Adding schema markup barely moved AI citations in a 2026 Ahrefs study

A 2026 Ahrefs study tracked 1,885 web pages that added JSON-LD schema markup between August 2025 and March 2026, then compared each one against thousands of matched pages that never made the change. On ChatGPT, citation counts moved by 2.2 percent. On Google's AI Mode, 2.4 percent. Both differences are small enough to be statistical noise. On Google AI Overviews, citations fell 4.6 percent, a result the researchers call real but can't fully explain. The finding that matters for GEO teams: for pages AI engines already cite, adding schema markup did not move citations, despite years of advice treating it as a top-tier technical lever.

What the Ahrefs study actually tested

Ahrefs' researchers identified 1,885 pages that first added JSON-LD schema sometime in that eight-month window, matched each one against three control pages from different domains with similar citation history that never added schema, then measured citation counts for 30 days before and after the change across Google AI Overviews, AI Mode, and ChatGPT. The primary method was a matched difference-in-differences analysis, run against 4,000 control pages in total and designed to strip out platform-wide trends rather than compare raw averages. Three further robustness checks, including a week-by-week event study, produced the same pattern. Every treated page already had more than 100 AI Overview citations in February 2025, so the results describe pages AI engines already trust, not pages trying to get discovered for the first time. The researchers pooled every schema type together, Article, FAQ, Product, HowTo, and Organization markup alike, and tested only static JSON-LD sitting in the page's HTML, not Microdata, RDFa, or schema injected by JavaScript.

The correlation that looked like causation

A separate Ahrefs analysis of six million URLs found AI-cited pages were almost three times more likely to carry JSON-LD schema than uncited pages, with 53 percent of cited pages using it. That gap, not the controlled test, is the real source of the advice telling brands schema drives AI citations. Correlation and causation diverge here because the kind of site that bothers to add structured data is usually the kind of site that already has deep topical coverage, consistent publishing, and other traits AI engines reward on their own terms. VizibleAI's own research into topical authority found a 220x citation gap between the narrowest and widest topical coverage, a difference schema markup alone could never plausibly account for. The matched difference-in-differences design exists specifically to strip that kind of confound out, and once it did, the schema effect nearly disappeared.

Schema markup correlates with AI citations because authoritative sites use both, not because adding one earns the other.

Google is already pulling back on schema rewards

Four days before Ahrefs published its report, Google ended FAQ rich results in search, continuing a retreat from structured-data rewards that started in 2023 and picked up speed through 2026. Google first restricted FAQ rich results to authoritative government and health sites back then, limited HowTo results to desktop, and later deprecated HowTo altogether. Course Info, Claim Review, and Estimated Salary markup followed in 2025, and Practice Problem structured data was cut in 2026. VizibleAI covered the FAQ restriction when it first landed in May 2026, arguing it would push AI citation rates higher for structured Q&A content; the pattern since then has only widened. Search Engine Journal's coverage of the Ahrefs study gathered reactions from several well-known SEO figures: Lily Ray flagged roughly 168,000 pages repeating the claim that FAQ schema is critical for GEO, and Joost de Valk compared the deprecation to earlier SEO spam cycles repeating themselves in AI search.

What schema markup still does

Schema markup hasn't become useless. It lost its case as a citation lever while keeping the jobs it was actually built for: rich result eligibility, voice assistant support, and entity recognition inside Google's Knowledge Graph. A separate test from research firm searchVIU, cited in Ahrefs' report, found that none of five AI systems tested used hidden JSON-LD, Microdata, or RDFa during live page retrieval; every one of them read only the visible HTML a human visitor would see. That's a narrower claim than it sounds, since it covers one stage of how these systems work, not whether schema still shapes what gets indexed or how a brand's entity gets recognized elsewhere. Product, Review, Event, and Video markup still earn visible rich results in classic search, and Organization and Person markup still feed the structured entity data that AI engines and Google's Knowledge Graph draw on when disambiguating a brand. For the mechanics of implementing JSON-LD correctly in the first place, VizibleAI's complete schema markup guide covers every type still worth shipping.

What actually predicts AI citations, according to the same research

The strongest predictors of AI citations in 2026 research have little to do with a brand's own markup: topical coverage depth, inline statistics tied to named sources, and mentions earned on third-party sites all outperform anything a company can add to its own HTML. A 2024 Princeton, Georgia Tech, and IIT Delhi study, which VizibleAI has broken down in detail, found that adding specific statistics and inline citations to primary sources lifted AI citation scores by double digits, with named expert quotations adding a further boost. Separately, Ahrefs' own reporting on press release pitches found owned newsroom pages consistently outperform wire distribution, reinforcing that where a claim gets repeated matters more than how it gets marked up. None of this requires ripping out existing schema. It does mean a team debating where to spend the next GEO budget should look past JSON-LD toward the pages, sources, and third-party coverage actually driving the citations it wants.

The practical takeaway for GEO teams

The right response to this data isn't removing schema markup, it's recalibrating what teams expect it to do: treat it as baseline technical hygiene, not a growth lever for AI citations. Zach Ball, CEO of link-building agency Page One Power, made the distinction plainly in his analysis of the Ahrefs findings: structured data helps AI systems read and classify a page, but it doesn't touch selection, the actual decision to cite one page instead of another. Ball's own business sits in earned links and mentions rather than markup, a detail worth weighing alongside his conclusion, but the Ahrefs data behind it stands on its own: 1,885 pages, four separate statistical tests, and the same near-zero result each time.