Does Schema Markup Help AI Visibility? Separating the Two Claims

Ask whether schema markup helps AI visibility and you will get two confident, contradictory answers.

The SEO position: obviously yes, structured data is how machines understand your content, and AI systems are machines. The sceptical position: obviously no, language models read text like a person does — they were trained on prose, not on JSON-LD.

Both are partly right, and they are answering different questions. Untangling them gives a usable answer.

Two mechanisms wearing one name

“AI visibility” spans at least two distinct systems, and structured data has a very different status in each.

Google’s AI surfaces. AI Overviews and related features are built by a company that has consumed structured data for over a decade, has documented parsers, and maintains an extensive structured data gallery describing exactly which types it reads and what they enable. Google’s guidance on its AI features situates these surfaces within the same systems as the rest of Search. Whatever influences Google’s understanding of a page plausibly influences what Google’s AI features say about it.

Conversational assistants answering from a language model. When ChatGPT or Claude composes an answer, the model is processing text. If the answer is grounded, a page was fetched and converted into something the model reads. There is no public documentation from these vendors describing JSON-LD parsing as a step in that pipeline, and the plain reading is that markup in a <script> tag is not what the model is reasoning over.

So: strong documented basis in the first case, no documented basis in the second. Any blanket answer conflates them.

Where schema does real work regardless

The interesting part is that the second case is not the whole story, because structured data influences AI answers through routes that do not require any model to parse it.

Entity resolution. This is the big one. Before an engine can say anything true about you, it has to know which entity you are — distinguishing your company from the seven other organisations sharing some version of your name. Organization markup with sameAs links to your authoritative profiles is one of the clearest disambiguation signals available, and Google’s organization structured data documentation describes exactly this use. Entity confusion is a top cause of engines getting brands wrong, and this is one of the few direct levers on it. Your brand is an entity before it’s a website covers why that matters more than it sounds.

Knowledge graph population. Structured data feeds knowledge bases. Knowledge bases inform search results, panels and downstream datasets. Those datasets get consumed by systems well beyond the one you marked up for. The path is indirect, slow, and real.

Forced precision. Marking up a product means committing to a price, an availability state, a currency. Marking up an organisation means committing to a founding date and a canonical name. Teams routinely discover during implementation that these facts are stated inconsistently across their own site — and the inconsistency, not the missing markup, was what confused the engines. The audit is often worth more than the artifact.

Machine-readable facts for anything that does parse. Not every consumer of your site is a chat model. Aggregators, shopping surfaces, agents and other intermediaries do parse structured data, and their number is increasing rather than decreasing.

Where it does not do what people hope

Being equally clear about the limits:

It is not a ranking or citation lever inside a chat answer. No evidence supports the idea that adding FAQPage markup makes a language model more likely to quote you. The mechanism is not there. If a page gets cited more after adding schema, look first at whatever else changed — usually the content was restructured at the same time, and restructuring is a known lever.

It does not rescue unclear content. Markup describes what is on the page. If the page never states the answer plainly, marking it up describes an absence in a standardised format. The GEO paper tested content modifications directly and found that adding quotations, statistics and cited sources raised visibility in generative engines, while keyword-style edits did not. Its finding is about the prose.

It is not a shortcut past authority. Nothing in schema.org asserts credibility. You cannot mark up expertise.

It does not help if it is wrong. Markup that disagrees with the visible page is worse than none — Google’s structured data policies treat mismatched markup as a violation, and an engine that catches a discrepancy has learned something about your reliability.

The practical answer

Implement structured data, for reasons that survive scrutiny, in this order:

1. Organization with sameAs, sitewide. Highest value, lowest effort, directly addresses entity disambiguation. If you do one thing, do this.
2. Product / Service for what you sell, with accurate pricing and availability. Increasingly consumed by agentic shopping surfaces — see agentic commerce.
3. Article with real dates and the author entity, for content you want treated as current.
4. FAQPage where genuine questions and answers exist on the page. Not as a wrapper around invented Q&A.

Then stop. Marking up everything markable has sharply diminishing returns, and each type is a maintenance liability that can drift out of step with the page.

Our schema generator and schema validator handle the mechanics, and the schema markup guide covers implementation properly.

How to tell whether it did anything

The measurement problem here is genuine, and worth being honest about rather than waving through.

Structured data changes are almost impossible to A/B test on your own site — you cannot serve half your crawlers a different Organization block and compare. What you can do is measure the thing schema is actually supposed to fix, which is not your visibility score but the accuracy of what engines say about you.

Record what the engines currently state as hard facts about your company: founding year, location, category, product line. Implement the markup. Wait a scan cycle or two, and check whether the wrong facts moved. That is a narrow, checkable claim, and it is the one the mechanism actually supports — unlike “our mention rate went up,” which has a dozen plausible causes.

Brand safety monitoring is the surface that tracks those factual assertions over time, because entity confusion shows up as confidently-stated wrong facts rather than as absence. If nothing changes after a couple of cycles, the honest conclusion is that entity resolution was not your problem — which is useful to learn early, since it redirects the effort to content.

The counter-argument worth taking seriously

The strongest sceptical case: structured data is a legacy SEO habit being carried forward into a paradigm that does not use it, and the entity-resolution argument is a rationalisation for continuing to do something familiar.

There is something to this. A lot of AEO advice is SEO advice with the nouns swapped, and it deserves the scrutiny. The entity argument survives it, though, for a specific reason: entity resolution is not a language model behaviour, it is an infrastructure behaviour, and the infrastructure — knowledge graphs, search indexes, reference datasets — demonstrably consumes structured data and demonstrably feeds what models retrieve. The claim is not “the model reads your JSON-LD.” It is “your JSON-LD shapes the corpus the model reads about you,” which is a weaker claim and a much better supported one.

The scepticism is correctly aimed at the inflated version — schema as a citation lever — and that version should be abandoned.

The one-line version

Structured data is infrastructure, not a visibility tactic. It helps machines agree on who you are, which is a precondition for anything else being right. It will not make a language model quote you, and any tool or consultant claiming otherwise is describing a mechanism that has never been documented by anyone.

Implement the four types above, keep them accurate, and spend the rest of your effort on the prose — which is the part with actual experimental evidence behind it. Structured data for AI visibility goes deeper on the reasoning.