This one goes round in circles every time it comes up, and I think the reason is that two different questions are being argued as if they were one.
Here is the contradiction. Both Microsoft and Google have gone on the record saying structured data helps their LLMs. And two separate pieces of testing found no relationship at all between schema and whether an AI cites a page. Most of what gets written on this picks a side and ignores the other half of the evidence. As far as I can tell both halves hold up.
The on-record statements first, since the sceptics tend to skip past them. Fabrice Canel, principal product manager at Bing, said at SMX Munich in March last year that schema markup helps Microsoft's LLMs understand content for Copilot. That was on stage, not a hedged blog post, and when it got written up he turned up in the comments and did not walk it back. About a week later Ryan Levering, who works on structured data at Google, said at Search Central Live in New York that a lot of Google's systems run considerably better with structured data, and specifically framed it around grounding their generative systems. Two companies that operate their own index, both on the record, within days of each other. OpenAI, Anthropic and Perplexity have said nothing publicly either way.
Now the other side. Search Atlas compared schema coverage against citation rates across OpenAI, Gemini and Perplexity and found no correlation, with heavily marked up domains performing no better than domains with almost nothing. And Mark Williams-Cook ran what I think is the cleaner test back in February. He placed a made up address inside deliberately invalid, invented JSON-LD, with no equivalent anywhere in the visible copy of the page. Both ChatGPT and Perplexity returned the address.
That last result is the one worth sitting with. If the model were parsing markup as structured data, invalid markup should have failed. It did not fail. Which means the script block is being read as text, the same as any other string on the page.
So both are true, and the reason is that they are describing different layers.
The first layer is inference time, the model reading a page during retrieval. At that point your JSON-LD is text. Tokenised, sometimes extracted from, but not understood as a graph. And as text it is arguably worse than a plain sentence, because the tokens are going on syntax rather than meaning. That is the layer the testing covers, and the testing is right about it.
The second layer happens much earlier, when Bing and Google's crawlers parse your markup into their knowledge graphs. Those graphs are what ground the answers. That is the layer Canel and Levering are describing, and they are right about it too.
Worth noting that Google's own documentation on AI features eligibility says no special schema is required to appear in AI Overviews, which looks like it contradicts Levering until you read it carefully. Not required, still helps the systems feeding it. Not the gate, the differentiator.
Where I do lose patience with the pro-schema argument is the standard comeback, which is that those studies measured citations rather than entity understanding. That may well be correct but as stated it is unfalsifiable, and if the benefit only exists somewhere nobody can observe it, it becomes difficult to tell a client what they have bought. The honest version is that the effect is real but upstream and indirect, so it needs measuring upstream. Bing Webmaster Tools has an AI performance report. Knowledge panel presence is checkable. Whether your entity resolves correctly is checkable. All of that is observable. I notice that nobody selling schema as an AI visibility service ever points you at any of it, they point you at citation counts, which is the one place the effect legitimately will not show up.
One stat I would like to see retired while we are here. SE Ranking found roughly 71 percent of pages cited by ChatGPT and 65 percent of pages cited by AI Mode contain structured data. It gets quoted constantly as proof. The confounding is enormous. Pages with schema are pages built by teams who also do server side rendering, sane internal linking, accurate dates and decent information architecture. That is a measurement of competence, not of markup.
Last thing, on the entity modelling work that has been getting attention lately. The underlying idea is sound, treating a site as a public declaration of your entities and the relationships between them rather than as pages that happen to carry markup. Building an ideal entity model for your sector and then auditing what you actually cover against it strikes me as a genuinely useful exercise and an underrated one. But I read a piece recently that described filling Schema.org's gaps by inventing custom entity types, and that part needs a caveat. Google and Bing parse known vocabulary and discard the rest. As an internal audit checklist, inventing entities is fine and useful. As markup you actually ship, it does nothing, and the piece did not draw that distinction.
Practical position I have landed on. Organization and Product with a stable @id and genuine sameAs references, since that is the entity resolution layer and it is the part with first party confirmation behind it. Skip the FAQ padding, that was a rich results play and the rich result has largely gone. And never let a fact exist only in markup. If it matters, it belongs in the visible copy as a sentence, with the markup as a second copy for the index rather than the primary one.
Caveats, because I would want them. Canel and Levering are solid, both first party and both reported by people who were in the room. The Williams-Cook test is small but the design is clean and I cannot come up with another explanation for the result. The Search Atlas work I have only seen summarised, I have not read the methodology. And a fair amount of the commentary surrounding all of this comes from companies selling AI visibility tooling, including some of the outlets publishing it.
The second layer is the part I am least confident about, because it has the least public evidence and the most vested interest attached. So the question I would actually like answered: has anyone here seen movement in the Bing Webmaster AI performance report after deploying or fixing Organization schema? That seems like the one place where this is testable with data most of us already have.
