First data: schema markup doesn't bring more AI citations
An Achtung.app analysis of 1,301 DACH brands across four verticals shows a counter-intuitive pattern: brands that put Schema-Markup (JSON-LD) on their homepage get cited less in AI answers, not more. Pooled across all verticals they get 0.49 fewer citations per brand on average (2.19 with JSON-LD vs. 2.68 without). The difference is measurable, but as the next breakdown shows, misleading.
Average AI citations per brand by JSON-LD type
Aggregated across all verticals. Negative differences dominate; FAQPage and Product trend positive but at small sample sizes.
Average AI citations per brand. Statistical tests and Bonferroni correction discussed in the caveat below.
Once we split by vertical, the effect mostly vanishes. In none of the four verticals (marketing, SaaS, finance, media) does the difference for the most important schema variant (Organization) clear the bar for a defensible claim. What looks like a negative effect from schema is actually a story about industry composition: marketing sites have the highest schema adoption (71%) and the lowest citation rate; SaaS sites get cited most often without using more schema. Averaging across all verticals mixes those effects and produces an apparent negative correlation that disappears once you control for category.
Average AI citations per brand (any JSON-LD), stratified by vertical
Within each vertical the effect shrinks substantially; the gaps narrow toward parity.
Average AI citations per brand. Statistical tests and Bonferroni correction discussed in the caveat below.
By JSON-LD type, no consistent picture emerges. WebSite (−0.67) and BreadcrumbList (−0.60) show the strongest negative pooled differences. FAQPage (+0.38) and Product (+0.94) trend positive but at n=88 and n=58 are too small to interpret. Organization (−0.26) narrowly misses conventional significance.
Even the two faintly positive directional readings, FAQPage and Product, are not an argument for schema. Both tend to live on pages with Q&A blocks or detailed product descriptions — exactly the visible text that search-grounded LLMs read anyway. If there's any effect there at all, it comes from the content, not from the JSON-LD.
The result echoes a parallel Ahrefs study on US data that also found no positive effect, using a difference-in-differences design (a before/after comparison against a matched control group) — −4.6% on AI Overviews, +2.4% on AI Mode, +2.2% on ChatGPT, all within noise. The most plausible mechanism: search-grounded LLMs like ChatGPT, Gemini, and Perplexity fetch pages live and primarily read the visible page text when composing an answer. JSON-LD blocks in the HTML head show no measurable direct effect at that step in the evidence we have today. Structured data remains useful for Knowledge Graph, rich results, and entity disambiguation in traditional search; just not as a lever for AI citations in this first analysis.
Caveat: the data are a snapshot, so correlation, not causation. Three further qualifications matter. First, the hygiene snapshot only covers the homepage; schema on product, FAQ, or study pages is not measured. AI answers also frequently cite third-party sources (Wikipedia, trade media, directories) rather than the brand itself, further weakening the link between homepage hygiene and citations. Second, 28 comparisons were run; under a Bonferroni correction (significance threshold 0.05 ÷ 28 ≈ 0.0018) only the pooled "any JSON-LD" and pooled WebSite effects remain clearly significant. Third, the same null result holds for the normalized citation rate (share of prompts where the brand surfaced): 4.7% with JSON-LD vs. 5.6% without, not just for absolute counts. All brands in the dataset already appear in at least one industry report, meaning they have some baseline AI presence. A causal claim needs data on brands that newly add or remove schema, measured before and after the change. Achtung.app's weekly hygiene snapshots have been running for two weeks; in 8 to 12 weeks a first before/after analysis should be possible.