗ — Googlebot and Google-Extended distinctions.
• Google: AI features and your website↗ — official guidance for AI Overviews and AI Mode.
• Schema.org getting started↗ — the structured-data vocabulary and its intended use.

Continue through the system

• 'AI SEO' Is Optimizing the Wrong Layer
• The AI Visibility Funnel: A Diagnostic for Where You're Losing Citations
• Reading Your AI Crawler Logs: GPTBot, ClaudeBot, PerplexityBot and Friends

Run the practical layer with the free AI visibility scan, explore the AI search library, or see the AI visibility benchmarks.