The published page is not necessarily the searchable document
Before paying for more structured-data work to improve AI answers, ask which version of your page—and which facts—the retrieval system will actually receive. Adding information to a page and getting it into the searchable input are separate jobs.
JSON-LD is machine-readable data embedded in a page. Extracted prose is the text a processing system keeps. A richer entity page can also expose readable facts about connected people, products or organizations. Those inputs may describe the same business without giving the system the same information.
A controlled study, “Structured Linked Data as a Memory Layer for Agent-Orchestrated Retrieval,” makes the difference unusually concrete. Its retrieval pipeline treated each document as a single text field, cut off at roughly 20,000 characters for embedding—the step that creates a searchable numerical representation. In 88% of its HTML-with-JSON-LD documents, the document exceeded that limit. The JSON-LD block began at a median position of character 18,510, leaving it vulnerable to partial or complete truncation.
This was not a clean test of what a dedicated structured-data parser could do. Some of the added markup could disappear before retrieval had a chance to use it.
Richer pages helped, but they also supplied more facts
The experiment covered 158 entities and 349 template-generated questions across editorial, legal, travel and e-commerce domains. In its standard retrieval-and-answer pipeline, adding JSON-LD to HTML produced a small, statistically significant improvement in model-rated answer accuracy: 3.89 versus 3.62 on a five-point scale. Completeness did not improve significantly after statistical correction. “Markup did nothing” would therefore be the wrong reading, even with the ingestion problem.
Enhanced entity pages scored substantially higher: 4.69 out of five for accuracy in the same pipeline. They bundled embedded JSON-LD, type breadcrumbs, visible links to related entities and instructions for agents.
Crucially, those pages also made linked-entity facts directly available where baseline pages sometimes supplied only opaque identifiers. The researchers changed both how information was presented and how much useful information was accessible. They did not isolate the effects of formatting, instructions, markup or added facts. The result supports the combined intervention, not a shopping list of independently proven features.
Ask to see the input, not just valid markup
The useful purchasing distinction is between adding structure and getting usable facts into the system. A valid JSON-LD block cannot help a particular retrieval pipeline through information that pipeline never receives. Conversely, a richer page may help because it exposes missing facts, without proving that its special formatting caused the improvement.
If you control an internal AI assistant, or your supplier can inspect its ingestion process, ask to see the actual document sent for indexing—not just the original HTML or a schema-validation report. Check whether embedded data survives extraction, where truncation happens, and whether important facts about related entities remain readable. Those checks separate a parsing problem from a content gap.
For external AI-search services, that internal view may not be available. This study does not establish how Google or Bing process your pages, and it did not measure live visibility, citations, traffic or sales. A proposal for more schema work should explain which input it expects to change and what evidence would show that change reaching the target system. Passing markup validation answers a different question.
How much weight to put on the result
These were model-judged answers in a controlled Vertex AI and Gemini experiment, not human assessments of live search results. Reference answers came from the same knowledge graphs used to construct the enhanced pages, creating a potential advantage for those pages. The study did not separate that advantage from presentation. Results also varied by domain, so the aggregate gain is not an expected uplift for an individual business.
All authors are affiliated with WordLift, whose Linked Data Platform supplied the knowledge graphs. Google Cloud provided credits, and the experiments ran on Google Cloud.