AI Retrieval Behavior Across Semantic HTML and Typed Tabular Representations
Study Track: HTML as Structured Memory Surfaces
Study ID: TPI-HSM-2026-04
Status: Experimental Study
Publication Date: September 2026
Study Type: Controlled observational field study
Research Question
When equivalent knowledge is published as an addressable machine-readable dataset within a web document, does the serialization method affect how AI systems discover, interpret, extract, reason over, resolve, and cite that knowledge?
This study extends TPI-HSM-2026-03, Publishing Knowledge Alongside Content, which examines the broader proposition that publishers can expose human-oriented explanations and machine-oriented representations of the same underlying knowledge simultaneously.
TPI-HSM-2026-04 narrows the inquiry.
It does not ask whether knowledge should be published separately from explanation.
Once a machine publication exists, does the way that knowledge is serialized matter?
Background
Web publishing has traditionally optimized information for human consumption. Tables, prose, headings, lists, and other document structures provide sufficient context for a person to interpret the information presented.
Machine consumers face a different task.
An AI system may need to identify the relevant factual domain, determine the unit of analysis, infer datatypes, normalize formatted values, establish geographic and temporal scope, distinguish related categories, resolve entities, and reconstruct relationships before it can safely use information extracted from a webpage.
Structured machine publications can reduce some of this reconstruction burden.
The preceding TPI study examines an architecture in which a webpage functions simultaneously as:
Human publication + Machine publication
This study examines a further variable within the machine publication:
Machine publication
|
+-- Semantic HTML representation
|
+-- Typed tabular representation
The underlying knowledge may remain substantially equivalent while its serialization changes.
Experimental Context
The study uses county-level Medicare coverage resolver pages as the experimental surface.
These pages describe multiple bounded factual domains, including Medicare Advantage plan availability and market characteristics, standalone Medicare Part D plan availability and costs, Medicare Supplement market characteristics, Medicare Special Needs Plan availability, county-level Medicare coverage inventories, and plan enrollment leaders.
Each factual domain is declared as a Schema.org Dataset.
Each Dataset exposes an addressable on-page distribution through a DataDownload and fragment-specific contentUrl.
The experiment preserves this discovery architecture while varying the representation contained within the referenced fragment.
Semantic HTML Serialization
One implementation represents machine-oriented facts using semantic HTML structures with namespaced keys.
Conceptually:
<dt data-key="medicare_options.ma.plan_count.value">
value
</dt>
<dd>16</dd>
The representation is explicit and inspectable, but values remain primarily document-oriented and their datatype and constraints must be inferred or established by the consuming system.
Typed Tabular Serialization
The experimental implementation represents equivalent observations using a typed tabular resource based on Frictionless Table Schema conventions.
Conceptually:
{
"schema": {
"fields": [
{
"name": "ma_plan_count",
"title": "Total Medicare Advantage plans",
"description": "Number of Medicare Advantage plans available in the county for the stated plan year.",
"type": "integer",
"constraints": {
"minimum": 0
}
}
]
},
"data": [
{
"ma_plan_count": 16
}
]
}
The second representation supplies both the observation and an explicit data contract describing how the observation should be interpreted.
Experimental Design
The study uses two independently published Medicare information properties as parallel observation surfaces.
Equivalent Medicare knowledge
|
Schema.org Dataset
|
DataDownload
|
contentUrl
|
WebMEM fragment
|
+------+------+
| |
Semantic HTML Typed Tabular
serialization serialization
| |
MedicarePlans Medicare.org
MedicarePlans.com retains the existing semantic HTML serialization.
Medicare.org receives the experimental typed tabular serialization.
The experiment is designed to preserve, to the extent practical, the underlying source data, factual domains, geographic scope, plan year, Dataset boundaries, fragment addressing, resolver purpose, and human-facing subject matter.
The principal experimental variable is the machine serialization.
Typed Identity and Scope
The experimental tabular representation explicitly publishes normalized identity and temporal dimensions including:
county_name county_fips state_name state_code plan_year
County and plan-year summary records use the composite key:
county_fips + plan_year
This allows the machine representation to distinguish the display identity of a location from its normalized geographic identity.
It also establishes the grain of the observation:
one county × one Medicare plan year
Plan-Level Resolution
Where a dataset contains individual Medicare plan observations, the experimental representation may additionally expose CMS plan identity components:
contract_id plan_id segment_id cms_plan_key canonical_plan_url
The composite CMS plan key represents:
contract_id-plan_id-segment_id
This allows an enrollment or availability observation to resolve from a county-level dataset to a canonical plan-level public resource.
The study therefore examines not only factual extraction but also whether explicit identifiers and continuation targets affect entity resolution behavior.
Hypothesis
The study does not assume that either serialization will universally outperform the other.
The working hypothesis is:
Machine-readable knowledge represented as typed, constrained tabular records may reduce the interpretation and transformation required for AI systems to operate on factual web content, while semantic HTML may retain advantages for systems primarily optimized for document and text extraction.
Different retrieval systems may exhibit different behavior.
The study therefore evaluates serialization effects rather than attempting to establish a universally preferred representation.
Observation Model
Retrieval behavior will be evaluated across five stages:
DISCOVERY
Can the system locate the intended dataset?
↓
INTERPRETATION
Can it determine what the fields and records mean?
↓
EXTRACTION
Can it recover the correct observations?
↓
REASONING
Can it filter, compare, calculate, or resolve using them?
↓
GROUNDING
Can it preserve geography, time, category, entity,
and source context in the resulting answer?
These stages are evaluated independently because successful discovery does not imply correct interpretation, and correct extraction does not necessarily imply correct reasoning or grounding.
Example Retrieval Task
Consider the question:
Which county has the lower minimum Medicare Advantage maximum out-of-pocket amount for the same plan year, and what is the difference?
A document-oriented retrieval process may require a system to:
locate pages → identify the relevant Medicare category → locate the appropriate values → interpret formatted currency → infer geographic scope → infer plan year → normalize the values → compare them → preserve source context
A typed tabular representation potentially permits:
select Dataset → resolve fragment → verify county_fips → verify plan_year → select ma_min_moop → verify numeric type → compare → cite originating resource
The study observes whether this reduction in interpretive operations produces measurable behavioral differences.
Measures
Discovery accuracy — whether the intended Dataset and fragment are located.
Extraction accuracy — whether requested values are recovered correctly.
Context accuracy — whether geography, plan year, Medicare category, and other applicable scope are preserved.
Schema utilization — whether systems correctly recognize and use declared types, constraints, keys, and enumerations.
Calculation accuracy — whether numerical operations performed on published observations are correct.
Entity-resolution accuracy — whether CMS plan identities and canonical plan resources are correctly resolved.
Grounding and citation — whether answers remain traceable to the originating document or dataset distribution.
Unsupported inference — whether systems extend the published data beyond what it supports.
What This Study Does Not Test
TPI-HSM-2026-04 does not attempt to establish that Frictionless serialization is universally superior to HTML, that search engines are required to consume embedded tabular resources, that Schema.org Dataset guarantees ingestion of a referenced distribution, that structured data guarantees citation or ranking, that observed retrieval differences are necessarily caused by a single markup feature, or that one serialization should replace all other machine representations.
The study concerns observed retrieval behavior under differing serialization conditions.
Why This Matters
Publishing structured knowledge does not end the machine-interpretation problem.
A machine representation can still require substantial reconstruction if its values, types, boundaries, identifiers, and relationships remain implicit.
This introduces a second-order publishing question.
The first is:
Should publishers expose knowledge separately from human explanation?
The second is:
How should that knowledge be represented so machines can use it with the least unnecessary reconstruction?
TPI-HSM-2026-04 investigates that second question.
The distinction can be expressed simply:
Some representations make facts readable. Others may make records operable.
The study will observe whether that distinction produces measurable differences in real AI retrieval systems.
Relationship to TPI-HSM-2026-03
TPI-HSM-2026-03
Publishing Knowledge Alongside Content
Question:
Does separating machine knowledge from
human explanation affect retrieval?
↓
TPI-HSM-2026-04
Serialization Effects in Structured Memory Surfaces
Question:
Once machine knowledge is published,
does its serialization affect retrieval and use?