28d • Entity Lab
The Duck Test: What Schema Actually Does in AI Search (With Receipts)
A fake t-shirt company called DUCKYEA settled the biggest argument in AI SEO. Sort of.
Quick setup. An SEO put a company address in JSON-LD only. The markup was deliberately broken — fake @context, made-up @type, nonsense properties. The address appeared nowhere in the visible text. Then he asked ChatGPT and Perplexity where the company was based.
They both returned the address.
Half the industry took the wrong lesson. So let's take the right one, because this changes how you should spend your schema hours.
What the duck actually proved
LLMs read your JSON-LD as plain text. Not as a graph. Not as structured data. They tokenize your entire HTML — script tags included — and "@type": "Organization" becomes character soup just like everything else on the page.
That's why invalid, fictional schema got extracted exactly the same as valid schema would. The structure did nothing. The text presence did everything.
So when someone sells you "schema optimization for ChatGPT rankings" — that mechanism doesn't exist. There are zero peer-reviewed studies showing structured data directly improves citation rates in ChatGPT or Perplexity. Zero.
But here's the other half nobody pairs with it
The platforms went on record. Both of them.
Fabrice Canel — the guy who runs Bing's crawling infrastructure — confirmed on stage at SMX Munich that schema helps Microsoft's LLMs understand your content. That covers Bing Copilot.
Google's structured data engineer Ryan Levering, same month: "A lot of our systems run much better with structured data." His reason? It's computationally cheaper than extracting facts from your prose. Then in April at Search Central Live Toronto, he said something even more specific: schema is used as context served to models when doing query fan-outs.
Translation: when Google's AI surfaces break your query into sub-queries, your structured data rides along as context.
So both camps are right. About different layers.
Layer 1: The retrieval layer. AI Overviews and Copilot sit on top of search indexes and knowledge graphs that have parsed schema for a decade. Your JSON-LD feeds entity reconciliation there. Confirmed by the engineers who build it.
Layer 2: Direct LLM reading. ChatGPT or Perplexity hitting your page at answer time sees your schema as text. Still useful — a clean key-value declaration of your address, price, or founder is harder to hallucinate around than a paragraph. But no graph parsing is happening.
Schema is not a cheat code. It's a translation layer. And it only translates for the systems built to listen.
What to actually do (this week, not someday)
- Keep your schema valid for Google and Bing. That's where the parsing happens, and AI Overviews inherit it.
- Stop redeclaring your entity differently on every page. Pick stable @id URIs for your Organization, your Person, your WebSite. Connect them. One canonical identity, referenced everywhere. This is the entity graph — and it matters way more than which schema plugin you use.
- Make sure every fact in your schema also exists in visible text. The duck proved LLMs read the page, not the markup. If it's only in JSON-LD, you're relying on tokenized luck.
- Audit your author/founder consistency. String on one page, Person node on another, three name spellings across the site? That's entity ambiguity, and ambiguous entities get skipped or described wrong.
- Run the test yourself. Ask ChatGPT, Perplexity, and Gemini what they know about your business. Save the answers. That's your baseline. We'll compare notes.
The bigger frame
AI engines retrieve entities, not vibes. Your human-facing brand can be strong while your machine-facing entity is weak — and most businesses have exactly that gap. The fix isn't more content. It's structure at the root.
SEO didn't disappear. The retrieval layer got added.
Sources if you want to go deeper:
- Mark Williams-Cook, "Schema, LLMs and the Low Bar for Evidence" (the duck experiment, original credit Richard Barrett) — markwilliamscook.substack.com
- Search Engine Land: Bing/Copilot schema confirmation (Fabrice Canel, SMX Munich)
- Search Engine Journal: Ryan Levering quotes, Search Central Live NY
- Suganthan Mohanadasan, "The Three Lives of Schema Markup" (Levering's Toronto fan-out quote)
Drop your baseline test results below. What do the engines currently believe about your business? Wrong answers are the most useful ones — post those first.
0
0 comments
The Duck Test: What Schema Actually Does in AI Search (With Receipts)
skool.com/alex-rodriguez-seo
A practical AI visibility lab for business owners, marketers, and operators who want to get found, trusted, cited, and selected.
Leaderboard (30-day)
1
+1
Powered by