“Add schema and you’ll get cited by AI” is one of the stickier promises in this space. We were skeptical, so we did the literal thing: we pulled the pages that AI cites most often in our own data and read their structured data tag by tag. The answer is more useful than the promise. Most cited pages do carry schema, there is a recurring stack worth copying, and the honest verdict is that schema is an amplifier, not a shortcut. Here is what the cited pages actually use, what it is doing, and where it does nothing.
First, what Google and the models actually do with schema
Set the expectation before the tactics. Google has said for years that structured data is not a ranking factor, and its guidance on AI features is that there is no special schema for AI Overviews or AI Mode; whatever you mark up should simply match what is visible on the page. So schema is not a switch that turns on AI citations.
It goes further than that. A 2026 controlled test found that ChatGPT and Perplexity read your JSON-LD as raw text: the model tokenizes the script block like any other string on the page rather than parsing it as structured data. So the model is not getting a private, machine-only briefing from your markup. It reads the same facts a human would, in your visible content and, incidentally, in the markup. That single fact explains everything the data shows next.
What the cited pages actually mark up
We read the structured data on a sample of the pages AI cites most in our measurements. The sample skews toward ecommerce and WordPress categories with some editorial and health pages mixed in, so read the types below as a pattern, not a census. Grouped by what kind of page they are:
| Page type | Schema the cited examples carried |
|---|---|
| Product, tool, or plugin page | Product plus SoftwareApplication, sometimes FAQPage |
| Best-of listicle | Article or BlogPosting, BreadcrumbList, a Person author, often FAQPage or ItemList |
| How-to guide | BlogPosting or Article with author and date; HowTo on the most thorough |
| Reference or docs | Often almost nothing beyond a BreadcrumbList |
| Health or finance explainer | A specific type such as MedicalWebPage, with a named reviewer |
The recurring spine across most of them was the same: the content-type entity (Article or BlogPosting for editorial, Product plus SoftwareApplication for tools), plus the sitewide trio of Organization, WebSite, and BreadcrumbList, plus a real Person author and a datePublished on the editorial pages. FAQPage showed up again and again on listicles and feature pages, and the single most decorated page we read stacked Product, FAQPage, HowTo, and VideoObject together.
Then the exception that keeps the rest honest. Two of the most-cited pages we read carried almost no content schema at all: one reference doc had only a breadcrumb, and one best-tools listicle had only a site-level tag. Both were cited well over a hundred times. Schema is clearly not required to get cited. Strong, extractable content on a trusted site wins with it or without it.
So what schema is actually for
Put the controlled test and our data together and they agree: schema is an amplifier, not a driver. On thin content it does nothing, which is why a perfectly marked-up weak page stays uncited. On content that already deserves to be cited it compounds, through two real mechanisms.
First, Google uses your markup to understand and confidently index your page as an entity, and Google’s index is what feeds AI Overviews, AI Mode, and Gemini. Clean markup helps you get into the surfaces that read Google, which is a different job for each engine, as the index map lays out. Second, the facts you mark up (the author, the date, the price, the answer to a question) are the exact facts a model lifts, and it will read them whether from your JSON-LD or your visible text. Marking them up forces you to state them clearly in both places.
That is also the one rule Google is firm on: your structured data must match the visible content. Schema is not a back channel to the model. It is a clean restatement of facts that have to be on the page for a human anyway.
The stack worth copying, by page type
- Every page: Organization and WebSite once at the site level, and BreadcrumbList on internal pages. This entity backbone was on almost every cited page we read.
- Article, guide, or listicle: Article or BlogPosting with a real, named author (a Person, with credentials for finance or health), plus datePublished and dateModified you actually keep current. Add FAQPage only for genuine questions and answers on the page, ItemList for a ranked list, and HowTo for real numbered steps.
- Product, tool, or plugin: Product plus SoftwareApplication with offers and price, and aggregateRating only when the ratings are real and shown on the page.
- Finance or health (YMYL): the specific page type plus a named reviewer. The models are conservative here, so the credential signals carry more weight than the markup does.
What not to bother with
- Schema theater: marking up FAQs, ratings, or content that is not visible on the page. Google can pull your rich results for it, and since the model reads the rendered text anyway, invisible markup buys nothing.
- Expecting schema to rescue thin content. Our two least-marked-up pages got cited on content and authority; plenty of immaculately marked-up thin pages never do. Fix the content first.
- Chasing a single markup type as an “AI signal.” Google has not tied FAQ markup to AI Overview inclusion, and FAQ rich results were largely retired in normal search, so add it for clarity, not for a phantom lever.
How to know it is working
Schema is an input to whether Google and the engines can confidently understand and surface you, not a number in itself. Judge it by whether the pages you marked up start getting named in AI answers for the prompts your buyers actually ask, engine by engine. That per-engine, did-it-get-cited view is the gap we built SurfacedBy to close. And the work that actually earns the citation, the specific, sourced, answer-first content that schema only amplifies, is the same craft we broke down in how to write content AI platforms cite.
The honest limits
This is a sample of the most-cited pages in our data, skewed toward a few categories, and schema presence here is correlational: well-built pages tend to have both structured data and citations. We did not run a controlled test that adds schema to a page and watches its citations climb, and the cleanest public attempt to do so found the engines reading the markup as plain text. So treat schema as table stakes for being understood, not as a lever you pull for citations. Mark up the facts that are genuinely on your page, keep them honest and visible, and put the real work into the content those facts describe.
