Every AI Engine Plays by a Different Rulebook

The Measurement

For every cited page that returned HTTP 200, we checked whether it carries structured data (JSON-LD schema markup) and which types. That gives us a schema fingerprint for each engine's citation pool: 2,355 pages for Google AI Overview, 2,870 for Perplexity, and 182 for ChatGPT, out of 5,090 AI-cited pages in total.

The headline: 80.8% of the pages ChatGPT cites carry schema markup, against 59.5% for Perplexity and 48.4% for Google AI Overview.

The engine with the least access to Google's ranking machinery leans hardest on what the page declares about itself in code.

What ChatGPT's Picks Have in Common

Break schema down by type and the picture sharpens:

ChatGPT leads on every type except one. FAQPage, where all three engines sit at the same 9%. That's a finding worth pausing on, because FAQ schema is the single most hyped AEO tactic of the last two years, and it's the one type that separates nothing.

Now look at what ChatGPT's top three types are actually saying, read together. Organization: this page belongs to a declared institution. BreadcrumbList: this page sits in a known place inside a site's hierarchy. Article: this page is a piece of published content. The pattern isn't "more markup." It's machine-readable identity. The pages ChatGPT quotes are the ones that state, in code, who published them and where they belong.

The Objection That Almost Killed This Finding

There's a real hole in the argument, and it deserves to be taken seriously, because it nearly sank the whole thing.

ChatGPT doesn't cite the same kinds of queries as the other engines. In our entire study it returned zero citations for informational queries and zero for YMYL. Its citation pool is 36% commercial, 40% transactional, 24% local. Commercial and transactional pages carry more schema by their nature. Online stores mark up products; comparison sites mark up articles and breadcrumbs. So maybe ChatGPT doesn't prefer schema at all. Maybe it just answers shopping questions, and shopping pages happen to be well marked up.

The test is to hold the query category constant and compare engines inside it. Inside commercial queries only, ChatGPT cited 67 pages, and 100% of them carry schema markup, against 52% for the pages Google chose for the same category. Inside transactional queries, 77% against 55%. Inside local, the gap narrows to 59% against 53%, with Perplexity highest at 68%.

The category objection explains part of the headline gap. It doesn't explain the phenomenon. Competing for the same kind of query, drawing from the same pool of candidate pages, ChatGPT keeps landing on the ones with structured data. In commercial queries, in our sample, it cited nothing else.

Why the Engines Diverge

We can only observe correlation here, but the pattern fits a simple mechanism.

Google AI Overview sits on top of two decades of ranking infrastructure: link graphs, click behavior, crawl history. It already knows which pages to trust before it reads a line of markup. Schema is one signal among hundreds, so its citation pool looks like the web at large, roughly half marked up.

ChatGPT's web citations come from a young retrieval system with none of that behavioral memory. When it fetches a page, the page's own declarations are a large share of the trust evidence available. A page that states "I am an Article, published by this Organization, sitting here in this site" clears a bar an unmarked page can't. The result is the pattern in our data: an 81% schema rate, concentrated in identity types.

Perplexity, running its own index with its own ranking layer, lands in the middle at 60%. The engines aren't applying different taste to the same shortlist. They're building different shortlists, with different evidence on hand, and schema matters most exactly where the other evidence is thinnest.

Adding Schema to a Page Doesn't Buy It Citations

The loudest recent answer to "does schema markup help AI citations" is a flat no, and it comes from a better-designed test than ours. Ahrefs tracked 1,885 pages that added JSON-LD schema between August 2025 and March 2026 and compared them against 4,000 matched pages that never did. Thirty days after the change, citations in Google AI Overviews were down 4.6% against the controls and statistically significant. Google AI Mode was up 2.4% and ChatGPT up 2.2%, neither distinguishable from noise. Their conclusion: adding schema produced no major uplift on any platform.

That result and ours answer different questions. Ahrefs measured what happens when you add schema to a page, a before-and-after on the same URL. We measured what the pages engines already chose have in common, a profile of a citation pool. A page can be far likelier than average to carry schema without schema being the thing that got it picked.

Their design also carries a boundary they name themselves. Every page in the treatment set already had 100+ AI Overview citations before any schema was added, so the test asks whether markup lifts a page that is already inside the consideration set. It says nothing about how a page enters that set, which is the population our citation pools describe.

One piece of evidence cuts against our own mechanism story. In October 2025 searchVIU published a test page with prices placed only in JSON-LD, hidden Microdata, and hidden RDFa, then queried ChatGPT, Claude, Gemini, Perplexity, and Google AI Mode. None of the five returned a price that existed only in the markup; every system read the visible HTML instead. If engines don't parse structured data at fetch time, schema cannot be doing the work in the moment ChatGPT decides what to quote. What it can still be is a marker of the kind of site that gets picked: one with the technical maintenance, entity declarations, and clean indexing that structured data usually travels with.

The confound we cannot rule out is ranking position. Our cut holds query category constant, not rank, and well-ranked pages both carry more schema and get cited more, so part of every rate on this page travels with the ordinary advantage of ranking. One piece of the finding resists that reading. ChatGPT has no Google ranking machinery to inherit, and it shares zero citations with Google AI Overview in this dataset, so its 80.8% cannot be a Google-rank proxy in the way Google's own 48.4% partly is. That is an argument, not a control. We did not collect rank for the ChatGPT pool and cannot settle it with the data we have.

SEO Expert Advice

• If ChatGPT visibility matters to you, expect the pages it quotes to be marked up. In our commercial-query sample, ChatGPT cited zero pages without schema. The minimum stack our data points to is Organization plus BreadcrumbList plus Article, or Product for commerce pages. Build it as basic machine-readable identity, and read it as the profile cited pages carry: the controlled test above found no citation lift from adding markup to a page.
• If Google AI Overview is your target, schema helps and is worth doing, but don't expect it to carry you. Half of Google's AI citations have no schema at all. It inherits Google's broader ranking judgment, so the classic work (authority, links, content quality) still dominates.
• If Perplexity is your target, you're on the middle path. Schema adoption among its citations runs meaningfully above the Google baseline, especially in local queries.
• On FAQ schema specifically: every engine cites FAQ-marked pages at the same 9% rate. Nothing in our data supports FAQ markup as an AI citation lever. If you add it, add it for a different reason.

None of this calls for a seven-layer schema stack. The types that separate the engines in our data are the identity types, and the implementation detail for everything else sits in technical AEO foundations.

And before any of that, know which engine your buyers actually use. The engines don't share citations, so winning one gets you nothing in the others by default.

Is your schema stack complete?

Run a free AEO audit and see which structured data your pages declare, and which engines are citing you today.

Run the Free Audit

Methodology

Dataset. 500 queries, 100 in each of five intent categories (informational, commercial, transactional, YMYL, local), US market, English. Citations captured for Google organic, Google AI Overview, ChatGPT (web search enabled), and Perplexity. All 8,397 unique URLs crawled; schema extracted from JSON-LD.

Sample for this article. 5,090 AI-cited URLs with HTTP status 200, of which Google AI Overview cited 2,355, Perplexity 2,870, and ChatGPT 182. (Engine pools overlap slightly, so they sum to more than the total.)

Limitations. This is a correlational design. We observe what cited pages carry; we did not add or remove schema on any page, so nothing here establishes cause. Ranking position is uncontrolled: we held query category constant, not rank, and part of every schema rate reported above travels with the ordinary advantage of ranking well. For a controlled read on causation, see the Ahrefs schema study discussed above.

Sample caveat. ChatGPT's 182 pages are a small pool next to Google AI Overview's 2,355 and Perplexity's 2,870, and the commercial-query cut inside it rests on 67 pages. Those are the counts the crawl produced, and they are why the ChatGPT percentages should be read as directional.

FAQ: Schema Markup and AI Citations