Ahrefs Test Finds Schema Markup Fails to Boost AI Citations

For years, the prevailing wisdom in search engine optimization has held that structured data, specifically JSON-LD schema markup, is a valuable tool for improving how content appears in AI-generated search results. A new, large-scale study from Ahrefs, however, introduces a significant complication to that narrative. The company’s latest report, which tracked nearly 1,900 web pages that added JSON-LD schema, found no clear evidence that implementing this markup leads to an increase in citations from major AI platforms, including Google AI Overviews, Google AI Mode, and ChatGPT. This finding challenges a widely held assumption and demands a closer look at what schema markup actually does for AI visibility. The study analyzed a massive dataset of 6 million URLs and revealed a compelling surface-level correlation: pages that are cited by AI are roughly three times more likely to include JSON-LD schema. This kind of statistical gap has often been interpreted as proof that structured data helps AI systems understand and reference content. Ahrefs, however, suspected a confounding variable. Websites that invest in schema markup tend to be higher quality overall. They typically have better content, stronger internal linking, and more authoritative backlink profiles. In other words, the correlation between schema and AI citations could simply reflect the fact that better sites use schema, not that schema itself causes the citations. To isolate the true effect of schema, Ahrefs designed a controlled difference-in-differences analysis. They identified 1,885 pages that added JSON-LD schema and matched each one with three control pages from different domains. These control pages had similar citation levels and never added schema. By measuring citation changes 30 days before and after the schema was implemented, and using their Brand Radar tool and Agent A to account for platform-specific trends, Ahrefs aimed to isolate the impact of the markup from other quality signals. The results were underwhelming for proponents of schema as an AI booster. Across the three major platforms studied, none showed a meaningful positive lift. For Google AI Overviews, pages that added schema actually saw a relative decline of 4.6 percent compared to the control group. This was a small but statistically notable drop. For Google AI Mode, the change was a paltry plus 2.4 percent, a figure indistinguishable from random variation. ChatGPT showed a similar pattern with a plus 2.2 percent change, also too small to be considered significant. The study further conducted three additional tests alongside the primary comparison, and all four analyses converged on the same conclusion: no clear positive or negative effect could be attributed to the addition of schema markup. The 4.6 percent decline observed in Google AI Overviews deserves careful contextualization. Ahrefs noted that both the treated pages and the control pages were already experiencing a downward trend in citations before the schema was added. The pages that received schema simply declined at a slightly faster rate. In absolute terms, the difference amounts to roughly 12 fewer daily citations per page, in a sample where most pages were receiving hundreds of citations each day. The report itself does not draw a definitive conclusion about this decline. It could reflect a small negative effect from the schema, perhaps because the structured data alters how the page is parsed in a way that is slightly less favorable for AI overview generation. Alternatively, it could be pure coincidence, a statistical artifact of tracking a large number of pages over a short period. Ahrefs explicitly states that the data does not allow for a firm determination either way. This nuance is critical for publishers who might panic over the negative figure. It is a small signal, not a confirmed penalty. The most important limitation of the Ahrefs study is that it only examined pages that were already highly visible to AI. Every page in the dataset had at least 100 AI Overview citations before any schema was added. These pages were already in the consideration set, being crawled, indexed, and surfaced by AI systems. This means the study cannot answer the question that matters most for the vast majority of publishers: does schema help pages that are not yet visible to AI get discovered? Ahrefs readily admits this limitation. It is possible that schema markup aids in the initial crawling, parsing, or indexing phases, helping AI systems understand a page’s content for the first time. However, the data simply does not cover this scenario. A separate experiment cited in the report adds another layer of complexity. The searchVIU experiment tested whether five major AI systems, including ChatGPT, Claude, Perplexity, and Gemini, actually use schema markup when fetching pages in real time. The result was stark: none of them did. They extracted only visible HTML, ignoring JSON-LD, Microdata, and RDFa entirely. It is important to note that this was a direct-fetch test, which is different from how schema might be used during training or indexing. But it does suggest that real-time retrieval, which is how many AI features work, does not rely on structured data at all. These findings have significant implications for content strategy. Schema markup is still valuable for traditional search features like rich snippets, knowledge graphs, and enhanced search result displays. It helps search engines understand the entity relationships on a page, which can improve click-through rates from standard search results. However, as a direct lever for increasing AI citations for pages that are already performing well, the evidence suggests it is not effective. Ahrefs interprets the strong correlation between schema and AI citations as a sign of overall site quality, not a causal link. The sites that use schema are simply better at most things that matter for AI visibility: content depth, authority, user experience, and link building. The report effectively separates the signal of quality from the signal of schema, and finds that schema alone contributes little to nothing. For publishers, the strategic takeaway is clear. If a page is already being cited by AI, adding JSON-LD schema is unlikely to increase those citations. The effort and resources spent on implementing schema for AI visibility might be better directed toward improving content quality, earning authoritative backlinks, or enhancing page experience. For pages that are not yet visible, the picture is less clear. Schema might still play a role in helping AI systems discover and understand new content, but that hypothesis remains unproven. The searchVIU experiment raises a practical concern about real-time retrieval, while the Ahrefs data shows no benefit for established pages. The logical next step for the SEO community is to conduct studies focused specifically on new or low-visibility pages to determine if schema helps them break into AI citation sets. The Ahrefs report is a valuable corrective to the hype around schema as a universal AI visibility tool. It provides rigorous, controlled data that separates correlation from causation, and it does so with transparency about its own limitations. The report does not claim that schema is useless. It does, however, convincingly demonstrate that for pages already visible to AI, adding JSON-LD does not boost citations. This shifts the burden of proof back to those who claim schema is a primary driver of AI discovery, and it encourages a more honest conversation about what structured data can and cannot do in the age of generative search. The future of AI citations will likely depend on factors that have always mattered in search: content quality, topical authority, user engagement, and robust link profiles. Schema markup remains a best practice for structured data representation and rich results, but it should not be viewed as a shortcut to AI visibility. The most effective strategy for publishers is to build genuinely excellent content that AI systems want to cite, and to use schema as a supporting tool, not a primary strategy.