Every GEO pitch deck I've sat through this year has the same slide: a screenshot of JSON-LD code, a green checkmark, and a promise that schema markup is what gets you quoted by ChatGPT. It's a tidy story. It is not, according to the person who actually builds the tools that measure this stuff, how any of it works.
I want to be precise about the claim under scrutiny here, because "schema is dead" isn't it. The claim is narrower: that structured data SEO — specifically the JSON-LD sitting in your page's script tags — is a proven, direct lever for getting cited by ChatGPT, Perplexity, or Google's AI Overviews. Mark Williams-Cook, who runs AlsoAsked and spends his working life inside search data, spent a May 28th Substack post pulling that pitch apart, piece by piece. I think he's right. I also think most of the GEO industry, agencies that look a lot like ours included, has been selling the wrong mechanism.
The claim everyone repeats
You've heard some version of it. Maybe in a webinar, maybe in a proposal, maybe from a vendor demoing a schema generator with a scorecard that turns green once you've added enough markup. The pitch is always some flavor of: structured data helps the AI understand your content, so it gets you cited more often. It's a satisfying claim because it sounds technical and it sounds actionable. You can point at a line of code and say, we did the thing.
The problem is that "understand" is doing a lot of unexamined work in that sentence. LLMs don't browse your live site and read your JSON-LD the way a rich-results validator does. They were trained, mostly, on text scraped from the web months or years before your prospect ever typed a prompt. What happens between the moment you publish that markup and the moment a model might cite you is a pipeline with several lossy steps in it — and at least one of those steps throws your schema away before a model even exists to read it.
“"Add schema and the AI will cite you." It might be the most repeated sentence in GEO right now, and it is not close to how any of these systems work.”
Where structured data SEO promises meet the pretraining pipeline
Start with how these models actually get built. A huge share of the open, foundational training data used across the industry runs through FineWeb, a web-scale dataset assembled specifically for LLM pretraining. FineWeb's extraction step — the part of the pipeline that turns a raw HTML page into clean text a model can train on — uses a tool called trafilatura. Trafilatura's whole job is boilerplate removal: strip the nav bar, the cookie banner, the ad slots, the scripts, and hand back the readable article underneath.
Here's the part that should stop the pitch in its tracks. Your JSON-LD schema doesn't live in the visible article. It lives inside a `<script type="application/ld+json">` tag, and script tags are exactly what a boilerplate-stripping tool is built to remove. Williams-Cook's point isn't subtle: the markup you're paying to have written and validated is, for a large share of the data an LLM ever trains on, discarded before the model exists. It's not that the model reads it and decides to ignore it. It's that it was never in the room.
And that's the generous read, where your schema at least survives the stripping step. Say it does. You still have to get through tokenization, the process that breaks all text into the small chunks a model actually operates on. Tokenization doesn't preserve nesting, doesn't preserve key-value structure, doesn't preserve the parent-child relationships that make JSON-LD useful to a machine parser in the first place. What comes out the other side is a flat stream of tokens. The model isn't looking at an object with a type field and a name field and an address field. It's looking at a sequence of characters that happens to contain some curly braces. Structure is exactly the thing tokenization is bad at preserving, which is a strange thing to build a citation strategy on when structure is the entire value proposition of schema in the first place.
The experiment that undercuts the theory
Williams-Cook didn't stop at the pipeline argument. He ran a test, something he calls "Duck Yea," where he fed a model a page carrying nonsensical, invalid JSON-LD: made-up schema properties that don't exist in any real vocabulary, nothing a parser could actually validate. If the theory holds — if models are reading and trusting your structured data — a broken, fabricated schema block should have thrown the answer off, or at least confused it. It didn't. The model returned the correct address information anyway.
That's the tell. If the model had actually been parsing the structured data, feeding it garbage should have broken something. It didn't break anything, because the model was never parsing the structured data. It was reading the same plain-text content a person scanning the page would read, the same way it reads most of the open web. The schema was inert scaffolding standing next to the actual information source, not a channel feeding into it.
Even Google's own stack disagrees with itself
If any company should be able to make structured data reliably feed an AI answer, it's Google. They own the Knowledge Graph. They own Google Business Profile. They own the index the AI Overviews model draws from. And yet Williams-Cook points out that AI Overviews sometimes contradicts the very Google Business Profile listing for the same business, inside the same company's own stack. If owning every layer of the pipeline, from the structured data source to the model generating the answer, still isn't enough to guarantee consistency, that's a hard case to make for the rest of us feeding schema into a system we don't own at all.
It's worth putting this next to a separate but related correction from this summer. Google confirmed in June that Search doesn't use llms.txt at all — it won't help or hurt your AI Overviews visibility either way — which lines up with whether llms.txt moves AI citations at all, a question we ran the numbers on ourselves and got the same answer: no measurable effect on AI citations. Different file, different mechanism, same underlying lesson. An assumed connection between a technical artifact and an AI system's behavior needs to be demonstrated, not just plausible-sounding.
What structured data SEO is actually still good for
None of this is an argument to rip your schema out. Williams-Cook is explicit about that, and so am I. Structured data SEO still earns real things: rich results in classic search — stars, FAQs, breadcrumbs — plus a clean, low-cost, unambiguous hint to every non-LLM system that still reads a script tag the way it was designed to be read. Those benefits are real, they're cumulative, and they cost you almost nothing once the markup is in place. What structured data SEO is not, based on how these models are actually built and how they actually behave under testing, is a proven, direct lever for getting cited by an AI answer engine.
| WHAT SCHEMA MARKUP ACTUALLY EARNS YOU | WHAT IT'S BEING SOLD AS IN GEO |
|---|---|
| Rich results in classic search — stars, FAQs, breadcrumbs, eligibility | A guaranteed AI citation lever |
| A clean, low-cost signal for non-LLM systems that still read script tags | A causal input inside LLM pretraining and inference |
| One compounding technical-SEO signal among many | The centerpiece of a GEO strategy |
That distinction matters more than it sounds like it should, because budget and attention are finite. If your team is spending real hours this quarter treating schema as the centerpiece of a GEO plan, that's hours not spent on the schema AI engines actually read — the specific types and structural choices that do show up in what these engines end up quoting, along with the content structure and corroboration underneath it. Schema earns you a supporting role. It was never going to be the lead.
What to do with your schema now
Keep shipping schema markup. Validate it, keep it accurate, keep it matched to what's actually on the page — Google's classic crawlers and rich-results eligibility still reward exactly that discipline, and it's cheap enough that there's no real argument for skipping it. Just stop selling it, internally or to a client, as the reason an AI engine is going to start citing you. Redirect the conversation you were going to have about markup toward the one that actually moves the needle: is the page reachable and renderable by the crawlers these models depend on, is the answer sitting in the visible text near the top instead of buried in a script tag, and is the page corroborated by sources these engines already trust.
That's the version of the work we lead with in our SEO services now, and it's the same recalibration we run in our SEO and GEO audits — the kind we did with the team at MarketCheck this year, and the same lesson JavaScript-heavy, headless sites already learned the hard way about the gap between what you shipped and what a crawler can actually see.
You don't need to feel behind for having schema at the top of your GEO checklist. Most of the industry put it there for a reason that sounded airtight at the time. What would be behind is still leading with it next quarter, after Mark Williams-Cook's breakdown of exactly how the pretraining pipeline treats it. Move it down the list. Put crawlability, visible-text structure, and third-party corroboration back at the top, where the actual evidence says the citations are coming from.
See where you are cited today
A free snapshot audit of your rankings and AI citations before we ever talk.
Josh leads work at the intersection of SEO and generative engines at Something Inc., helping B2B brands get ranked and cited across every major AI engine.