Schema Markup for AI Search: Which Schema Types Matter in 2026

Most “which types matter” lists still open with FAQPage and promise a citation boost. Both are out of date. Google removed FAQ rich results on May 7, 2026, and the two real studies on schema and AI citations look like opposites until you read their fine print. If generative engine optimization is new to you, start there, then come back.

Does schema markup for AI search actually work?

Not as a citation lever, and Google says so itself. Its guide to optimizing for generative AI features, last updated on 10 July 2026, states that “Structured data isn’t required for generative AI search, and there’s no special schema.org markup you need to add.” Schema helps engines confirm facts they would otherwise guess, and the two best experiments on it show little beyond that. Our guide to optimizing for Google AI Overviews covers what does move Google’s surface.

2.2%

rise in ChatGPT citations for pages that added JSON-LD, a change indistinguishable from zero. Google AI Overviews fell 4.6% in Ahrefs’ study of 1,885 pages.

The two best studies on this disagree, and that is the honest starting point. Ahrefs tracked 1,885 pages that added JSON-LD schema against 4,000 matched control pages from August 2025 to March 2026. The result across engines: Google AI Overviews down 4.6%, AI Mode up 2.4%, ChatGPT up 2.2%. The last two are indistinguishable from zero, and Ahrefs could not pin the AI Overviews dip on schema because both groups were already declining. Every page already had 100 or more AI Overview citations, so the test says little about pages AI has never cited. Ahrefs put it flatly: “Adding schema produced no major uplift in citations on any platform.”

Otterly ran a different kind of test: a single-brand experiment on its own SaaS website, using three pages, 319 tracked prompts and seven AI platforms from December 2025 to March 2026. Brand coverage on those pages swung hard in both directions, with AI Overviews up 611% and ChatGPT down 71%, which is how two studies asking the same question ended up with opposite headlines. Otterly’s own write-up adds that competitors rose in the same period, which it ties to a Google algorithm change, and its summary also cites +1,500% for AI Overviews appearances across the whole site, a separate measure from the +611% on the three tested pages.

What the studies actually agree on

They agree on less than either headline suggests. Neither shows a reliable lift on ChatGPT or Perplexity, and both are cautious about Google: Ahrefs measured a small decline in AI Overviews, and Otterly’s gain arrived while competitors rose too. The vendor line is more positive. Google said in April 2025 that structured data gives an advantage in search results, as reported by Search Engine Land, and Microsoft’s Fabrice Canel said in March 2025 that schema helps its LLMs understand content for Copilot. Google’s July 2026 guidance for generative AI search adds that no special markup is needed, and the measured lift for AI citations is small to nil.

Why prevalence is not proof

Here is the trap most pages fall into. Ahrefs found that 53% of AI-cited pages use JSON-LD schema, almost three times the rate of non-cited pages. That reads like a smoking gun until you get to Ahrefs’ own caveat, that the gap could easily be correlation rather than cause. Sites that add schema also tend to publish stronger content and earn more links, so prevalence is a hint rather than a verdict.

Which schema types for AI search matter most?

Mark up facts AI cannot safely infer. Tier 1 is Organization, which says who you are, and Product/Offer, which states price, availability, and rating. Tier 2 is Article, BreadcrumbList, WebPage and Person, which describe page structure and authorship. Everything else is optional. For an agent or brokerage, RealEstateAgent takes the Tier 1 slot, as the real estate section below explains.

One more correlation cross-check. An empirical study from rocketblue (formerly Spotlight) counted the schema types on 5,499 websites cited by AI models. The most frequent, from ImageObject and ListItem down to Organization and Product, are mostly page-furniture types a normal CMS emits on its own, which is prevalence again and no ranking of what wins.

Tier 1: facts AI cannot guess

Two types earn their place before anything else. Organization tells the engine your name, logo and sameAs links to profiles like LinkedIn, Clutch and Crunchbase, so it can resolve who you are and attribute a claim to you, while Product and Offer state the numbers a machine will not invent: price, priceCurrency, availability, GTIN, brand and AggregateRating. Guessing wrong is worse than not answering, so AI leans on the markup when it exists.

Tier 2: page-structure hygiene

The rest is clean-up. Article and BlogPosting confirm the author and publish date, which supports freshness for time-sensitive queries, BreadcrumbList describes where the page sits in your site, and WebPage and Person round out the entity picture and feed E-E-A-T signals. None of it wins a citation on its own, and one rule holds across every type: only mark up what is visible on the page, because schema-only facts do not get extracted and hiding them breaks Google’s own guidelines. Our technical rail on llms.txt covers the sibling question of what crawlers can reach in the first place.

Is FAQ schema deprecated in 2026, and should you remove it?

FAQPage is demoted, not deleted. Google removed FAQ rich results from Search on May 7, 2026, so the markup no longer earns a visible result, but you do not need to remove it, and a clean question-and-answer block still helps AI because engines pull the answer as text whether or not it sits inside a FAQPage wrapper. That is the correction almost every ranking page has missed.

The phase-out began in 2023, per Search Engine Journal, and the rich results stopped appearing on May 7, 2026. Rich Results Test support was scheduled to end in June 2026 and Search Console API reporting in August 2026. Search Engine Journal’s report is direct about what to do next: FAQ structured data “can stay in place. The markup won’t cause problems, but it also won’t produce visible results in Google Search.”

What you should not do is keep listing FAQPage as the number one schema for AI search, which most guides still do months after it lost its feature.

What schema can ChatGPT and Perplexity actually read?

Not much, directly. In Otterly’s test, six of seven AI platforms could not fetch or correctly interpret schema markup when asked directly. Gemini was the lone exception that retrieved it. When Otterly hid a fact only inside FAQ schema, none of the platforms used it to answer, and Google’s own surfaces were the only ones whose numbers rose. ChatGPT, Copilot and Perplexity lean on readable on-page content, not your JSON-LD. If your plan is “add schema so ChatGPT cites us,” the data says that plan is thin.

Those numbers come from Otterly’s experiment, 319 prompts across seven platforms; Claude was also tested and does not appear in its results table. Their conclusion, verbatim: “6 out of 7 AI Search platforms were unable to fetch or correctly interpret schema markup when directly asked.” Read the pattern, not each cell. Only Google’s two surfaces rose, Otterly reads the AI Overviews gain as a traditional SEO effect on SERP-feature eligibility, and ChatGPT spiked in December and January before correcting, with no isolated link to the markup.

For the engines that ignore your JSON-LD, the work is different. Our guides on how to get cited by ChatGPT and how to get cited by Perplexity go into the answer-capsule and third-party-mention work that actually moves those two.

What structured data do AI shopping answers need?

For AI comparison and shopping answers, product data needs GTIN, brand, availability, price, priceCurrency and AggregateRating, the core Product and Offer fields. Miss any one of those and your products can go invisible in AI shopping queries. SEOptimer notes that Product, Offer, and Brand markup is what keeps products showing accurately in AI-generated shopping results.

This is where markup is most likely to pay, since a shopping answer that names a price, a rating and a stock status cannot afford to guess: GTIN and brand make the item unambiguous, price, priceCurrency and availability are the numbers a buyer compares, and AggregateRating is what AI surfaces in “best” and comparison answers. On a real estate site the closest equivalent is the listing page, which gets its own section below.

Can schema markup hurt your AI visibility?

Yes, when it does not match the page. Otterly found that no AI platform could answer a question from schema-only facts, and Google AI Mode even reported schema types that were not on the page at all. An empty Product tag with no price and no rating declares a commercial fact and then fails to state it, which a machine reads as noise.

Warning: Never mark up something the page does not visibly show. Structured data describing content a human cannot see is a policy violation, and the specific case worth naming is aggregateRating on your own Organization: reviews you control make the page ineligible for star results and put a manual action on the table.

The underused point is that the properties are what carry meaning: price, rating, author and date. A Product label with nothing inside it tells an engine to expect commercial facts, then delivers none. Half-filled schema from a default plugin or website template is an easy failure to miss on agent and brokerage sites, and it is invisible until someone reads the raw output. So the test to run is whether every property is populated with a real value that also shows on the page.

A simple rule for deciding what to mark up

Mark up the facts AI cannot safely guess, treat the rest as hygiene, and use schema to verify facts, since neither study shows it winning citations on its own. Four steps, in order.

1. Verify, do not decorate. If AI could get the fact wrong, mark it up. If AI reads it fine from the page, markup adds little.
2. Entity and money first. Organization and Product/Offer are the facts with a right answer AI must not guess. Everything starts here.
3. Structure is hygiene, not a lever. Breadcrumb and WebPage help the engine parse the page. They do not win citations.
4. Populate or skip. Empty schema is worse than none. Fill every property or leave the tag out.

In our 500-query Houston study, an estimated 94.0% of the Houston Association of Realtors’ 50,000-plus members were never named in 2,000 AI answers. The studies above found little reliable citation lift from schema, so if your firm is missing from AI answers, the reasons your brand is invisible to AI run deeper than a missing tag. The Bright Realty case study shows what closing that deeper gap looked like for one real estate network.

Which schema should a real estate agent or brokerage use for AI search?

For an agent or brokerage, the Tier 1 entity is RealEstateAgent, a LocalBusiness subtype, on the homepage and every agent bio. Listing markup comes last, because Google has no rich result for property listings and a syndicated listing rarely wins the citation.

Houston study

43.6%

of all citations in our Houston data went to portals and agent referral networks, against 8.4% to 10.4% for a firm’s own domain.

Photo: Jeffrey Eisen, Pexels

The rule above still holds, but for an agent the facts AI cannot guess are different ones, and our AI search guide for real estate agents covers the wider playbook around them. Google’s search gallery lists a Local business feature, which RealEstateAgent falls under as a LocalBusiness subtype, and no feature for real estate listings. Here is the order we use:

1. RealEstateAgent on the homepage and each agent page. The name, address and phone in the markup must match the visible page and the Google Business Profile, areaServed should be real City or Place objects rather than one comma-separated string, and a license number in identifier must also be visible on the page.
2. Person for each agent, linked to the business with worksFor and to real profiles with sameAs, such as the state licensing lookup and the brokerage roster.
3. Organization for the brokerage, connected to its agents by @id references so the graph resolves to one firm.
4. FAQPage on neighborhood guides and process pages, with every question and answer visible as text.

The aggregateRating warning above applies here too, so never copy your Google Business Profile stars into your own markup. That profile matters to AI answers on its own terms: Local Falcon’s Realtor AI Visibility Index, a vendor measurement, reported that Google Business Profile supplied 36.6% of AI citations on “best realtor near me” queries across the 100 largest US cities, one of the sourced figures on our real estate SEO statistics page.

The listing-schema trap

RealEstateListing looks like the Product and Offer of real estate, and that resemblance is the trap. On an IDX site the same listing, with the same MLS description, appears on other agents’ sites and on the portals, so markup on a syndicated listing page rarely makes yours the version an engine picks. Our Houston data points the same way: a firm’s own domain was cited in 8.4% to 10.4% of the answers naming it, while portals and agent referral networks took 43.6% of all citations.

So mark up only the listings you index and add something unique to, such as your own photography or a written take on the property. Never mark up a page set to noindex, and change the status when a home sells, because a sold listing still marked as available states a false fact.

Where schema markup for AI search fits, and where it does not

Schema verifies the facts you already state on the page, and it will not tell you whether AI names your brand today. The Free AI Visibility Audit answers that second question.

It checks where you stand across ChatGPT, Perplexity, Gemini and Google AI Overviews, so you know whether the problem is your markup, your content, or the fact that no trusted source mentions you yet.