If you are adding schema markup to earn AI citations, the best controlled evidence says stop expecting a payoff. Ahrefs tracked 1,885 pages that added JSON-LD against 4,000 matched controls. Citations in Google AI Overviews fell 4.6 percent relative to controls; AI Mode and ChatGPT moved within statistical noise. Don't add schema to chase citations. Don't rip out what you have, either.

The advice everyone repeats

Ask five agencies whether you need schema markup for AI search and you will hear the same answer five times: yes, add it. FAQ schema, Article schema, Organization schema appear near the top of every optimization checklist. The statistic usually quoted behind that advice comes from Ahrefs: across an analysis of 6 million URLs, pages cited by AI were almost three times more likely to contain JSON-LD than pages that weren't.

That number travels well on LinkedIn, and it has a familiar problem. Sites that add schema tend to be the same sites that invest in technical SEO, publish authoritative content, and earn links. The markup may ride along with all of that rather than cause any of it. In May 2026, Ahrefs ran the test that separates the two.

How the experiment worked

The design was straightforward. Ahrefs used its crawler history to identify 1,885 pages that added JSON-LD for the first time between August 2025 and March 2026. Each treated page got three matched control pages from different domains, about 4,000 in total, with similar citation levels and no schema. The team counted citations in Google AI Overviews, Google AI Mode, and ChatGPT for 30 days before and 30 days after each page's change, then compared the groups with a difference-in-differences analysis, which strips out platform-wide trends such as AI Mode growing and AI Overviews contracting during the same period.

One detail matters for reading the results: every treated page was already heavily cited. Each had at least 100 AI Overview citations before the schema went on. This experiment measured what happens to pages already inside the answer engines' consideration set, not whether schema helps a forgotten page get discovered.

What it found

AI source Change vs. matched controls Ahrefs' verdict
Google AI Overviews −4.6% Small but statistically significant decline
Google AI Mode +2.4% Indistinguishable from zero
ChatGPT +2.2% Indistinguishable from zero

Read the table and the fair headline is that adding schema produced no measurable gain on any platform. The −4.6% on AI Overviews deserves a sentence of caution before anyone quotes it as a scare number: it works out to roughly 12 fewer citations per page per day, on pages that were receiving hundreds, in a sample where both treated and control pages were already declining together for reasons that had nothing to do with markup. Ahrefs states plainly that it cannot pin that small gap on schema. Four separate statistical tests told the same story: no clear positive or negative effect.

Why engines treat your schema as plain text

The mechanism explains the null result. In February 2026, Mark Williams-Cook published an experiment built to catch the difference between reading a schema block and parsing it. He created a page for a fictional t-shirt company and placed its address only inside deliberately invalid JSON-LD: a made-up @context URL, a made-up @type, invented properties, no address anywhere in the visible text. When he asked ChatGPT and Perplexity for the company's address, both returned it correctly. Perplexity even volunteered that it had found the answer in "the page's embedded structured data." The models had read the block as ordinary text with unusual punctuation, and never checked whether any of it was valid schema.

A searchVIU experiment, cited in Ahrefs' report, points the same direction from another angle. When ChatGPT, Claude, Perplexity, Gemini, and Google AI Mode fetched a page in real time, all five extracted only the visible HTML. JSON-LD, hidden Microdata, and hidden RDFa were ignored during direct retrieval.

The observational studies point the same way

Controlled experiments are rare in this field, so it is worth checking whether the big observational datasets contradict them. They don't. SE Ranking analyzed 216,524 pages across 129,000 domains and found pages with FAQ schema averaged 3.6 ChatGPT citations versus 4.2 for pages without, a slight negative correlation (observational study). First Page Sage modeled the drivers of AI recommendations across 4,213 commercial prompts and assigned schema and structured data 0.9% of modeled influence in standard AI search, against 80.4% for authority signals (vendor-modeled analysis, not an experiment).

One nuance from the same First Page Sage report is worth keeping: in agentic search, where an AI agent verifies facts before acting, structured data's modeled weight rose to 4.6%. Pricing, availability, and booking facts genuinely help a machine complete a task. That is an argument for structured data as functional plumbing for agents, not as a citation lever.

What this study cannot tell you

The limits, most of which Ahrefs lists itself. Every page studied was already cited heavily, so the result says nothing about pages AI never surfaces; for those, markup might still help crawling, parsing, or indexing, and Ahrefs says so explicitly. All schema types were pooled together, so one type could behave differently from another. The window was 30 days, which would miss slow effects. The study covered JSON-LD in served HTML only, not Microdata, RDFa, or schema injected by JavaScript. And because this was a matched observational study rather than a randomized trial, the small AI Overviews decline cannot be cleanly attributed to schema. That is also why calling JSON-LD a "liability" overstates the data. The defensible claim is narrower: no evidence of a payoff where it is most often sold.

What to do instead

Three actions, in the order the evidence supports.

First, stop budgeting for new schema as an AI visibility play. If a proposal sells FAQ schema as an AI citation lever, it runs against the best controlled evidence available. The timing argument has also collapsed: Google removed FAQ rich results from Search entirely on May 7, 2026, so the old "own more space on the results page" pitch no longer applies.

Second, do not panic-delete. Google's guidance says unused structured data causes no problems, and markup still feeds traditional search infrastructure and Knowledge Graph entity disambiguation. Williams-Cook, no fan of the schema-for-GEO hype, still recommends keeping it for exactly that reason. Deleting working markup on the strength of one negative correlation would repeat the original mistake in reverse.

Third, spend the effort where the evidence points. First Page Sage's modeling puts 80.4% of influence on authority signals, far above schema's 0.9%, and Ahrefs' correlational work keeps finding that cited pages belong to sites that are strong overall. The practical version is getting named: our guide to checking whether ChatGPT recommends your product takes an afternoon and no tools, and why being named by AI beats ranking on Google covers what to do about it.

If you want a first read on your own site, the free instant check runs seven technical checks on your homepage and shows every result on the page.

Run the free instant check

A note on this page's own schema

Yes, this page carries Article and FAQPage JSON-LD. That is not a contradiction of anything above. Our markup targets traditional search infrastructure and entity clarity, and we do not claim it as an AI citation lever. The argument here is against adding markup while expecting citations in return, not against keeping a technically sound site.

Frequently asked questions

Does adding JSON-LD schema increase AI citations?

The strongest available evidence says no. Ahrefs tracked 1,885 pages that added JSON-LD against 4,000 matched control pages: citations in Google AI Overviews fell 4.6% relative to controls, while AI Mode (+2.4%) and ChatGPT (+2.2%) moved within statistical noise. No platform showed a meaningful uplift.

Should I delete the schema markup I already have?

No. Google's guidance says unused structured data does not cause problems for Search, and FAQPage markup can stay in place after the May 2026 rich-result removal. Schema still serves traditional search infrastructure and entity disambiguation. The evidence argues against adding new markup specifically to chase AI citations, not against keeping a technically healthy site.

Why don't AI engines read schema markup the way Google's crawler does?

In a February 2026 experiment, ChatGPT and Perplexity correctly extracted an address that existed only inside deliberately invalid, invented JSON-LD. The models read the block as plain text rather than parsing its structure. A searchVIU test cited in Ahrefs' study found five AI systems extracted only visible HTML when fetching pages in real time.

What actually improves AI citations instead?

The modeled evidence concentrates on authority: First Page Sage attributes 80.4% of influence in AI answers to authority signals, against 0.9% for schema. Ahrefs' correlational finding that cited pages were about three times more likely to have JSON-LD points the same way: the markup rides along with strong sites, it does not make them strong.

This page carries FAQ schema. Isn't that a contradiction?

No. We keep Article and FAQPage markup on this page for traditional search infrastructure and entity clarity, and we do not claim it as an AI citation lever. The argument in this article is against adding new schema while expecting citations in return, not against maintaining valid structured data.

Sources