Does schema markup help AI citations? Matched tests find no lift

The claim that adding schema markup makes AI systems cite a page is Overstated. Across 1,885 pages that added JSON-LD, AI citations moved +2.4% on Google AI Mode and +2.2% on ChatGPT, both statistically indistinguishable from zero Ahrefs · 2026B. The apparent benefit in large datasets disappears once rank or domain is held constant. AI systems read the text a visitor can see, and compliant markup only restates that text. The one category that beats the baseline is Product and Review markup carrying prices, ratings and specifications, where the attributes appear to do the work.

Adding schema markup does not raise AI citations in matched tests

The only causal test of the claim found no lift. Ahrefs tracked 1,885 URLs that moved from no JSON-LD to JSON-LD between August 2025 and March 2026. Each was matched to three control pages from other domains with similar prior citation levels Ahrefs · 2026B. Google AI Overviews showed a small 4.6% decline over the same 30-day windows. Because the pages were matched on existing citation levels, the result describes pages AI systems already cite, not pages none has yet retrieved.

A rank-controlled study reached the same answer by another route. Across 730 ChatGPT and Gemini citations over 75 commercial queries, schema presence stopped predicting citation once Google rank was controlled. The corrected odds ratio was 0.678 (p=.296). Rank did the work instead: position-1 pages were cited 43% of the time, falling to 5% at position 7 Fischman · Growth Marshal · 2026B. The study is a single-author preprint with a modest sample, and its direction matches the matched test.

Schema correlates with AI citations because strong domains use it

The correlation behind the claim is real and does not survive a domain control. In the Ahrefs dataset, 53% of AI-cited pages already carried schema, which the authors read as a marker of overall site quality Ahrefs · 2026B. Across about two million citations from ChatGPT, Claude, Google AI and Gemini, structured data and FAQ blocks looked positive in pooled data Moore · Discovered Labs · 2026B. Under domain fixed effects, which compare pages within the same site, those effects reversed or collapsed to zero.

The match between page and prompt predicted citation instead. Prompt-content alignment was the strongest page-level predictor (beta +0.37), and domain-level AI authority outweighed page features roughly sixfold Moore · Discovered Labs · 2026B. The study is an AEO agency's preprint, not peer-reviewed, and covers B2B SaaS pages only.

The strongest evidence on the other side is associational. An audit of 1,702 citations across 1,100 URLs found structured data among the three on-page properties most strongly tied to citation arXiv · 2025A. Its models used domain-clustered standard errors rather than domain fixed effects, so they cannot separate markup from the sites that deploy it.

AI systems read visible text, and Google forbids markup for hidden content

The mechanism explains the null result. A controlled experiment placed product prices only in hidden JSON-LD, Microdata and RDFa. It then asked five AI systems to read the page. None of ChatGPT, Claude, Perplexity, Gemini or Google AI Mode extracted the hidden values, and all five read visible HTML only searchVIU · 2025B. The result is cross-verified through two independent reports because the primary page resisted direct fetch.

Google's rules close the gap from the other side. Structured data must be a true representation of the page, and marking up content readers cannot see can trigger a manual action Google · 2026A. A fact held only in markup is therefore unreadable to these systems and a policy violation at the same time.

Visible does not mean above the fold. Google's October 2026 revision of its helpful-content guide counts tabbed or expanded sections, such as product specifications, as main content Google · 2026A. An accordion a visitor can open is page text; a JSON-LD block is not. Google also states that site owners need no extra structured data to appear in its generative AI features Google · 2026A. That is the platform's own position rather than an independent measurement.

Bing and Copilot consume schema, which is not the same as citing it

Structured data still reaches some models. Microsoft's Fabrice Canel confirmed at SMX Munich that Bing and Copilot use schema markup to help their language models understand content Schwartz · Search Engine Land (reporting Microsoft/Fabrice Canel) · 2025B. That is a conference statement reported by trade press, not a study, and no other major platform had made an equivalent confirmation at the time.

Inside retrieval pipelines the measured gain from markup is small. In a test harness across 349 queries, JSON-LD moved plain HTML from 3.62 to 3.89 on a 1-to-5 accuracy scale. An enhanced entity page reached 4.69 Volpini et al. · WordLift · 2026A. The authors sell the entity-page format, and the harness scored answer accuracy, not live citation. Consumption and citation lift are separate questions, and only the first has a clear yes.

Product and Review schema is the exception, because the attributes do the work

One schema category beat the baseline in the rank-controlled study. Product and Review markup carrying pricing, ratings and specifications was cited 61.7% of the time, against 41.6% for generic types Fischman · Growth Marshal · 2026B. Those pages also state a price and a rating in visible text. The markup type is therefore a likely proxy for the facts.

Controlled retrieval work points the same way. Across 252,000 paired trials in six models, explicit price information and a recent timestamp helped consistently, while formatting-only edits had little impact Vishwakarma et al. · Sprinklr · 2026A. The authors work for a vendor and used a synthetic corpus, so the result awaits independent replication.

Practitioner tests rarely isolate markup. In one vendor case study, a team added a Key Takeaways block with schema to a single page. Its citation rate rose from 1.45 to 2.33, against 1.61 to 1.72 for one control page Debski · Peec AI · 2026C. Schema and new copy changed together, so neither can take the credit. It is one page pair, Tier C; treat as provisional.

Google retired the FAQ schema rich result in 2026, but most sites lost it in 2023

The dated facts are narrow. FAQ rich results stopped appearing in Google Search on 7 May 2026, and Google removed the FAQPage documentation on 15 June 2026 Google · 2026A. Search Console API support for the feature ended in August 2026, while FAQPage remains valid Schema.org vocabulary Google · 2026A.

Most commercial sites had lost the feature years earlier. Google restricted FAQ rich result eligibility to authoritative government and health sites in August 2023 Google · 2026A. A retailer or software company therefore had no route to that snippet for nearly three years before the retirement.

FAQPage also had a vocabulary problem: it describes a page whose purpose is question and answer. Declaring it on a product page that merely carries an FAQ section misdescribes that page. Google requires structured data to be a true representation of the page content Google · 2026A. The retirement removed a display feature. Whether FAQ content earns citations is a separate question, decided by the evidence inside each answer, as the library's article on evidence-dense FAQs sets out.