If your AI citation KPI is flat and the team wants to “fix it with schema,” Ahrefs’ newest test is the constraint: adding JSON-LD didn’t create lift across Google AI Overviews, AI Mode, or ChatGPT. Not a little lift. No meaningful lift.
That’s uncomfortable because schema is one of the few “clean” technical levers marketers can ship without rewriting the whole site. Add markup, validate it, move on. Simple.
But the data—at least in this setup—doesn’t reward that instinct.
Primary takeaway (one move): treat schema as hygiene, then run a fan-out query coverage experiment to earn citations from the sub-queries AI systems actually pull from.
The Ahrefs result: correlation showed up, causation didn’t
Ahrefs went after a question a lot of teams hand-wave: do AI systems cite pages more often because those pages have schema, or do schema-heavy sites just tend to be better overall?
In the test summarized by Search Engine Journal’s Matt G. Southern, Ahrefs tracked 1,885 pages that added JSON-LD schema. Each treated page was matched against three control pages from different domains with similar citation levels that never added schema. They measured citation changes 30 days before vs. 30 days after the schema addition, and used a matched difference-in-differences approach (via Ahrefs’ Brand Radar tool and Agent A) to account for broader platform trends.
Here’s the part that matters operationally: no platform showed a meaningful citation increase after schema was added.
The reported deltas were small: Google AI Overviews: −4.6% (a statistically notable decline relative to controls), Google AI Mode: +2.4% (too small to separate from noise), and ChatGPT: +2.2% (also too small to separate from noise). Three additional tests ran alongside the primary comparison and landed in the same place: no clear effect.
So yes, schema is more common on cited pages. But when Ahrefs isolated the variable—“add schema to pages that are already cited”—it didn’t behave like a growth lever.
But the context, however, is where teams usually misread the lesson.
Why this matters now: AI citations aren’t just “rank tracking, but new”
Even before schema enters the chat, Ahrefs’ broader AI Overview citation research has been sending a louder signal: the citation set doesn’t map neatly to the top of the SERP.
In Ahrefs’ early-2026 analysis (as covered by DesignRush), only about 37.9%–38% of cited URLs ranked in the top 10 organic positions for the exact query, while roughly 31.2% landed in positions 11–100 and about 31.0% ranked beyond the top 100 (or didn’t appear in standard results for that exact query). In the organic-only follow-up, the split was similar: Top 10 = 37.1%, 11–100 = 26.2%, beyond top 100 = 36.7%.
Ahrefs’ explanation is the part demand gen leaders should care about: Google uses “fan-out queries,” breaking one prompt into multiple sub-queries and pulling citations from those branches. That means the page that gets cited may be the one that best answers a sub-question, not the one that ranks #1 for the head term.
Translation: schema can be perfectly implemented and still irrelevant to the actual selection problem. The selection problem is coverage and relevance across the fan-out.
What “schema didn’t move citations” does—and doesn’t—prove
This Ahrefs test is useful because it’s constrained. And that’s also its limitation.
Every page in the dataset had 100+ AI Overview citations before schema was added (per the SEJ summary). These pages were already in the consideration set: crawled, indexed, surfaced, cited. If schema helps with discovery or eligibility for pages not yet visible to AI systems, this design won’t show it.
It also pooled schema types together. If some schema types matter more than others, the average effect can wash out. And the 30-day window may miss slower-moving effects. Those are real constraints, not excuses.
There’s another tension worth holding without forcing a fake “winner.” Practitioner studies have reported the opposite pattern. Digital Applied, for example, analyzed 1,000 AI Overviews and reported schema-marked pages were cited 2.3× more often, and adding HowTo schema increased that to 2.8×. WP Riders has also claimed pages with schema were 36% more likely to appear in AI summaries/citations.
Then you have the skeptical line: Search Engine Land (Dec 2024), citing a SearchAtlas study, argued schema coverage showed no correlation with AI citations. Their framing is blunt: relevance and authority win; schema may help extraction/understanding, but it doesn’t reliably drive citations by itself.
So what’s the operator read? This: schema might correlate with citation performance because high-quality sites implement it. Ahrefs’ controlled add-schema test suggests schema isn’t the independent variable teams want it to be—at least not for already-cited pages.
Run it this week: a fan-out coverage experiment (not a schema project)
If you only change one thing, change this: stop treating AI citations like a single-keyword ranking problem. Treat them like a coverage problem across sub-queries.
The hypothesis (make it falsifiable): If we publish and internally link a set of sub-query pages that map to AI fan-out branches for our top 10 revenue-driving topics, then AI Overview citation count and citation diversity (unique cited URLs) will increase over the next 6–8 weeks because we’ll have more “best answer” candidates for the sub-questions AI systems assemble.
Setup: pick 10 topics tied to qualified pipeline (not traffic). For each topic, generate 8–12 sub-queries using the SERP itself (People Also Ask, related searches) plus the questions that show up inside AI Overviews when they appear. Build a simple map: head term → sub-questions → best-fit page target.
Launch: ship 3–5 new pages per topic (start smaller if bandwidth is tight), each written to answer one sub-question cleanly. Give each page a single job. Add internal links from the head-term page and between adjacent sub-questions. Keep schema implementation consistent, but don’t change it mid-test; schema is not the variable.
Readout: track AI Overview presence/citations separately from rank. That separation is the whole point. Ahrefs’ citation distribution data is the warning label: a page can be cited even when it isn’t top 10 for the tracked keyword.
Success = lift in AI citations for the topic cluster (directional, not definitive) plus an increase in the number of distinct URLs cited from your domain. Guardrails = no drop in non-branded organic conversions to the cluster and no meaningful increase in bounce/back-to-SERP signals on the new pages. Stop-loss = if conversions to product-qualified actions fall materially for the cluster over two consecutive weeks, pause publishing and audit intent mismatch.
Trade-off (name it): this will reduce “content density” on your head-term page and may split links and engagement across more URLs. That’s the cost. The upside is more entry points into the citation set.
The kicker: schema is still worth doing—just not for the reason people want
Schema isn’t useless. It’s just not a reliable knob for AI citations in the way teams hope, especially once a page is already being cited.
The cleaner mental model is boring but effective: implement schema because it improves machine readability and supports rich results and entity understanding. Then win citations the hard way—by showing up across the fan-out branches where the real selection happens.
Ahrefs ran the test many teams avoid because it might kill a comfortable belief. It did. And that’s fine. In 2026, the teams that keep shipping are the ones that can drop a pet tactic and still keep the pipeline plan intact.
