Which schema types correlate with AI Overview citations

Markup earns its keep when it carries a fact someone might quote.

Most arguments about schema and AI citations ask whether. The better question is which. Practitioners who marked up client sites with care have reported no noticeable increase in AI citations against sites without it, and others note that AI Overview results cannot be reproduced consistently, because the answer shifts with each user's context and location.

So a single yes or no was never going to hold. Look closer, though, and the types begin to separate.

Search Engine Land published a small test in September 2025 that made the point quietly. Three pages, three keywords, all submitted to Google on Aug. 29, 2025. The page with complete Article, FAQPage and BreadcrumbList markup ranked for 6 keywords and was the only one to reach an AI Overview. The page with incomplete markup and a wrong date format ranked for 10 keywords and never appeared in one. More keywords, no citation. The authors called their result "promising, but inconclusive", which seems fair for three pages.

One practitioner described FAQ markup as "a small clarity boost," worth having where answers are tight and intent is plain. That reads about right to me. The research section below walks through these findings. Start with the figure that splits them most.

Which schema types correlate most with AI citations?

Populated Product and Review markup correlates most strongly with AI citation. Question-shaped FAQPage follows, loosely. Generic Article, Organization and BreadcrumbList schema trail some way behind.

Before the numbers, here is how I would sort the markup on any site:

1. List every schema type the site ships, template by template.
2. Mark the ones that carry real values: prices, ratings, specifications, answers.
3. Mark the ones that only declare a type and a name.
4. Put new effort into the first group, and leave the second alone for a while.

The clearest type-level figure comes from Kurt Fischman's February 2026 study, as EPR reported it. Pages with Product or Review schema and populated fields (pricing, aggregateRating, specifications) were cited at 61.7%. Pages carrying only generic types such as Article, Organization or BreadcrumbList were cited at 41.6%. He looked at 730 AI citations from ChatGPT and Gemini across 75 commercial queries, and the gap held after controlling for the higher Google rankings of schema-bearing pages.

That last detail is the useful one. It separates citation from ranking, which most correlation work never tries to do.

A 6-million-URL correlation study, as Search Engine Journal reported it, points the same way at a coarser grain. AI-cited pages were almost 3x more likely to carry JSON-LD, and roughly 53% of them ran schema. Ahrefs reads that as a mark of well-kept sites rather than a schema effect, which is fair. A separate dataset in the same discussion put pages with JSON-LD at 38.5% cited, against 32.0% without.

The question-shaped types are murkier. One B2B SaaS team rolled out FAQPage across 40 product and category pages and saw AI Overview appearance rise from 14% to 22% of tracked queries, though its rankings and new posts climbed over the same weeks. A vendor case study reported a 233% rise in AI brand mentions after adding FAQs and schema together, measured in ChatGPT rather than AI Overviews. Both point up. Neither isolates the markup.

The common assumption is that schema is one switch. It is several, and they are not wired to the same light. An analysis of 4 sources shows the same slope: the more specific the fields, the stronger the link to citation. Open one of your product templates and count the empty fields. An empty AggregateRating is a label on a jar with nothing in it.

AEO Content has run 26,577 AI-visibility audits, and I would rather every one of them asked which schema than whether. If you want the reasoning behind how we weigh studies like these, our page on research-driven content built on evidence sets it out. For now the ordering is plain enough. The research section below walks through why.

If schema correlates with citations, why do controlled tests show no lift?

The largest controlled test asked a narrower question. It added pooled schema to pages already heavily cited and watched for 30 days. A type-level correlation can survive that.

In May 2026, Search Engine Journal reported on an Ahrefs test that tracked 1,885 pages adding JSON-LD against 4,000 matched controls. No platform moved in any meaningful way. Google AI Overviews fell 4.6%, AI Mode rose 2.4% and ChatGPT rose 2.2%, and Ahrefs called both gains too small to tell apart from random variation.

The decline sounds worse than it is. It came to about 12 fewer daily citations per page, in a sample where most pages received hundreds, and treated and control pages were already sliding before any markup went in. The data was analyzed four different ways. All four pointed the same direction, which is to say nowhere much.

The fine print is where the two sets of results stop quarrelling.

• Every page was already cited. Each had 100+ AI Overview citations before schema arrived, so the test is silent on pages not yet seen.
• All types were pooled. FAQPage, Product and Article went into one bucket, and a type-level effect could sit unnoticed inside the average.
• The window was 30 days. Slower effects would not show.
• Live fetch skips markup. A searchVIU experiment cited in the same report found five AI systems fetching pages in real time read only visible HTML, ignoring JSON-LD, Microdata and RDFa.

So the test measured whether schema pushes an already-visible page higher. It does not, or not by much. Whether markup helps a page get chosen at all is a separate question, and the test was never built to answer it.

A small Search Engine Land experiment sits on the far side of that line. Three new pages, three keywords with matched metrics. Only the page with complete Article, FAQ and Breadcrumb schema appeared in an AI Overview, peaking at Position 3, while the page with incomplete schema ranked for more keywords, topped out at Position 8 and surfaced in no AI Overview at all. The page with no schema was crawled but never indexed. The authors called it promising but inconclusive. I would go further toward modesty: three pages is a handful of pebbles, not a beach.

Practitioners on technical SEO forums add a fair point of their own. Nobody among them has isolated schema as a single variable and then tracked citations before and after. Understanding a page and choosing it as a source are, they note, two different acts.

Held side by side, the results are less contradictory than the headlines make them. Strong correlation by type. Near-zero lift from bolting pooled markup onto pages that already win. If the fetch mechanics are new to you, the guides in our AEO knowledge base are a reasonable place to begin, before deciding what to put inside the markup and on which pages it belongs.

How should a B2B SaaS team build schema that supports AI citations?

Fill the fields that carry facts, write question markup from questions people actually ask, connect every entity through a stable ID, and keep the markup honest as pages change.

1. Populate Product, Service and Review markup first. Prices, ratings, specifications. Start with the templates your sales team links to most.
2. Write FAQPage answers from real support questions. Not from a keyword tool. The answer in the markup and the answer on the page should be the same words.
3. Declare each entity once, in a single @graph, with a stable @id. Organization, WebSite, each Person. Reference them everywhere else rather than redeclaring them.
4. Audit the markup whenever the page changes. Staff leave, prices move, products retire. The schema should follow them out.

Cap Puckhaber's client notes show what happens when the last two steps are skipped. On a financial services site with 340 published articles, the schema listed eleven author bylines while the writing staff numbered six. Three writers who had left were still active authors in the structured data eight months after their last article. Moving to one @graph with six Person entities, each on a stable @id, removed the phantom bylines within one sprint.

None of that threw an error. The pages still passed the Rich Results Test, which confirms eligibility and says nothing about whether the nodes connect to one another.

His healthcare case is the encouraging one for identity markup. Adding knowsAbout declarations for six clinical specialties took the site from AI Overview citations on 2 of 6 specialty terms to 5 of 6, four months later. A B2B SaaS client that added worksFor, sameAs and knowsAbout to its twelve most active authors saw four of them named in AI-generated answers for branded queries within ten weeks, with no article copy changed. In Puckhaber's own words, "The fix isn't more schema. It's connected schema, built once and referenced everywhere."

Identity types earn their keep when they connect. Check each entity against the people and the company it claims to describe. That is the quiet correction to all the effort sites put into a lone Organization block on the homepage.

AEO Content has scored 11,000+ domains across 15 sectors and 28 categories, and the team behind that scoring includes a full-stack engineer and content infrastructure architect with 20 years of building enterprise systems. Alex and I bring a combined 40+ years of SEO and content infrastructure experience to the work. I still think of schema as plumbing. Dull, mostly hidden, and expensive when one joint is wrong.

If you are weighing where markup sits among the rest of your AEO work, our answers on AI search optimization go through the common questions in order. What none of the steps above can tell you is whether they worked on your own pages, and that needs a test design before it needs more markup.

What should you mark up first if you want AI Overview citations?

Mark up the facts a buyer would ask about: product, price, reviews and real questions. Generic Article and Organization tags can wait, because they rarely carry anything an engine can quote.

A 2023 ShreyWebs explainer drew the line before AI Overviews made it urgent. Product markup, it said, carried price, availability and reviews. Article markup was only "possibly enhancing" visibility. The hedge was telling. The research section above walks through these findings.

What I cannot yet give you is the audit figure this article set out to find: how often each schema type sits on a page that Google AI Overviews actually cite. That count would settle it. Until it exists, I measure the outcome instead: AEO Content's promise is to get you named in AI answers in 90 days, or we work free until you are.

One result of ours went from a quiet site to 10 booked calls in 3 weeks. It was measured in calls, not markup. That is the only ledger I trust, and it brings the question back to your own pages and the fields still sitting empty on them.