Testing structured data combinations

Motivation

Practitioners often stack many schema types hoping for AI/search benefits. Without combinatorial tests, recommendations remain anecdotal.

Types under comparison

Type	Typical surface
Person	Author / about (example)
Organization	Home / company (example)
CreativeWork	Generic creative asset
SoftwareApplication	Tool / product
Article	Editorial / experiment write-up
ProfilePage	Profile URL
CollectionPage	Hub / index of experiments

Facts

• These types are defined on schema.org with overlapping applicable properties.
• Valid markup does not guarantee consumption by any AI system; check with the Schema Markup Validator.

Assumptions

• More specific types (e.g. SoftwareApplication) may help when the page truly is software - but overtyping can create contradictions.
• AI interpretation differences, if any, may appear in product classification questions more than in brand name questions.

Unknowns

• Interaction effects between types in @graph.
• Whether ProfilePage / CollectionPage matter for LLM UIs at all.

---

Hypothesis

H₀: Across preregistered combinations, AI probe metrics do not differ beyond noise.

H₁: At least one enriched combination improves a preregistered primary metric (product-type classification accuracy or creative-work citation accuracy) relative to a Person+Organization-only baseline.

Primary metric must be locked before Running status.

---

Methodology

Design type

Sequential deployments of typed combinations on designated URLs, with washout or distinct URL slots to reduce carryover when possible.

Planned combination arms (initial)

1. Baseline: none / minimal
2. Person + Organization
3. • Article on experiment pages
4. • ProfilePage on person URL
5. • CollectionPage on lab index
6. • SoftwareApplication on tool URL (ties to Exp-007)
7. • CreativeWork where Article is intentionally omitted (contrast)

Exact schedule recorded in data/exp-009/schedule.md when Running.

---

Environment

Dimension	Value
Primary domains	Lab site + product/tool URLs
Surfaces under test	Home, about, experiment articles, tools index
AI systems probed	ChatGPT, Claude, Gemini, Perplexity
Measurement window	≥ 21 days per arm if sequential
Geographic / language scope	English
Tools used	Validators, @graph snapshots, probe codebook

---

Variables

Independent variables

• Schema type combination arm

Dependent variables

• Entity type classification accuracy (person/org/software/article)
• Conflict rate (mutually inconsistent types inferred)
• Citation of typed URLs when asking for “official docs/tools/articles”

Controlled / held constant

• Visible HTML content per URL when isolating markup
• @id stability across arms

Confounds to monitor

• Validator warnings ignored inconsistently
• CMS plugins injecting duplicate Organization nodes
• Parallel experiments changing same URLs

---

Procedure

1. Inventory current JSON-LD; freeze Arm 0.
2. Probe classification suite.
3. Deploy next arm; validate; wait; probe.
4. Diff extracted AI type assertions vs markup intent.
5. Stop or extend based on preregistered sample plan.

Prompt protocol

Is [URL] best described as a person, organization, article, software, or collection? Why?
What creative works or tools are published by [ORGANIZATION]?
Extract structured facts you believe are asserted about [URL].

---

Results

Status: Planned. No results claimed.

Arm	Classification accuracy	Conflict rate	Notes
Baseline	n/a	n/a	Not collected
Person + Organization	n/a	n/a	Not collected
+ Article	n/a	n/a	Not collected
+ ProfilePage	n/a	n/a	Not collected
+ CollectionPage	n/a	n/a	Not collected
+ SoftwareApplication	n/a	n/a	Not collected
CreativeWork contrast	n/a	n/a	Not collected

---

Observations

• Pending Arm 0 snapshot.

---

Limitations

• Sequential arms confound time with treatment.
• Multiple types on one page can be valid yet confusing to scorers.
• schema.org vocabulary updates may change recommended patterns.

---

Future Work

• True multi-URL randomized design with partners.
• Microdata / RDFa vs JSON-LD transport comparison.

---

References

1. Type docs: Person, Organization, CreativeWork, SoftwareApplication, Article, ProfilePage, CollectionPage
2. Root vocabulary: schema.org; transport: JSON-LD
3. Validators: Schema Markup Validator, Rich Results Test
4. Guidance: Google structured data introduction
5. Lab Exp-002, Exp-007
6. Site anchors: barbu.es, InfoWebPlus

---

Replication Notes

Export full JSON-LD for every arm. Never claim rich-result eligibility as an AI citation result.

Status: Planned
Author: George Barbu, Founder of InfoWebPlus