Schema Markup for AI Search: Does It Actually Get You Cited?

Schema markup is still worth adding to your site. It just won't get you cited more often by ChatGPT, Perplexity, or Google's AI Overviews. A study of 1,885 pages that added JSON-LD schema between August 2025 and March 2026 found no meaningful citation lift on any of the three AI platforms tested, and a companion test found the likely reason: these systems read what's visibly on the page, not the hidden markup underneath it. That's close to the opposite of the standard AEO advice, which usually tells you to add schema first.

Key Facts

• Across 1,885 pages that added JSON-LD schema, matched against 4,000 control pages, citation changes were statistically indistinguishable from noise on Google AI Mode and ChatGPT, and slightly negative on Google AI Overviews (Ahrefs, May 2026).
• A companion test of five AI systems found none of them used JSON-LD, hidden Microdata, or hidden RDFa when fetching a page directly; all five read only the visible HTML (cited in the same Ahrefs piece).
• Google fully retired the FAQ rich result on May 7, 2026, and removed its documentation on June 15, 2026. HowTo rich results were retired in 2023. Both schema types are still valid markup; neither produces a Google Search rich result anymore.
• GPTBot, ClaudeBot, and PerplexityBot do not execute JavaScript by default, so content that only exists after a client-side render, including schema injected by JavaScript, is often invisible to them.

What Schema Does and Doesn't Do for AI Search

Start with the table, because this is the part most AEO advice gets backwards.

Two different jobs are getting conflated in most schema advice: helping a page get found and formatted by traditional search, and helping a brand get quoted by a generative answer. Schema still does real work on the first job. The evidence says it does close to nothing on the second.

What the Study Actually Found, and What It Didn't

Ahrefs tracked 1,885 web pages that added JSON-LD schema between August 2025 and March 2026, then matched each one against a pool of roughly 4,000 control pages from different domains with similar pre-period citation levels that never added schema. They measured citation changes in the weeks before and after the schema addition date.

Read that precisely. The study doesn't say schema is worthless for everything, and it doesn't test every AI surface that exists. It tested one specific question: if a page didn't have JSON-LD schema and then got it, did that page start getting cited more often by Google AI Overviews, Google AI Mode, or ChatGPT? The answer, in this sample, was no on two platforms and a small negative on the third. Ahrefs put it plainly: you can't tell from this data whether schema did a tiny bit of good or nothing at all, because the lift doesn't clear the noise floor.

One more thing worth naming directly: Ahrefs sells Brand Radar, an AI visibility product, so it has a commercial interest in this category. A finding that undercuts a popular AEO tactic (schema) runs against that interest rather than flattering it, which is a reason to trust the finding more, not less.

Why the Mechanism Makes Sense

The study also points at a mechanical explanation, and the explanation holds up on its own logic. A companion test cited in the same Ahrefs piece checked whether five AI systems (ChatGPT, Claude, Perplexity, Gemini, and Google AI Mode) actually use schema markup when fetching a page in real time. None of them did. During direct retrieval, every system extracted only the visible HTML content. JSON-LD, hidden Microdata, and hidden RDFa were all ignored.

That tracks with how these systems are built. A large language model answering a question from a fetched page is doing something closer to reading the page the way a person would: the words, headings, and visible text, not a hidden data block meant for machines. Schema was designed for a different reader, a search engine's structured-data parser, and generative answer systems mostly aren't using that parser when they decide what to quote.

The practical implication is the single most useful idea in this article: put the fact in visible text, not only in markup. If your JSON-LD says a plan costs $49 a month but the page itself just says "contact us for pricing," an assistant quoting that page says "contact us for pricing," because that's what it read. The $49 exists only to a parser the assistant isn't running. The same goes for a spec, a release date, a step in a process, or an address. If a fact lives exclusively in structured data, it is effectively invisible to the thing you're trying to get cited by.

This also reframes how to think about AEO versus SEO as disciplines: a tactic can be real SEO practice and still do nothing for AI citation, and schema is the cleanest example of that gap found so far.

What Schema Genuinely Still Earns

None of this makes schema a wasted afternoon. It still does real, verifiable work, just not the work most AEO content promises.

Rich results are real, but the list changed in 2026. Google's current structured-data gallery still supports Product, Organization, Review and AggregateRating, Article, LocalBusiness, Event, and Breadcrumb rich results. FAQ and HowTo are gone. Google stopped showing the FAQ rich result on May 7, 2026, removed the FAQ report from Search Console and the Rich Results Test in June 2026, and pulled FAQ support from the Search Console API in August 2026. HowTo lost its rich result earlier, retired from mobile in August 2023 and desktop in September 2023. Google's own position is that the markup itself "is still valid markup that Google will continue to use to better understand pages," but neither schema type produces a visible search feature anymore on most sites. Before 2023, FAQ rich results were narrowed to mostly government and health sites anyway, so most businesses had already lost the practical benefit well before the full retirement.

Entity and organization signals still feed the knowledge graph. Organization and Person schema help a search engine, and by extension an assistant pulling from that engine's index, resolve who a company or author actually is. This is a supporting signal, not a citation driver on its own, and it pairs with the brand-consistency work covered in our GEO vs AEO vs SEO comparison: the same name, address, and facts repeated consistently across your site and the web.

Product and Offer schema do a genuinely different job: feeding a structured shopping surface directly, not getting read during a text citation. Platforms building AI shopping features ask merchants to submit Product schema with price, availability, and identifiers like GTIN or MPN through a feed, the same pattern as Google Merchant Center. That's direct, structured ingestion by design, the opposite of the page-reading scenario the citation study tested. It still depends on the visible price and availability matching what the feed claims, so the "put it in visible text too" rule applies here as well.

The Client-Side Rendering Trap

There's a second, bigger version of the "schema is invisible" problem, and it applies to your entire page, not just your markup.

GPTBot, ClaudeBot, and PerplexityBot do not execute JavaScript as their default behavior. They fetch the raw HTML response and extract whatever text is already there. If your pricing, your product descriptions, your docs, or your blog body only render after client-side JavaScript runs, the content these crawlers receive is often an empty shell, regardless of whether you also added perfect schema. Googlebot is the exception: it runs a headless Chrome rendering pipeline and sees the fully rendered page, which is also the index that Google's AI Overviews and AI Mode draw from.

One honest caveat worth flagging rather than hiding: server-log analysis published in late September 2026 by one site owner found OpenAI's crawlers (GPTBot and OAI-SearchBot) beginning to issue JavaScript-framework prefetch requests against their site starting around September 25, consistent with some form of rendering, after months of pure raw-HTML behavior (DEV Community, September 2026). That's a single site's server logs, not a confirmed OpenAI policy change, and the behavior looked selective rather than universal. Treat it as an early signal to watch, not a reason to assume any AI crawler reliably renders JavaScript today. Until a vendor documents that behavior directly, the safe assumption stays the same: if a fact only appears after JavaScript runs, assume the assistant never saw it.

What to Do Instead

The fix for both problems, invisible schema and invisible client-side content, is the same instinct: make the important fact exist in plain visible text in the initial HTML response, every time, without exception.

Want to track whether any of this is working? Our ranked AEO tools guide and the broader AI visibility tools comparison cover what's worth paying for. Best GEO Tools in 2026 sorts largely the same vendors by whether they only report or also act, and how to measure AI visibility covers the free methods first. Already fixed the content and crawlability basics and weighing paid placement instead? Noble vs Anchorial and should you pay for AI citations cover that different purchase. For the terminology question underneath all of this, see GEO vs AEO vs SEO. The other popular tactic with thin evidence has its own write-up in llms.txt Explained, and How to Get Cited in AI Answers covers what does have documented support. And for tooling that helps produce clear, visible-text-first pages, our AI tools for SEO content roundup covers that layer.