I need to say something that’s going to annoy a lot of agencies selling “AI schema audits” right now. Adding JSON-LD to your site is not going to get you cited in ChatGPT, Gemini, or Perplexity. That’s not my opinion. It’s what Google itself said in May 2026, and it’s what the one large scale test on the subject found too.
Here’s the context. On May 15, 2026, Google published its first official guide on optimizing for generative AI features, written with input from John Mueller on the Search Relations team. The message was blunt for a company that usually hedges everything: AI Overviews and AI Mode run on the same crawl, the same index, and the same quality systems as regular Google Search. There’s no separate AI index. No special schema markup that unlocks a citation slot. If Googlebot can’t crawl and understand a page normally, that page isn’t showing up in an AI answer either, no matter how clean the structured data is.
Then there’s the study. Ahrefs ran a controlled test across 1,885 pages, adding schema markup to some and leaving others untouched, then tracking AI Overview citations over time. The pages with new schema didn’t gain citations. They actually lost about 4.6 percent. That’s a small number, and I wouldn’t read too much into the direction, but it’s enough to say plainly: schema markup validation and clean JSON-LD are not a lever that pulls citations toward you. When a Redditor asked Mueller directly whether schema helps with LLMs, his answer was honest to the point of being unsatisfying: yes, no, and it depends, mostly useful for data driven features like Shopping listings, mostly just enrichment everywhere else.
So why am I still writing about schema markup for a client that wants AI visibility? Because “it doesn’t cause citations” is a different statement than “it doesn’t matter.” I’ve gone through a fair number of these audits over the years, and the businesses that skip structured data entirely tend to have a second, bigger problem hiding underneath it: their site is genuinely hard for any system, human or machine, to understand at a glance.
Structured data markup, and JSON-LD specifically since it’s the schema markup language Google recommends over the older microdata format, still does real work. It tells a system with certainty that a business name is a business name, that a price is a price, that a service area covers a specific set of cities instead of forcing the crawler to guess from surrounding text. Bing’s AI features use it. Google pulls it during the retrieval step that feeds its AI Overviews, even if it’s not the deciding factor in whether you get quoted. And for anything with a price, availability status, or booking date, like product schema on an ecommerce catalog, it’s closer to mandatory than optional, because that’s exactly the “data driven feature” case Mueller pointed to.
There’s also a separate, more basic issue that has nothing to do with whether schema markup causes citations, and it trips up more sites than people realize. AI companies run two different kinds of bots. GPTBot and ClaudeBot exist to collect training data. OAI-SearchBot and Claude-SearchBot are the ones doing live retrieval for search answers. A lot of businesses, trying to opt their content out of AI training, block anything with “GPT” or “Claude” in the user agent string inside robots.txt, and end up blocking the search bot right along with the training bot. At that point the schema markup seo question is irrelevant. The bot never gets to the page.
One more practical snag worth mentioning: schema added through Google Tag Manager loads via JavaScript, and not every LLM crawler renders JavaScript the way a browser does. If your JSON-LD only appears after a script fires, some crawlers see a page with nothing on it. Server side rendering avoids that problem entirely.
A quick word on FAQ schema markup, since I get asked about it constantly. Google scaled back the visual FAQ dropdown in regular search results earlier this year, and that led a lot of people to assume FAQ schema markup stopped being useful altogether. It didn’t. Retrieval bots still read it during indexing, whether or not Google chooses to display the dropdown, and it remains one of the cleanest ways to hand an AI system a direct question paired with a direct answer instead of making it dig through paragraphs. I’d still put it below Organization and Product schema on the priority list, but I wouldn’t rip it out either.
Validation matters more than most people give it credit for. A json ld schema markup block with even one property typo can fail silently, meaning it sits on the page looking fine to a human while returning nothing useful to any system reading the code. Run new markup through the Rich Results Test and the Schema.org validator before publishing, not after a client asks why nothing changed. And check it again anytime a developer touches your templates, because that’s when structured data breaks most often without anyone noticing for weeks.
None of this is glamorous work, and it won’t produce the overnight AI visibility a lot of pitches promise. It’s closer to plumbing. Fix the robots.txt, render the markup server side, keep Organization and Product schema accurate, run it through the Rich Results Test after every redesign, and skip anything sold as an “AI schema” shortcut that Google itself doesn’t recognize. That’s the same groundwork the technical SEO team at Adams Internet Marketing walks through with clients before making any bigger promises about AI search.
Key Takeaways
- Google’s own May 2026 guidance states there’s no separate AI index and no special schema markup that guarantees an AI citation.
- An Ahrefs test across 1,885 pages found no citation gain from adding schema, and a slight drop, which matches Mueller’s “it depends” answer on whether schema helps LLMs.
- Schema markup still matters for data driven features like Shopping and product listings, and for helping any system, human or machine, understand a page with certainty.
- Blocking GPTBot or ClaudeBot in robots.txt to opt out of AI training can accidentally block the separate search bots (OAI-SearchBot, Claude-SearchBot) that power AI citations.
- JSON-LD loaded through Google Tag Manager may be invisible to crawlers that don’t render JavaScript, so server side rendering is safer.
Highlights
- Google has said directly that AI Overviews and AI Mode use the same crawl and index as standard search, not a separate AI specific one.
- The one large scale causal test on schema and AI citations found no measurable benefit.
- FAQ schema markup and product schema still carry value for data extraction, even where visual rich results are limited or deprecated.
- The more common cause of AI invisibility isn’t missing schema, it’s a robots.txt file quietly blocking the wrong bot.