Schema and proof in GEO: why LLMs lower the bar for evidence
Summarize this article with AI
« $8,000 in schema for zero citations »: the call that froze me
Tuesday morning. The phone buzzes on my desk in Southeast Asia. On the line, an e-commerce director. His voice is calm, but tense. « Stéphane, I did everything I was told. Schema Product, Offer, AggregateRating, Organization… I spent $8,000. I don’t have a single citation in ChatGPT Search. Nothing in Gemini. Nothing in Perplexity. »
On his site, 327 pages. The schema is clean, validated by Google’s tool. But the pages are disconnected. No contextual linking. No silo logic. The content is good, but it floats in an ocean without a lighthouse.
And this isn’t an exception. Since January 2025, I’ve observed the same pattern across 11 out of 14 GEO audits. Massive investments in markup, with an AI promise… and zero results. Why? Because people apply a magic recipe without understanding how LLMs actually digest a page.
Mark Williams-Cook from Candour wanted to settle it. In an experiment covered by Search Engine Journal, he built the most absurd page possible to see if LLMs fell for it. They did. And what that reveals is far graver than a simple technical bug.
I’ve built systems that run without me since 2016. Schema is a tool, not an end goal. But when I see $8,000 swallowed by markup that nobody correlated to an architecture, I know it’s time to explain the real mechanics.
A duck, a fake address, and deliberately broken JSON-LD
The experiment is simple and effective. Mark Williams-Cook created a web page entirely dedicated to ducks. No connection to any business. In the <head>, he injected a JSON-LD block containing a fictional address: « 123 Duck Street ». The JSON-LD was intentionally invalid—broken syntax that the official validator rejects. In the page body, not a single mention of this address.
He then asked several large language models: « What is this company’s address? »
100% of the LLMs answered correctly. Perplexity, Gemini, ChatGPT Search… all restituted the address of the « duck company ». Several even cited the « structured data » as their source, as if they’d read the Schema.org manual.
On LinkedIn, part of the GEO community erupted. « Proof that LLMs exploit schema! » In reality, it was the opposite. The schema was broken. The LLMs didn’t validate it, didn’t interpret it semantically. They read the plain text, curly braces and all, the way they’d read a line in an HTML comment or a <meta> tag.
« The schema was deliberately broken. The LLMs still returned the address. Not because they understood the structure, but because they were reading the text. » — Mark Williams-Cook
That’s the illusion. Schema is sold as a machine language that AIs « understand ». For a language model, JSON-LD is just a string of words with curly braces and colons. Not a semantic contract. Not a layer of structured meaning.
LLMs don’t understand schema. They digest it as text.
An LLM doesn’t interpret JSON-LD the way Google’s structured data API does. It doesn’t validate each Schema.org property. It doesn’t build an entity graph. It reads the source code as a single text stream. JSON curly braces become tokens, just like <h1> tags or bolded words.
I redid this test with three client sites. The first: an e-commerce with 22,000 pages. Its headquarters address appeared only in the <script type="application/ld+json">—no visible text. ChatGPT, Gemini, and Perplexity never returned that address across 15 requests. The address was buried in a heavy header.
Then I placed the same address in the footer, as visible HTML, as plain text. Result: 100% restitution. A schema snippet was ignored, but the visible text was cited word-for-word.
The logic is straightforward. LLMs aren’t looking for a PostalAddress. They’re looking for textual proximity between the word « address » and a string that resembles a location. Schema provides that string, but with no more weight than if you’d written it in a paragraph. AI visibility comes from readable text, not markup.
So why is schema still prominent in GEO discourse? Because it’s a visible, billable technical layer. But its effect on LLMs is overstated. The duck test proves it: an absurd page, broken schema, and restitution that « experts » mistook for validation.
Why I still deploy schema for my clients, despite everything
Don’t make me say what I didn’t: schema isn’t useless. It remains an effective tool for classical Google Search. And it’s Google Search that, in many cases, indirectly feeds conversational agents. A site that gains SEO positions, that wins rich snippets, that clarifies its entities in the Knowledge Graph… that same site has a better chance of being cited by Gemini when a user asks a question.
Schema helps Google’s machine understand a page’s topic, content type, products, reviews. It boosts CTR, reduces crawl time, and strengthens perceived relevance. For my clients, clean deployment of Article, FAQ, Product, or LocalBusiness schema averages 12 to 18% more SERP clicks within 90 days. That’s concrete.
But watch out: don’t confuse this gain with a magic wand for LLMs. I’ve seen too many projects where 29 schema types are added across 400 pages without touching internal linking, hoping for an explosion of ChatGPT citations. The result is always the same: +2% AI mentions, no additional traffic, and a fat agency bill.
I’m not saying « stop schema ». I’m saying: stop selling it as the heart of GEO. Use it for what it is—a structuring veneer, not the engine. The engine is semantic architecture. No JSON-LD replaces that.
The article reveals a striking pattern: in 78% of audits, the schema deployed on sites had no detectable impact on AI citations. This donut visualizes the proportion of sites where schema remains invisible to large language models.
Schema exploitation in AI: the 2025 reality
In 11 out of 14 audits, schema was not utilized by LLMs
The real engine of AI visibility: my 14 audits tell you what it is
Since 2016, I’ve built more than 1,300 semantic cocoons. And what I see, audit after audit, is that AI growth doesn’t come from enriched schema. It comes from a structure where each pillar page gathers its satellites, where internal linking mirrors the user’s mental journey, and where thematic authority grows because signals are coherent.
Take a concrete example. An e-commerce client, 945 pages. Already invested $12,000 in advanced schema: Product, Offer, AggregateRating, events, breadcrumbs… nothing was missing. But when I audited it in April 2025, its architecture was a plate of spaghetti. No pillar pages. Mixed catégories. Product cards isolated, with no links to buying guides.
We dropped schema. We spent 4 hours mapping the domain.
We built 4 thematic silos. We merged 87 thin pages into 23 consolidated pieces. We implanted internal linking with contextual page-to-page links, respecting cocoon logic. Then we let it run.
Result by January 2026: 37,000 organic sessions per month, up from 14,000 before. +164%. And 12 generative queries ranked on transactional keywords, generating 11% more sales. Schema wasn’t touched. The difference was architecture.
I’m not alone in seeing this. In Mark Williams-Cook’s experiment, even broken JSON-LD was picked up by LLMs because the text was there. What matters is what’s readable, understandable, and credible in its organization. No markup gives you that. Only structure does.
Next time someone sells you schema for AI, ask this question
The GEO market is young. Evidence is thin. As the duck test shows, the bar for qualifying proof is too low. An LLM regurgitates an address in broken JSON-LD, and people deduce that schema is an AI lever. You have to read it the opposite way: the LLM just read text where there was text, regardless of the braces.
When a consultant tells you « look, the LLM restituted the schema address, so our GEO package works », ask: was the schema valid? Was the test run on a page where the address appeared nowhere else? Did you test multiple LLMs across varied contexts with different data volumes? You’ll know if the approach is serious.
Schema remains a useful tool for Google, for content clarity, for your CTR. But for LLMs, it’s a weak signal. The strong signal is your information structure, your thematic authority, and your pages’ ability to answer a question without the user clicking three times.
I’m not selling you the method. I’m showing you the pages. Your site’s next step isn’t a new schema type. It’s an audit of your current architecture. Because it’s silos, not tags, that AIs will remember.
In your opinion, how many pages on your site today are disconnected from their thematic pillar?
Your AI visibility audit in 37 minutes
I show you the silos that matter, the pages to consolidate, and where to start generating AI traffic without a line of useless schema. No jargon, no promises—just what your site deserves.
Book a strategic call — 45 minFrequently Asked Questions
Has schema markup become useless for SEO?
No, it’s still useful for Google and rich results. It clarifies content and boosts CTR. For LLMs, it’s nearly nothing. Use it as a structuring veneer, not as an AI visibility engine.
Should I stop implementing schema on my site?
Absolutely not. Keep the useful types (Article, Product, FAQ, etc.). But to generate citations in ChatGPT or Gemini, bet on semantic architecture and internal linking.
How do I verify if schema is exploited by LLMs?
Run the duck test. Create a page where info only appears in the schema, even if broken. Query 3 LLMs. If they answer, they’re reading text, not structure. That alerts you to their actual processing.
What’s the #1 factor for being cited by generative AIs?
I build thematic authority through silo architecture. Pillar pages linked to satellites. Content that matches long-tail queries. Navigation that follows the user journey. Schema? Far behind.
How long does it take to see results in GEO by betting on architecture?
For my clients, first AI citations appear 3 to 5 months after semantic cocoon creation. Organic traffic continues growing over 8 to 12 months. Impact strengthens as authority consolidates.
