What you publish is not necessarily what gets indexed

Before paying for more structured data to improve how AI systems answer questions about your business, ask a simpler question: what version of your page—and which facts—will the retrieval system actually see? Putting data on a page and getting it into a searchable index are two different problems.

Consider three ways information can exist on or come from a webpage: JSON-LD (machine-readable data embedded in the code), extracted prose (the plain text a scraper keeps), and a richer entity page (which explicitly spells out facts about connected people, products, or organizations in the visible text). All three might describe the same business, yet hand entirely different information to an AI pipeline.

A controlled study, “Structured Linked Data as a Memory Layer for Agent-Orchestrated Retrieval,” shows just how much this gap matters in practice. The researchers used a retrieval pipeline that treated each document as a single block of text, cutting it off at roughly 20,000 characters before generating embeddings (the numerical representations used for search). In 88% of their HTML documents containing JSON-LD, the total length blew past that limit. Because the JSON-LD block started at a median position of character 18,510, it was vulnerable to partial or complete truncation.

This wasn't a clean test of what a dedicated structured-data parser could do. Some of the added markup could disappear before the retrieval step ever saw it.

Richer pages performed better, but they also contained more facts

The experiment tested 158 entities and 349 template-generated questions across editorial, legal, travel, and e-commerce sites. In a standard retrieval-and-answer setup, adding JSON-LD to standard HTML yielded a small, statistically significant bump in model-rated answer accuracy: 3.89 vs. 3.62 on a five-point scale. Answer completeness didn't show a statistically significant improvement after correction. So even with the truncation problem, it would be wrong to say the markup did nothing.

However, “enhanced entity pages” scored much higher, hitting 4.69 out of 5 for accuracy in the exact same pipeline. These pages bundled together embedded JSON-LD, type breadcrumbs, visible links to related entities, and specific instructions for AI agents.

More importantly, these enhanced pages spelled out facts about linked entities directly in the text, whereas the baseline pages sometimes supplied only opaque IDs. In short, the researchers changed both the formatting and the actual volume of useful information on the page at the same time. Because they didn't isolate the variables—formatting, agent instructions, markup, or raw added facts—the study supports the combined intervention, not that each individual feature is worth buying on its own.

Inspect the ingested input, not just the validator report

When buying SEO or data-engineering work, separate the act of adding structure from the act of getting usable facts into the system. A perfectly valid JSON-LD block is useless to a retrieval pipeline that chops it off before indexing. Conversely, a richer page might improve AI answers simply because it finally states the missing facts out loud, not because of its clever formatting.

If you run an internal AI assistant—or if your vendor can look under the hood of their ingestion pipeline—ask to see the exact text string sent for indexing. Don't settle for looking at the raw HTML or a passing grade from a schema validator. Check whether your embedded data survives text extraction, see where character limits cut the document off, and verify that key facts about related entities are still readable at the end of the pipe. That is how you tell a parsing problem apart from a content gap.

For outside AI-search engines like Google or Bing, you may not be able to inspect the internal pipeline. And this study doesn't prove how commercial search engines process your pages, nor did it measure live search visibility, citations, traffic, or sales. Still, if someone pitches you on more schema work, they should be able to explain which specific input they expect to change and how they know that change will survive ingestion. Passing a markup validation test answers a completely different question.

Caveats and context

Keep in mind how these results were graded: an AI model judged the answers inside a controlled Vertex AI and Gemini environment. These were not human evaluations of live search results. Furthermore, the “ground truth” reference answers were pulled from the exact same knowledge graphs used to build the enhanced entity pages, creating a potential home-field advantage for those pages that the study didn't separate from the formatting changes. Performance also swung noticeably depending on the industry domain, so the average score bump isn't a guaranteed lift for any single site.

All of the paper's authors are affiliated with WordLift, which provided the Linked Data Platform and knowledge graphs used in the test. The experiments were run on Google Cloud using Google Cloud-provided credits.