A product page can contain valid JSON-LD and still fail to supply an answer someone needs. A model can cite a page without having used its markup. Those two observations make “does schema work for AI?” too broad a question to guide an implementation.
The practical decision is where to put essential facts and how to test whether they are available. Put the answer in readable page content, maintain accurate structured data for the uses that support it, and inspect retrieval separately from citation. A passing markup validator closes only one part of that job.
What Google actually requires
Google's guidance for AI Overviews and AI Mode says a supporting page must be indexed and eligible for a search snippet. It recommends making important information available as text and keeping structured data consistent with what readers see. It does not require special schema.org markup for these AI features. [1]
That establishes a requirement, not a universal explanation of model behavior. It does not show that structured data can never contribute to a search system. Nor does it tell us whether a particular ChatGPT or Claude session retained a JSON-LD script. A statement about Google's eligibility rules cannot answer a question about another product's extraction pipeline.
Google also supports JSON-LD for eligible rich-result features, while making clear that valid markup does not guarantee those features will appear. Its quality rules call for current, relevant information that represents visible page content. [2] There is no reason to abandon those uses because the separate question of AI citation remains unsettled.
Four stages that need different evidence
Consider a hypothetical supplier whose minimum order is 250 units. The number appears in a script but not in the product description. A developer verifies the JSON and declares the page machine-readable. The buyer still cannot find the answer, and an extracted text document may not contain it either.
Now put the minimum order in the visible specifications and keep the structured representation consistent. This solves a concrete content problem without requiring a claim about which system reads which script. The fact is available to a person and to a reader that receives the visible text.
| Stage | Evidence to retain | What it cannot establish |
|---|---|---|
| Page delivery | Rendered content and the delivered document | That a search index is current |
| Index eligibility | Indexing and snippet eligibility checks | That an assistant selected the page |
| Content retrieval | Returned text, URL, timestamp and tool details | That every retained fact shaped the answer |
| Answer and citation | Exact prompt, response and cited URL | Which representation supplied the fact |
The table is a diagnostic model, not a disclosed architecture shared by all assistants. A product may use several routes, change providers, or reuse older material. Keep that uncertainty in the test record rather than collapsing every route into “the model read our website.”
Fetching a URL is not the same as reading its HTML
Anthropic's API documentation illustrates the distinction. Its web-fetch tool retrieves webpage content and returns a document to the conversation; the example response represents web content as plain text. Supported versions can also filter fetched content before it reaches the model's context. [3]
This documents an extraction step and, for some versions, a filtering step. It does not establish a blanket rule that every JSON-LD fact is removed in every Claude product. The API tool documentation is also not a specification for the consumer application's entire browsing stack. “Fetched successfully” and “this particular fact reached the answer” require separate observations.
The same distinction limits negative tests. If an assistant fails to repeat a fact, possible explanations include failure to retrieve the page, an old index entry, extraction, selection, or answer generation. The missing answer alone does not identify the stage responsible.
A citation advantage would not settle the mechanism
Suppose pages with schema are cited more often in an observational study. Those pages might also have better templates, clearer specifications, stronger internal links, or more editorial maintenance. Even a model that adjusts for measured differences cannot automatically remove differences that were never recorded.
An indexing benefit is one possible explanation for an association. Direct use of the markup is another. A content-quality difference is a third. Choose among them with evidence about the intermediate steps, rather than treating the citation outcome as a view inside the pipeline.
Conversely, a controlled live-fetch result describes the tool and conditions tested. It should not erase evidence about an indexed search route the experiment never exercised. Apparently conflicting findings can be about different operations.
An audit a content team can run
Start with a small set of important pages and a fact inventory. For each page, record the customer question, the exact answer, its visible location, its markup location if applicable, and the owner responsible for keeping it current. Prioritize facts that change a decision: compatibility, service area, eligibility, availability, and limits.
Check the visible answer first. A reviewer should be able to find and understand it without opening developer tools. Then compare the structured fields with the rendered page and validate the supported markup. Treat a mismatch as a content defect even if both versions are syntactically valid.
If testing AI retrieval, record the product, model or tool version where available, date, prompt, URL, and whether the task used an explicit URL fetch or a search. Save returned source text when the tool exposes it. Score fact retrieval, factual accuracy, and citation as separate outcomes. Repeat the same conditions before interpreting a change, and use a comparison group if the aim is to estimate an effect.
Do not hide critical facts in markup to manufacture a live production experiment. Use a controlled test page for extraction questions, and keep public customer information complete. A test of markup's contribution should hold visible copy and other material variables stable.
For a release decision, require agreement between the visible facts and structured data, successful delivery of the intended page, and an honest record of what was actually observed. Schema remains useful structured information. A citation promise needs its own evidence.