Ahrefs just published one of the more honest pieces of research to come out of the AI Search space in 2026.
In “We Tracked 1,885 Pages Adding Schema. AI Citations Barely Moved.“, Louise Linehan and Xibeijia Guan tracked 1,885 pages that added JSON-LD between August 2025 and March 2026, matched them against 4,000 control pages, and ran a difference-in-differences analysis to isolate what Schema actually did to AI citation frequency across Google AI Overviews, AI Mode, and ChatGPT.
The result: no meaningful uplift. A +2.4% change in AI Mode and +2.2% in ChatGPT, both statistically indistinguishable from noise. A small but significant −4.6% in AI Overviews that they explicitly say they can’t confidently attribute to Schema itself.
They used four separate tests, all pointing in the same direction. The methodology is solid. Their honesty about the limitations is even more solid.
So let me say clearly: Ahrefs is right. Within the scope of what they tested, the data is credible, and the conclusion holds.
The problem is that the scope of what they tested is much narrower than the conclusion most people will take away from the headline.
What the study actually proves
The headline — “AI citations barely moved” — will travel far. It will end up in LinkedIn carousels, conference slides, and client emails as shorthand for “Schema doesn’t matter for AI Search.” That reading is wrong, and the Ahrefs team essentially says so inside their own article if you read it carefully.
Here’s the specific claim the study successfully tests: adding JSON-LD to pages already heavily cited by AI does not produce a short-term increase in citation frequency.
That’s a real and useful thing to know. It refutes a weak version of the Schema argument that has been circulating in the GEO/AEO space: the idea that structured data is a magic citation lever you can pull for immediate visibility gains. It deserved to be tested, and it didn’t hold up.
But notice what the study does not test.
Every single page in the dataset already had 100+ AI Overview citations before treatment. These pages were already in the consideration set, aka already crawled, already surfaced, already “known” to the AI systems being measured. As Ahrefs themselves acknowledge: “If a page is already getting picked up, our data suggests that adding schema isn’t going to push it higher.“
This is a critical scope constraint. It’s like testing whether adding a label to a bottle already on the supermarket shelf makes customers pick it up more often. The shelf placement already happened. The question of how the bottle got on the shelf in the first place is a different study.
The Retrieval-Time confusion
The Ahrefs article cites a searchVIU experiment showing that during direct retrieval, five major AI systems — ChatGPT, Claude, Perplexity, Gemini, and Google AI Mode — extracted only visible HTML content and completely ignored JSON-LD, hidden Microdata, and hidden RDFa.
This is real. But it describes retrieval-time behavior, which is one specific moment in a multi-stage pipeline. It does not describe what happens during training-data ingestion, during Google’s indexing and entity parsing processes, or during Knowledge Graph construction, all of which happen well upstream of the moment an AI retrieves a page to answer a query.
Conflating these two things — “the AI ignored JSON-LD when fetching this page to answer a question” and “JSON-LD has no effect on how AI systems understand or represent this entity” — is a category error. A meaningful one.
Google’s own documentation is explicit that structured data helps it understand page content and gather information about entities such as people, books, or companies. That understanding happens at index time, not at query time. The searchVIU finding doesn’t contradict this; it just operates at a different layer of the stack.
Schema Sacrifice: infrastructure, not a lever
This is where I want to introduce a distinction that the study — through no fault of its own — was not designed to test.
When I talk about Schema Sacrifice in the context of entity SEO — meaning the deliberate use of structured data not as a rich result tactic but as a mechanism for Knowledge Graph node reinforcement and entity disambiguation, so that an organization or concept becomes machine-resolvable across multiple contexts — I am not describing a citation frequency optimization. I am describing an infrastructure decision.
The analogy that fits better than any other: Schema Sacrifice is less like placing an ad and more like registering a company. You don’t register a company expecting the act of registration to immediately drive revenue. You register it because, without that registration, the company doesn’t exist as a verifiable entity in any official system. The downstream benefits — opening bank accounts, signing contracts, being recognized by third parties — depend on the registration being complete and accurate. But you can’t measure those benefits by asking “did revenue go up in the 30 days after registration?”
The Ahrefs study measured a 30-day citation window on already-cited pages. It cannot speak to whether structured data — particularly Organization, Person, and sameAs properties — helps an entity become machine-resolvable in Google’s Knowledge Graph for the first time, or helps disambiguate an entity that was previously confused with others. These are different functions, operating at a different layer, with a different measurement window.
The Schema type pooling problem
Ahrefs acknowledges this limitation themselves: they pooled Article, FAQ, Product, HowTo, and Organization schema types together. This is methodologically necessary given the study’s scale, but it matters a great deal for interpretation.
Article schema and Organization schema are not variations of the same tool. They serve fundamentally different purposes. Article schema is a content signal — it helps systems understand what a page is and what it contains. Organization schema is an entity identity signal — it helps systems understand who is making a claim and how to connect that entity to its other representations across the web.
Measuring the citation impact of the Organization schema is a bit like measuring whether a company’s VAT registration number makes customers trust their product reviews more. The mechanism is different. The value is real, but it’s not expressed in citation counts over 30 days.
A future study that separates schema types — particularly isolating entity-identity schemas from content schemas — would be far more informative.
What the correlation finding actually tells us
Here is the finding that deserves more attention than it got: at the start of the study, pages cited by AI were almost three times more likely to have JSON-LD than non-cited pages.
Ahrefs immediately and correctly cautions that correlation is not causation. The sites that implement structured data also tend to invest in technical SEO, build better content, earn more links, and do “all the other things that get pages cited.” Schema might just be riding the wave of broader site quality.
That’s a fair methodological concern. But there’s a complementary reading that the study doesn’t fully explore: the same sites that implement structured data tend to be the same sites that have invested in their entity presence — their Knowledge Graph entries, their organization schema, their sameAs connections to authoritative external sources. Those things are genuinely correlated with AI visibility, not because of a direct citation mechanism, but because entity clarity is one of the inputs into whether a system treats a source as trustworthy and well-understood.
The correlation might not be spurious. It might be pointing at a real signal that the study’s 30-day citation measurement window was too narrow, and too focused on already-cited pages, to detect.
The legitimate caveat territory
There are things the Ahrefs team flags in their caveats section that I think deserve more prominence in the main argument.
First: they note that pages adding JSON-LD often change other things at the same time — links, content, technical fixes. Isolating the schema effect from a broader site improvement is genuinely difficult, and they’re right to flag it.
Second: they note the study covered JSON-LD only, and that AI crawlers appear to treat JavaScript-injected schema differently from HTML-embedded schema. This matters. If a significant portion of the treated pages were injecting schema via JS in ways that Google’s crawler hadn’t fully processed within the 30-day window, the measurement window might be capturing the delay rather than the absence of an effect.
Third: they explicitly acknowledge that for pages not yet cited by AI at all, schema might still play a role in getting crawled, parsed, or indexed in the first place. They just can’t test that with this dataset.
These aren’t minor footnotes. They’re the boundary conditions of the entire study.
So what should you actually do?
The practical takeaway is not “stop implementing Schema,” and it’s also not “add schema and watch your AI citations climb.” It’s more precise than either.
For pages already performing well in AI search, Schema is not a short-term citation amplifier. The Ahrefs data supports this clearly, and you should not be selling Schema to clients as an immediate AI visibility lever. If that’s the pitch, the pitch is wrong.
For entities that are not yet well-established in Google’s Knowledge Graph — organizations without a Knowledge Panel, brands with disambiguation problems, individuals whose entity record conflates them with others — structured data, implemented correctly and with genuine attention to entity identity signals (Organization, Person, sameAs, about, mentions), remains one of the most important investments you can make. Not because it will move your citation count in 30 days, but because entity disambiguation is the foundation on which everything else is built. AI systems that can’t resolve who you are with confidence are far less likely to treat your content as citable, regardless of its quality.
And for everyone: the searchVIU finding that AI systems ignore JSON-LD during real-time retrieval is important to understand, but it is not the argument against Schema. It’s the argument against expecting Schema to do work at query time that it was never designed to do. Its work happens earlier in the pipeline, at the indexing and entity resolution stage, where the representation of your entity in machine-readable systems is either clear or it isn’t.
A note on the research itself
Ahrefs continues to produce the kind of empirical work that this industry needs more of. The “Xarumei” experiment that Mateusz Makosiewicz ran — testing whether AI systems could be fooled by detailed fabricated brand narratives — directly inspired my own thinking on how AI search handles entity verification and what Google’s Knowledge Graph is actually defending against.
Good research that tests a narrow question and is honest about that narrowness is more valuable than broad claims with thin data. This study qualifies. The appropriate response is not to dismiss Schema, but to use the study to dismiss a specific, weak version of the Schema argument, and to hold the stronger version to a higher standard of evidence than we currently have.
The stronger version may turn out to be right. The 30-day citation window on already-cited pages was not the test that would tell us.
