Does Schema Causally Lift AI Citations? No.

The retraction nobody repeated

In February 2026 a study reported that schema markup was negatively associated with AI citation. Odds ratio 0.546, p below .001. Pages carrying schema got cited less often. It was a fantastic headline. It was also wrong, and the person who found it is the one who said so.

Kurt Fischman re-ran the model with query-clustered standard errors and organic rank position in the equation. The negative effect evaporated into nothing. He published the reason himself.

This finding proved to be a methodological artifact: Google's ranking algorithm systematically enriches top-10 organic results for schema-bearing pages, inflating schema prevalence in the non-cited control population.

Read that sentence twice, because it is the whole argument in one line. The pages that rank are the pages that have schema. Any study that measures schema without measuring rank is measuring rank.

The argument is always run as all or nothing, which is why it never resolves. One camp says schema is dead because Ahrefs found nothing. The other says schema is more important than ever because Bing said it helps. Both camps are arguing about a claim neither has bothered to state precisely.

So let me state it. There are four separate claims hiding inside "schema helps AI":

1. Structured data helps an engine parse a page and resolve which entity it is about.
2. Structured data makes a page eligible for a specific enhanced search feature.
3. Structured data feeds product, merchant and feed-driven surfaces.
4. Structured data increases the probability that a generative engine selects your page as a source for its answer.

Claims one, two and three are documented by the engines themselves. Claim four is the one being sold, and it is the only one with no documentation and no supporting experiment. Separating them is the entire job of this post.

What the engines say in their own words

Start with the party that has the most to gain from you adding markup, because Google's own language is quietly devastating to claim four.

Google's AI optimization guide, last updated 10 July 2026, says it plainly: "Structured data isn't required for generative AI search, and there's no special schema.org markup you need to add." The AI features documentation repeats it: "There's also no special schema.org structured data that you need to add."

Google does not stop at neutrality. It tells you why to keep doing it anyway: structured data "helps with being eligible for rich results on Google Search." That is Google endorsing claim two while declining to endorse claim four, in the same paragraph, on the same page.

The single technical prerequisite Google does state for AI features is one almost nobody sells: a page "must be indexed and eligible to be shown in Google Search with a snippet." A stray nosnippet or data-nosnippet zeroes out AI eligibility in a way no amount of JSON-LD can rescue. If you are buying anything mechanical, buy a crawler and snippet access audit before you buy markup.

Now the strongest counter-argument, because it deserves a fair hearing. At SMX Munich in March 2025, Microsoft's Fabrice Canel said something the schema industry has been quoting ever since. Be precise about what the record actually is, because almost nobody repeating it has checked. There is no transcript and no verbatim Canel sentence in circulation. What exists is David Mihm's write-up of the talk, which Barry Schwartz then reported at Search Engine Land: Canel confirmed that schema markup helps Microsoft's LLMs understand your content.

Take it seriously anyway. A Microsoft product manager said it on stage and the trade press stood it up. But look at the verb, because the verb carries the entire load. Understand. Not select, not cite, not rank. It is a claim about parsing, which is claim one. Microsoft's own AI Performance announcement in February 2026, written by four of its AI product managers, gives formatting guidance for AI citation and never once mentions structured data. What it recommends instead is this: "Clear headings, tables, and FAQ sections help surface key information and make content easier for AI systems to reference accurately."

The asymmetry is the finding. When engines describe schema, they describe comprehension. When they describe citation, they describe content. Nobody in the industry seems to have noticed that the two lists do not overlap.

Three studies, three designs, one result

There are now three reasonably serious attempts to measure this, and their value comes from how differently they are built. A single design can be fooled. Three designs failing to find the same effect is harder to dismiss.

The Ahrefs design is the strongest of the three because it has a treatment and a control and a before and an after. The AI Mode and ChatGPT numbers sit close enough to zero to be noise. The AI Overviews decline is small, statistically significant, and unexplained, which Ahrefs says out loud rather than spinning into a story.

Adding schema produced no major uplift in citations on any platform.

SALT.agency comes at it from the opposite direction. Instead of watching pages add schema, Dan Taylor's team looked at 107,352 sites that had already won an AI Mode citation and asked what markup they carried. If exotic schema types were the edge, the winners would be full of them. They were not.

Schema is a hygiene factor (at best) for AI Mode visibility, not a differentiator.

Fischman's contribution is the one the industry should be reading most closely and is reading least. He is the only one of the three who explicitly modelled organic rank alongside schema, and the result is stark: schema presence is statistically nothing, while every single position of organic rank changes citation odds by roughly 24%. Position one pages were cited in 43% of the queries they appeared in. By position seven that fell to 5%.

One honest caveat on all three. Ahrefs studied pages that already had 100 or more AI Overview citations before treatment, so its data cannot speak to a page with zero existing AI visibility. Fischman's work is a self-published preprint by a practitioner, not peer-reviewed literature. SALT's design is observational and cannot establish direction. None of these is a randomised controlled trial. What they are is three independent teams, using three incompatible methods, all failing to find the thing the industry sells.

The three layers schema is documented to occupy, and the fourth it is sold for

Here is the version of this argument I wish someone had drawn for me two years ago.

Use this graphic on your site

Free to republish with a link back to this page. Copy the embed code:

<a href="https://josephtimpson.com/insights/does-schema-help-ai-citations"><img src="https://josephtimpson.com/assets/infographics/does-schema-help-ai-citations.svg" alt="A four layer stack showing three documented functions of structured data and a detached fourth layer labelled AI citation selection with no documented mechanism" width="1200" style="max-width:100%;height:auto"></a><p>Graphic by <a href="https://josephtimpson.com/insights/does-schema-help-ai-citations">Joseph Timpson</a></p>

Layer one is parsing and entity disambiguation. Google's introduction to structured data says it uses markup "to understand the content of the page, as well as to gather information about the web and the world in general." That second clause is the Knowledge Graph, and it is the most underrated sentence in the documentation. This layer is real, and it is the reason entity clarity work still belongs in a technical scope.

Layer two is rich result eligibility. Google's structured data policies draw the line with unusual precision: "Using structured data enables a feature to be present, it does not guarantee that it will be present." The feature gallery currently lists 31 supported types. Schema.org's core vocabulary contains 823. The other 792 types produce nothing in Google Search, which should tell you how narrow this layer actually is. FAQ and HowTo used to be on that list and are not any more.

Layer three is merchant and feed surfaces. Product markup makes a page "eligible for display in merchant listing experiences on Google Search, including the shopping knowledge panel, Google Images, popular product results, and product snippets," per Google's merchant listing documentation. For ecommerce this is not a nice-to-have, it is the plumbing.

Layer four is AI citation selection. It has no documentation, no stated mechanism, and no supporting experiment. It is drawn detached in the graphic above because that gap is the argument.

There is one more piece of evidence for the appearance-only reading, and it is the sharpest one available. Google says a structured data manual action "means that a page loses eligibility for appearance as a rich result; it doesn't affect how the page ranks in Google web search." Even the penalty for abusing schema is confined to appearance. If markup were a retrieval input, the punishment would touch retrieval. It does not.

The confound, stated cleanly

Every study that finds schema helps is measuring the same three things and calling them one.

The sites that implement structured data are not a random sample of the web. Look at where markup actually clusters, the way the 2025 Web Almanac's SEO chapter tracks it: large publishers, ecommerce platforms, and CMS installs where the SEO plugin emits Organization and Article markup automatically on day one. Those are the same properties that dominate AI citations.

So when a vendor shows you the share of ChatGPT-cited pages carrying structured data, the number may well be accurate and it still proves nothing. Those pages are on managed CMS platforms with editorial teams, established entity footprints, and link authority accumulated over years. The markup and the citation share a cause. Neither caused the other.

Daniel Cheung screened the literature on this question and reached the cleanest formulation of it I have read.

Any schema study that doesn't control for rank is mostly measuring rank.

His review of the schema and citation evidence concluded there is "no credible evidence that schema markup earns AI citations," and he identified the pattern that makes this whole literature legible: the studies that found an effect are the ones missing controls, and the studies with controls found nothing or slightly negative. That split is not a coincidence. It is a diagnostic.

I should be straight about my own position here, because I have been on the wrong side of it. I have built and sold a productized schema service. I still sell it, and I still ship JSON-LD on client sites in volume. What changed is the sentence I am willing to put next to it on a proposal. It used to be about AI visibility. It is now about rich result eligibility and entity disambiguation, because those are the two claims I can defend with a primary source.

Where the evidence is not null

The intellectually honest version of this post has to include the place the data does not support me, so here it is.

Fischman found one real split. Pages carrying Product or Review markup with concrete attribute values populated, actual prices, actual aggregate ratings, actual specifications, were cited at 61.7% versus 41.6% for pages carrying generic wrappers like empty Article, Organization or BreadcrumbList. That difference was significant at p = .012, and he reported it as most useful to lower-authority domains.

This is worth taking seriously, and it is worth reading carefully, because it does not say what schema vendors will claim it says. The signal is not in the markup. It is in the facts inside the markup. A page with a price, a rating and a spec table has extractable claims on it. A page with an empty Organization block has decoration.

One trap before anyone rushes to add Review markup on the strength of that number: Google's review snippet documentation, updated 24 July 2026, states that if the entity being reviewed controls the reviews about itself, its LocalBusiness or Organization pages are ineligible for the star review feature. The homepage testimonial widget wrapped in AggregateRating is the most common violation in local SEO, and it was already not doing what people thought.

Which means the practical instruction the finding generates is not "add more schema types." It is "put concrete, checkable numbers on the page," and you get most of that benefit from visible on-page content whether or not you wrap it in JSON-LD. That maps directly onto what the original GEO research found in 2023, where adding statistics, quotations and cited sources were the three top-performing interventions, and keyword stuffing scored below doing nothing.

Also worth saying: nobody has good measurement here anyway. AI answers are non-deterministic enough that a single before-and-after screenshot is worthless as evidence, which is why sample size discipline matters more in this discipline than in classic SEO.

What the schema budget buys instead

If you have a fixed number of technical hours and you were going to spend them on markup for AI reasons, here is where the published evidence actually points. Ranked by strength of evidence, not by how good it sounds in a pitch.

Four things to do with the hours.

First, fix access. Verify that AI crawlers reach your pages and that nothing is suppressing snippet eligibility. This is boring, mechanical, cheap, and it is the only item on this list that Google explicitly states is a requirement.

Second, chase subquery rank rather than head-term rank. Ahrefs' analysis of 1.4 million ChatGPT prompts found cited page titles matched the engine's internally generated fan-out queries more closely (cosine 0.656) than they matched the user's original prompt (0.602). Surfer's fan-out research found pages ranking for both a main query and its fan-outs were 161% more likely to be cited. That is a query fan-out coverage problem, not a markup problem.

Third, earn mentions off your own domain. Ahrefs' 75,000-brand correlation study put branded web mentions at Spearman 0.664 against backlinks at 0.218, and Aleyda Solis measured 84% to 93% of AI citation weight sitting on third-party properties across 15 SaaS brands. Correlational, and both authors say so, but it is a stronger evidence base than schema has. See why brand mentions out-correlate backlinks.

Fourth, put real numbers and real named quotes in the content. It is the one intervention that has both benchmark support and the Fischman attribute-richness result pointing the same direction.

Note what is not on this list: llms.txt, which has its own controlled null result and an explicit statement from Google that it does nothing. The schema argument and the llms.txt argument are the same argument wearing different hats, and they belong together in any honest myths audit.

The verdict, and how to answer the client

Keep the schema. Change the sentence next to it.

Structured data is a legitimate deliverable with three documented jobs, one of which (entity disambiguation) is arguably more valuable in an entity-driven retrieval world than it was five years ago. Microsoft has been reported as saying it aids comprehension. None of that is in dispute and none of it needs defending.

What cannot survive contact with the evidence is the causal claim. When a client or a prospect asks whether schema will get them cited in ChatGPT, the accurate answer has three parts, and it takes about twenty seconds:

Aimee Jurenka, writing at Search Engine Land, put the distinction better than the people arguing either extreme.

This doesn't mean schema is useless, it means schema alone doesn't drive citations.

That is the whole position. Schema is not dead and it is not the mechanism. It never claimed to be the mechanism. The industry claimed that on its behalf, then went looking for numbers to justify the claim, and found correlations it forgot to control.

The strongest evidence in the entire literature is a number about rank, published in a study about schema, by an author who went back and corrected his own headline. Start there. Everything else on the insights index is downstream of getting that ordering right, and if you want the ordering applied to a specific site, that is what the working session is for.