Structured Linked Data as a Memory Layer for Agent-Orchestrated Retrieval (Volpini, Raad, Gamba & Riccitelli, 2026)
arXiv 2603.10700, submitted 2026-03-11, 33 pages, with a reproducibility appendix (datasets, evaluation framework, page templates released). All four authors work at WordLift (Rome); two are its CEO and CTO, and the winning page format and the Linked Data Platform that serves it are WordLift's product. Google Cloud supplied credits. T3 for that reason — the maker measuring the maker's thing — and ingested at the curator's instruction on 2026-09-14 after two research passes had declined it on the bar. The numbers were already on growth edge 1; this page holds the design.
The experiment
Four sites, 158 entities, 349 queries: WordLift's own blog (16 entities / 22 queries), Express Legal Funding (32 / 111), SalzburgerLand tourism (54 / 79), BlackBriar e-commerce (56 / 137).
Three ways of publishing the same page: plain HTML with the JSON-LD stripped; HTML + JSON-LD
as the site actually ships it; and an enhanced entity page — a natural-language summary, visible
links to related entities with dereferenceable URIs, llms.txt-style instructions to agents,
schema.org type breadcrumbs, and the JSON-LD embedded. Two ways of retrieving: standard RAG
(Vertex AI Vector Search 2.0, K=10, Gemini 2.5 Flash writes the answer) and agentic RAG (Google
ADK, ReAct loop, three tools: document search, entity-link traversal up to two hops, and a
knowledge-graph search). Seven conditions, 2,443 evaluations, scored 1–5 for accuracy and
completeness by Gemini 3.0 Flash as judge.
The result, in one table
| condition | accuracy | vs plain-HTML standard RAG |
|---|---|---|
| C1 plain HTML, standard | 3.62 | — |
| C2 HTML + JSON-LD, standard | 3.89 | Δ +0.27, d = 0.18 |
| C3 enhanced page, standard | 4.69 | Δ +1.04, d = 0.60 (+29.6%) |
| C4 plain HTML, agentic | 4.36 | agent compensates for missing structure |
| C5 HTML + JSON-LD, agentic | 4.40 | |
| C6 enhanced page, agentic | 4.70 | Δ +1.04, d = 0.61 (+29.8%) |
| C6+ enhanced with richer navigation | 4.85 | Δ +0.06 over C6, d = 0.08, not significant |
Hypothesis 1 — that JSON-LD alone helps — is rejected. The reason is mechanical and is the
paper's most useful finding: the vector store embeds each document as one text field truncated at
about 20,000 characters, and "82% of plain HTML and 88% of JSON-LD documents exceeded" it; the
<script type="application/ld+json"> block starts at a median position of character 18,510,
right at the cut. The markup was mostly never indexed. What worked was link materialization:
writing the graph's facts (a hotel's coordinates, cuisine, what contains it; an article's about and
mentions) out as readable prose on the page. Gains were largest where the graph knew things the
HTML did not — SalzburgerLand Δ +2.47, the WordLift blog Δ +2.73 — and near zero for the e-commerce
site (Δ +0.07), whose product facts were already visible.
What the authors concede, and what a reader should add
The paper's own limitations section is unusually complete: the enhanced format changes layout and adds graph-derived facts at once, with no ablation separating the two; Vertex's flat-text ingestion is not how a production search engine handles JSON-LD (which it extracts as a separate signal); every stage — retrieval, generation, judging — is a Gemini model; the ground truth is derived from the knowledge graph, so the evaluation rewards formats that present that graph; one retrieval system; 349 queries.
Add to that what the affiliation implies: the "enhanced entity page" is WordLift's product output, the Linked Data Platform is WordLift's, and one of the four test sites is WordLift's own blog. The design is sound and released; the conclusion — publish your graph as prose and the agent will read it — is also the pitch.
What it does to this spoke's edge
Growth edge 1 asks for evidence that any AI system consumes the artifacts this spoke's proposals publish. This is the first source in the spoke to run that test, and it answers in a direction the proposals did not want: the machine-facing artifact (schema-org JSON-LD, the thing entitymap and yoast-schema-aggregation exist to aggregate) was not consumed by a mainstream flat-text RAG pipeline, because it sat past the truncation point. The facts got through when written as text a human could read. That is the same shape as the GEO result recorded on the edge: prose moves generative retrieval, markup does not, at least in the systems tested. The edge stays open for a neutral party's version of this experiment, and for a retrieval system that parses JSON-LD out before embedding — the case the paper explicitly does not cover.
Related
structured-data-for-ai · schema-org · entitymap · yoast-schema-aggregation · knowledge-representation · synthesis