Schema Won’t Buy LLM Citations — Use It Anyway

TL;DR: LLMs are not parsing your JSON-LD as schema — they’re reading it as slightly oddly punctuated text. A controlled experiment with deliberately broken, nonsense schema proved LLMs returned the fake address anyway, citing “structured data” they never actually validated. Schema still has value, but selling it as a magic citation lever is not supported by the current evidence.

What Schema Was Actually Built to Do

Schema.org structured data is a shared vocabulary — built cooperatively by Google, Microsoft, Yahoo, and Yandex — designed to let webmasters embed machine-readable information directly into pages. The core job is disambiguation. Natural language is messy. “Apple” is a fruit, a technology company, a record label, and the surname of someone’s gerbil. Plain text forces a search engine to guess context. Schema eliminates that guesswork by tagging strings with explicit types: this string is a streetAddress, this entity is an Organization, this identifier resolves to a specific node in a knowledge graph.

That explicit structure feeds Google’s Knowledge Graph — the entity-and-relationships database behind knowledge panels, carousels, and People Also Ask results. Schema is one of several inputs, but it’s a low-noise one, which is precisely why search engines have pushed adoption for over a decade.

The problem is that most of the current “GEO best practice” conversation conflates what schema does for traditional search engines with what it supposedly does for large language models. Those are different systems with different architectures, and the distinction matters significantly if you’re spending client budget on it.

The Duck Experiment: What It Actually Proved

SEO researcher Mark Williams-Cook built a test page for a fictional company called DUCK YEA T-SHIRTS. The visible page content mentioned no address. Buried in the HTML head, inside a <script type="application/ld+json"> block, sat a completely fabricated schema: a made-up @context URL that resolves to nothing, a made-up @type called MallardEnterprise, and invented properties including flockName, waddleStyle, nestingGrounds, and quackVolume. Not a single property existed in the Schema.org vocabulary.

He then asked ChatGPT and Perplexity where the company was based. Both returned the fabricated address. Perplexity explicitly credited the “embedded structured data.” Neither flagged invalid types. Neither rejected the nonsense vocabulary.

A genuine schema-aware parser would have rejected this outright. Instead, both systems did what LLMs always do: read the text of the page, found a string that looked like an address, and returned it. The JSON-LD wrapper was just text with curly braces. The disambiguation — the entire point of schema — was invisible to them.

This isn’t fringe skepticism. It’s a reproducible test. If the LLM were genuinely parsing schema as designed, invalid @types and fabricated properties would produce a failure state, not a confident answer. They produced a confident answer.

Why Schema Probably Doesn’t Survive LLM Training

There are two mechanisms by which schema could theoretically influence an LLM: getting baked in during pre-training, or being read at query time via live page fetching. Both are weaker than the LinkedIn carousel crowd suggests.

On training: large-scale pre-training pipelines strip HTML and boilerplate before a single GPU sees the data. The widely used FineWeb dataset — 15 trillion tokens derived from 96 Common Crawl snapshots — uses trafilatura specifically because it extracts main page text and discards markup. JSON-LD lives in a <script> tag. Trafilatura ignores <script> tags by design. The schema on your product page is almost certainly being discarded alongside your analytics snippet and cookie banner.

Even if some JSON-LD did survive ingestion, it would be tokenized — broken into sub-word chunks — which dissolves the structural relationships entirely. The string "@type": "Organization" becomes a sequence of tokens no different from the same words appearing in a blog post about schema. The disambiguation collapses at the first step of training. And for any individual business, a single mention of a streetAddress on a single product page is a negligible signal in a fifteen-trillion-token corpus. The model won’t recall it, because it was never reinforced enough to settle into the weights.

On query-time fetching: this is more plausible in principle. RAG-based systems that retrieve pages at answer time could, in theory, parse structured data from those pages. But there is no published evidence, no leaked papers, and no public confirmation from any frontier lab that this is happening. The argument “they probably should do this, therefore they are doing this” is conjecture dressed as strategy.

What This Means for High-CAC Vertical Operators

For operators running acquisition campaigns in forex, legal, iGaming, or crypto — verticals where a single converted lead can be worth hundreds or thousands of dollars — the schema conversation has a direct budget implication. If an agency is billing for “AI citation optimization via schema implementation” at meaningful cost, you need to see the working.

The Ahrefs study tracking 1,885 pages that newly added JSON-LD found essentially no effect on AI citations across ChatGPT, AI Mode, and AI Overviews. The study has methodological caveats — it tested already heavily cited pages, where schema’s disambiguation value is lowest — but it is currently the most rigorous public data available. “LLM returned content that appeared in the schema” is not evidence the schema was used. It’s evidence the LLM read the page.

Where schema investment does make sense for high-CAC operators: new brands and challenger entities with thin web footprints. A new forex broker, a recently launched crypto exchange, or a law firm without a knowledge panel — these are the entities where schema’s disambiguation value is real. You’re not buying a citation today. You’re establishing a resolvable node in the entity graph so you’re a candidate when the system eventually sorts out how to wire structured data into LLM answer pipelines. That’s a longer-horizon bet, but it’s a legitimate one.

Operators running iGaming acquisition campaigns in particular operate under brand-collision risk — generic terms, similar operator names, geographic variants. Schema on entity pages won’t buy you an AI Overview citation this quarter, but it is part of the infrastructure work that supports long-term entity resolution. Same logic applies to forex broker lead generation, where brand credibility signals compound over time. For legal operators, a law firm’s digital presence depends heavily on being an unambiguous, resolvable entity — not just ranking for keywords.

If you’re unsure where schema fits in your current setup, a proper technical marketing audit will surface the gaps worth fixing versus the line items worth cutting.

Google’s Own Systems Aren’t Talking to Each Other

There’s an uncomfortable data point buried in the original research. A single Google SERP showed the AI Overview confidently stating a business was open, listing address and hours — while the Google Business Profile knowledge panel on the same page displayed a large red “Permanently closed” banner for the exact same business.

Google Business Profile data is structured, user-edited, and as close to authoritative as any business-hours signal gets. Google owns the model, the index, the knowledge graph, the crawler, and the AI Overview surface. If full vertical integration over every part of the stack can’t reliably wire structured business data into AI answers, the claim that OpenAI or Anthropic has built a richer entity pipeline that defers to your Organization schema deserves serious skepticism.

This doesn’t mean structured data is worthless. It means the infrastructure connecting structured data to LLM answer generation is still being built — by everyone, including Google. Being early to proper schema implementation is still the right move. Paying a premium specifically for “LLM citation via schema” is not yet a purchase backed by evidence.

What to Actually Do with Schema Right Now

Implementation cost is low. Downside is essentially nil. The case for schema is cumulative: if the frontier labs do build entity pipelines that ingest structured data — and they probably will — the groundwork is already done. That’s the right framing.

What you should stop doing: presenting “the LLM returned content from the schema” as proof the schema was used. Run the duck test. If your evidence doesn’t survive a deliberately invalid schema, it’s not evidence. The same applies to vendor pitches — ask for the working, in writing, before the invoice.

Prioritize schema where disambiguation actually matters: new entities, brands with name collisions, organizations without a knowledge panel, personal entities competing with more established namesakes. That’s where the asymmetric upside sits.

For operators running crypto exchange lead generation or CDL recruitment campaigns, the more immediate citation-driving levers are content authority, structured citation consistency, and performance channel management that builds brand search volume — signals that LLMs can actually observe in the text of the web. Schema supports all of that work. It doesn’t replace it.

Schema is useful infrastructure. It is not a shortcut into AI-generated answers, and anyone selling it as one right now is working from a remarkably thin evidence base. Use it. Price it accordingly.

Originally reported by Search Engine Journal, June 2026.