The practical brief

The finding

In one retrieval pipeline, enhanced entity pages earned higher answer-quality scores than ordinary HTML or HTML with added JSON-LD markup. [c1] [c2] [c3] [c5]

Why it matters

A valid markup report and a useful answer are different deliverables. Before buying an AI-visibility package, ask your supplier to show the information its system actually ingests and the answers it produces.

Same retrieval method, three page formats
Page formatMean score out of 5Standard deviation
Plain HTML3.621.82 [c1]
HTML + JSON-LD3.891.70 [c2]
Enhanced entity page4.690.95 [c3] [c5]
Table 2, standard RAG conditions C1–C3. Scores are an AI judge’s answer-accuracy ratings from 1 to 5. ± values are standard deviations, not confidence intervals. The enhanced pages also add linked facts. Study source

What to try

BLURSOR’s practical interpretation

Choose one important customer question. Confirm that the necessary facts are visible on the page, then ask your supplier for the retrieved passage and the resulting answer. Use that example to identify what is missing.

Study boundary

The enhanced pages added linked facts as well as changing presentation. The experiment cannot isolate a formatting benefit or predict Google or ChatGPT visibility.

Make the missing fact easy to find

A useful supplier demonstration starts with a question your customer would ask. For example: which locations do you serve, what does a service include, and what would rule out a purchase? Write the expected answer before inspecting the AI output.

If a needed fact exists only inside markup, a download or a linked page, ask how the system obtains it. Request an example from the system you are paying for. A generic promise that structured data improves visibility is not a substitute for that demonstration.

  • Keep the visible answer and the markup consistent.
  • Ask which exact passage reached the answer generator.
  • Agree on a useful answer test before approving a wider rewrite.

Evidence [c5] [c6]

The ingestion limit matters

This system flattened pages into text and truncated long inputs. Added JSON-LD was often partly or fully cut before indexing. A system that separately parses structured data may behave differently.

Evidence [c6]

The authors disclose a circularity risk

Reference answers came from the same knowledge graph that supplied the enhanced pages. The authors acknowledge that closer matching to those facts may help scores. They also did not isolate the extra information from the formatting changes.

Treat the result as a prompt to inspect your own evidence path. It does not justify a blanket removal of schema, a guaranteed visibility percentage, or a claim that one page template will work everywhere.

Evidence [c4] [c5]

The source and its limits

Structured Linked Data as a Memory Layer for Agent-Orchestrated Retrieval

Andrea Volpini, Elie Raad, Beatrice Gamba, David Riccitelli · 2026-03-11

349 queries in a WordLift study using Vertex AI retrieval and an AI judge; comparison uses standard RAG only.

  • One retrieval stack, not a public-engine visibility test.
  • The authors work for WordLift, whose knowledge graph infrastructure is used.
  • Table means are reproduced directly; inconsistent reported deltas elsewhere are not redrawn.

Evidence behind this briefing

[c1] Plain HTML: mean accuracy 3.62, standard deviation 1.82.

Section 4, Table 2, C1 · Read the study

[c2] HTML with JSON-LD: mean accuracy 3.89, standard deviation 1.70.

Section 4, Table 2, C2 · Read the study

[c3] Enhanced pages: mean accuracy 4.69, standard deviation 0.95.

Section 4, Table 2, C3 · Read the study

[c4] The authors disclose potentially circular knowledge-graph-derived reference answers.

Section 5.5, KG-derived ground truth · Read the study

[c5] Additional facts and page presentation were changed together.

Section 5.5, Information content versus presentation format · Read the study

[c6] Flat-text ingestion could cut added JSON-LD before indexing.

Section 5.5, Flat-text ingestion architecture · Read the study