Ahrefs's 1,885-Page Schema Study Killed the AEO Industry's Main Sales Pitch
Ahrefs added JSON-LD schema to 1,885 pages between August 2025 and March 2026 and compared the result to 4,000 control pages. AI Overview citations dropped 4.6%. AI Mode rose 2.4%. ChatGPT rose 2.2%. All three are statistically indistinguishable from zero, which directly contradicts the 2.5x-to-2.8x citation lift the AEO software industry has been selling since GPT-5 shipped.
What Ahrefs actually tested, and why the design matters
The test wasn't a casual audit. Ahrefs selected 1,885 pages that added JSON-LD between August 2025 and March 2026, then compared their AI citation rates against 4,000 control pages that didn't add schema in the same window. Every page in the study had at least 100 AI Overview citations in February 2025, so this is data on pages already in the citation pool. Not pages trying to break in.
That matters because most "schema lifted our citations 2.5x" case studies you'll see on LinkedIn this week come from sites that were nearly invisible to AI engines beforehand. They added schema and a dozen other things, and the citation count went from 4 to 10. The lift came from being small enough to grow, not from the JSON-LD. Ahrefs's control group removes that confound.
The numbers, from Ahrefs's own write-up:
- Google AI Overviews: -4.6% (small decline relative to controls)
- Google AI Mode: +2.4%
- ChatGPT: +2.2%
The researchers stated it plainly: "Adding schema produced no major uplift in citations on any platform." For pages already in the citation rotation, JSON-LD is, at best, statistical noise.
Why the AEO software industry is selling a number that doesn't exist
Walk through almost any AEO tool's pricing page right now and you'll see some version of "pages with schema get cited 2.5x more often" or "comprehensive schema implementations drive 2.8x higher AI citation rates." Those numbers come from correlation studies on already-cited pages, not from controlled tests of pages adding schema.
The causation runs the wrong way. Authoritative sites that invest in technical SEO also invest in schema. Citation engines aren't picking those pages because of the JSON-LD. They're picking them because the page ranks well, has clean structure, and has been around long enough to accumulate brand signal. The schema is along for the ride.
You can see the same dynamic in earlier Ahrefs research. Their April 2026 study of 4 million AI Overview URLs found that 38% of cited pages also rank in the top 10 for the same query, down from 76% in last year's version of the same study but still the strongest single predictor. Ranking is doing the work. Schema is dressing.
What's actually moving citations in May 2026
If schema is noise, what isn't? Three things are showing up in study after study, and none of them are on-page JSON-LD:
Brand search volume. The strongest single predictor in earlier Ahrefs work turned out to be branded search demand. Not backlinks. Not schema. The frequency with which humans type your brand into a search box is the signal that LLMs lean on hardest when deciding who to cite as authoritative on a topic.
Third-party mentions, not first-party pages. Wikipedia, Reddit, and G2 entries are pulling more citation weight than your own marketing site. We covered this in March: if your brand has thin coverage on those three properties, you're not going to schema your way into more AI citations. You're going to need PR, community presence, and review platform investment.
Title and H1 phrasing that directly answers a question. Ahrefs's earlier ChatGPT-citation study found that pages whose titles match the prompt language verbatim earn more citations than pages with higher domain authority and worse title matches. That's a copy decision, not a markup decision.
None of these are sold as a $99/month AEO tool because they're harder to package. Brand demand takes years. Reddit reputation takes a community manager. Title rewrites are unsexy.
The case for keeping schema anyway
I want to be honest about the other side. Ahrefs's study tested pages that already had 100+ citations. If your site is brand new or has near-zero AI visibility, the test doesn't speak to your situation directly. Schema might still help an engine understand what your page is about when nothing else does.
It's also still useful for the classical rich-results purposes (recipe cards, product pricing, review snippets) which haven't gone away. Those are downstream of Google's classic SERP, not AI Overviews, and they still earn measurable CTR gains.
And if your dev team is going to ship JSON-LD anyway because it's part of the modern CMS template, the cost of leaving it on is roughly zero. The opportunity cost is what matters. If a marketing team spends a Q1 sprint "implementing comprehensive schema" instead of doing the harder work of brand-demand building or community presence, the schema sprint has actively traded down.
The AEO vendor pitch decoder
From what I've seen, here's the rule of thumb for evaluating any vendor selling AI citation tooling this year:
- If the case study compares a site that added schema to itself before adding schema (no control group), the lift number is essentially meaningless. Pages that get attention from a marketing team for any reason tend to gain traffic in the following months.
- If the lift is stated as a multiple (2x, 3x, 5x) without a base rate, ignore it. Going from 4 citations to 10 is a 2.5x lift and statistically nothing.
- If the vendor can't tell you the control-group methodology, they didn't have one. Ahrefs is the rare AEO research source that runs actual controls. Most of the rest are running correlation studies dressed up as causal claims.
The honest pitch for AEO tooling is monitoring, not lift. Knowing where your brand is being cited, which prompts trigger it, and how that's changing week-over-week is a real product. Promising 2.5x citation increases from adding JSON-LD is not.
What I'd do this week with a 3-hour budget
If you have a single afternoon and you want to actually move AI citations, in priority order:
- Audit your Wikipedia, Reddit, and G2 presence. If you don't have a notable Wikipedia entry, a credible Reddit footprint in your space, and at least a few hundred verified G2 reviews, fix that before you touch schema. Pick one of the three and start.
- Rewrite the H1 and title tag on your top 10 commercial pages to match the actual prompt phrasing people use. Not the keyword. The prompt. Type the question into ChatGPT and Perplexity, see how it phrases the answer, mirror that.
- Don't add schema you weren't already adding. Or do, but don't expect anything from it.
That ordering will feel wrong if you've been pitched the AEO playbook. It's the ordering the data supports.
One study, but a hard one to argue with
A single 1,885-page test isn't the final word on schema and AI citations. Replication matters, and other research groups will run their own versions of this over the next six months. Search Engine Journal's coverage of the earlier Ahrefs top-10 study suggests the picture is already shifting fast. But Ahrefs ran a clean before-and-after with controls, on pages that were already in the citation pool, across three of the most-watched engines, and got a result that is roughly zero. That's harder to dismiss than the correlation studies the AEO industry has been leaning on.
I think the cleaner read of where AI citations sit in May 2026 is that they're being assigned by the same signals that drove organic ranking three years ago, with brand demand pulled up the stack. Schema is along for the ride, occasionally helpful, mostly noise. Anyone selling you a "comprehensive JSON-LD audit" as the AEO answer for 2026 is fighting last year's argument with someone else's data.
Notice Me Senpai Editorial
