Schema untuk Dataset: Apa yang Harus Ada sebelum Menyebut Halaman sebagai Dataset

Ada halaman berisi tabel.

“Pakai Dataset schema aja.”

Ada PDF dengan angka.

“Dataset.”

Ada artikel dengan 20 statistik.

“Dataset juga.”

Ada dashboard.

“Pasti Dataset.”

Belum tentu.

Dataset bukan label untuk “halaman yang punya data”.

Schema.org mendefinisikan Dataset sebagai badan informasi terstruktur yang menggambarkan topik tertentu. Google Dataset documentation juga memberi contoh yang cukup luas: tabel atau CSV, kumpulan tabel, file proprietary yang berisi data, koleksi file yang bersama-sama membentuk dataset, structured object, image yang menangkap data, sampai file machine-learning tertentu.

Jadi scope Dataset memang luas.

Tetapi tetap ada perbedaan antara:

Data.

Artikel yang membicarakan data.

Report yang menganalisis data.

Dashboard yang memvisualisasikan data.

Dataset landing page.

File distribusi dataset.

Kalau semuanya disebut Dataset, structured data kehilangan meaning.

Pertanyaan pertama bukan “ada angka nggak?”

Pertanyaannya:

“Apakah ada body of structured information yang memang menjadi object tersendiri?”

Artikel Bukan Dataset Hanya karena Ada Tabel

Contoh:

Judul:
“Tren AI Search Indonesia 2026.”

Artikel punya tabel:

ChatGPT usage.
AI referral.
Query behavior.

Tetapi angka berasal dari berbagai source eksternal.

Tidak ada dataset original.

Tidak ada file.

Tidak ada method collection sendiri.

Page adalah Article atau Report depending format.

The table is presentation.

Not necessarily Dataset.

Sebaliknya:

“Dataset Observasi AI Visibility Indonesia Juli 2026.”

Page menjelaskan:

1.000 observations.

Defined query set.

Date range.

Fields.

Methods.

CSV download.

License.

Creator.

Version.

Now Dataset is real.

The difference is data object autonomy.

Can someone describe, cite, download, reuse, or inspect dataset independently from prose?

If yes, stronger Dataset case.

Metadata Bukan Data Itu Sendiri

Google's Dataset documentation explicitly distinguishes dataset metadata from dataset contents.

Metadata describes:

What dataset is about.

Variables.

Creator.

Publication.

License.

Distribution.

It doesn't need to include every data value.

This is important.

Dataset landing page can be HTML.

The actual data may be CSV.

JSON.

Excel.

Parquet.

ZIP.

API.

The HTML page describes.

Structured data marks metadata.

`distribution` points to downloadable forms.

So don't dump 100,000 rows into JSON-LD.

That's not the job.

Schema is catalog metadata.

Dataset is underlying information body.

Minimum Identity: Name dan Description

Google's Dataset documentation lists `name` and `description` as required properties for its Dataset Search support.

That makes sense even outside Google.

A dataset should have a unique descriptive name.

Not:
“Data 2026.”

Better:
“GEO.or.id AI Search Citation Observation Dataset, Indonesia, July 2026.”

Description should say:

What it contains.

Coverage.

Period.

Subject.

Basic method.

No marketing fluff.

If two datasets both named “AI Data”, identity becomes weak.

Dataset naming should work like product naming.

Stable.

Specific.

Version-aware.

Creator Harus Jelas

Who created data?

Person?

Organization?

Research team?

Government agency?

Publisher?

Scraper?

This matters for provenance.

Google recommends `creator`.

For institutional dataset, Organization often appropriate.

If research collaboration:
multiple creators possible.

Don't assign CEO as creator because founder wants personal entity visibility.

Creator should reflect actual authorship/data stewardship.

If collection automated by system but Organization owns project, visible methodology can explain.

Provenance is trust.

Not schema decoration.

Publisher dan Creator Bisa Berbeda

University researcher creates dataset.

Repository publishes.

Creator:
Researcher/lab.

Publisher:
University repository.

Or NGO creates data.

Government portal republishes.

Keep roles.

Don't force one Organization into every field.

This helps citation and provenance.

Dataset ecosystem often involves republication.

Identity mapping matters.

Period dan Coverage Harus Ada secara Editorial

Schema properties can represent temporalCoverage and spatialCoverage.

But before markup, landing page should clearly explain.

Temporal:
January–July 2026.

Spatial:
Indonesia.

Topic:
AI search observations.

Population/scope:
Defined query panel.

If user can't tell what period and area data covers, dataset page isn't ready.

A dataset without scope is easy to misuse.

Especially when numbers get cited by AI.

“Indonesia data” from Jakarta-only sample needs limitation.

Metadata must protect against overgeneralization.

Variables Harus Didefinisikan

Dataset says:

citation_rate.

mention.

recommendation.

What do these mean?

If column definitions absent, reuse becomes dangerous.

Provide data dictionary.

Field.

Type.

Definition.

Allowed values.

Unit.

Missing value.

Example:

`brand_mentioned`: boolean.
Meaning: brand name/entity present in final answer.

`owned_citation`: boolean.
Meaning: at least one source URL belongs to brand-controlled domain.

`query_id`: stable ID.

This is not required by every schema consumer.

But it is required by good dataset publishing.

Semantic clarity matters more than markup completeness.

Methodology Harus Terlihat

How collected?

Manual?

API?

Browser test?

Logged in?

Logged out?

Location?

Model?

Repeat count?

Sampling?

Data cleaning?

If dataset supports research claim, method is essential.

Without method, “dataset” becomes spreadsheet with unknown origin.

Schema cannot rescue.

Create methodology section or linked methodology page.

Version method.

If query set changes, document.

If platform behavior changes, note.

Dataset should be reproducible enough for intended use.

Tidak Semua Dataset Harus Downloadable Publik

Some data may be restricted.

Privacy.

License.

Commercial agreement.

Security.

You can have dataset metadata without public raw download in some contexts.

But if Google Dataset Search discovery is goal, distribution details are particularly useful when download exists.

Google documents `DataDownload`, where `distribution.contentUrl` identifies the download location and `encodingFormat` describes format.

If no public distribution, don't invent URL.

Maybe dataset page describes access request.

Maybe aggregate only.

Public structured data should reflect actual access.

Jangan Publish Sensitive Raw Data demi Schema

Person-level data.

Customer email.

Precise private location.

Health.

Financial account.

Employee records.

Prompt logs containing personal info.

Don't publish because “dataset needs distribution.”

Privacy/legal basis first.

Anonymization.

Aggregation.

Access control.

Data minimization.

Dataset SEO is never reason to expose personal data.

If public dataset cannot be safely released, publish metadata/aggregate if appropriate or keep private.

Entity strategy doesn't override ethics.

License Harus Nyata

Can people reuse?

CC BY?

CC0?

Custom license?

All rights reserved?

Unknown?

Don't put Creative Commons URL just because Google recommends license information.

License is legal statement.

Get approval.

Visible page should state license.

If no reuse allowed, say.

If attribution required, give citation format.

Licensing clarity is one of the most useful dataset metadata fields.

Not a ranking trick.

Versioning Bukan Optional untuk Data yang Berubah

Dataset v1.

v2.

Monthly edition.

Rolling dataset.

Which?

If page updates in place, old research citation may no longer reproduce.

Define strategy.

Immutable snapshots:
2026-07.

2026-08.

And maybe latest alias.

Or version:
v1.0.

v1.1.

Document changes.

Dataset identity should survive updates but maintain historical references.

Internal ID.

DOI if institutional workflow supports.

`identifier`.

sameAs only for same dataset in canonical repository.

Do not sameAs materially different versions if identity differs.

Google Dataset guidance specifically distinguishes simple republication from significantly changed datasets. For changed derivatives, `isBasedOn` is more appropriate than asserting sameAs.

This is good provenance discipline.

Distribution Harus Dipisahkan dari Landing Page

Landing:
HTML description.

Distribution:
CSV file.

JSON file.

ZIP.

API endpoint maybe as access mechanism depending model.

One dataset can have multiple distributions.

CSV full.

JSON subset.

Excel.

Parquet.

Google's documentation notes that different distributions do not necessarily have to be exactly equivalent in all details.

But page should explain.

If CSV excludes sensitive columns while Parquet internal is full, don't imply same public content unless appropriate.

Distribution metadata should be honest.

encodingFormat helps.

contentUrl points actual resource.

No fake download.

Dataset vs DataFeed

Schema.org DataFeed is subtype of Dataset for a single feed providing structured information about entities/topics.

If page represents continuously updated feed, DataFeed may be more precise.

Example:

Real-time station measurements.

Product feed.

Activity feed.

But not every API is DataFeed.

Again, semantics.

What is object?

Static dataset?

Series?

Feed?

Catalog?

Choose accordingly.

Dataset vs DataCatalog

Repository with 500 datasets:
DataCatalog.

Individual data package:
Dataset.

Don't mark catalog landing page as one giant Dataset if it primarily lists many independent datasets.

Catalog can contain datasets.

This hierarchy is useful.

Example:

GEO.or.id Data Lab:
DataCatalog.

AI Citation Observation July:
Dataset.

AI Crawler Logs Sample:
Dataset.

Each has own page.

Clean.

Dataset vs Report

Report:
Interpretation.

Findings.

Narrative.

Dataset:
Structured information.

They often live together.

Report based on Dataset.

Dataset supports Report.

Don't collapse.

A report can cite dataset.

Dataset page can cite research paper/report.

Google Dataset guidance even includes `citation` for related academic publications and makes clear that it is not for citing the dataset itself.

This distinction is valuable.

Build relation.

Not identity confusion.

Dataset vs ScholarlyArticle

ScholarlyArticle:
Academic/scholarly article.

It may include data.

Supplementary dataset can be separate Dataset.

Don't call paper Dataset.

Don't call CSV ScholarlyArticle.

If one research landing page hosts both:
WebPage.
ScholarlyArticle.
Dataset.
Relationship.

But only if each object truly exists.

No need create 10 nodes for aesthetics.

Apa yang Harus Terlihat di Dataset Landing Page?

Minimum practical content:

Dataset name.

Description.

Creator/publisher.

Purpose.

Coverage.

Collection period.

Methodology.

Variables/data dictionary.

Version.

Update frequency/status.

License.

Citation guidance.

Distribution/access.

Limitations.

Contact.

Related report/publication.

This may be 800 words.

Could be 2,000.

Length not key.

Completeness for reuse is key.

If page only says:
“Download our exclusive dataset here.”

Thin.

Not enough.

Dataset Search 2026: Nuance yang Harus Dipahami

Google continues documenting Dataset structured data specifically for making datasets easier to find in Google Dataset Search.

Google's Search documentation updates have also clarified that Dataset structured data is used by Dataset Search, not as a general Google Search structured-data appearance.

This distinction matters.

Do not sell Dataset schema as standard blue-link/rich-result ranking tactic.

If institutional goal is data discoverability:
use it.

If page is just blog article:
don't force.

If AI search goal:
Dataset architecture may make evidence easier to publish and verify, but there is no guarantee it directly increases AI citation.

Measure downstream separately.

Dataset Schema Tidak Membuat Data Credible

A fabricated CSV with perfect JSON-LD is still bad data.

Credibility comes from:

Method.

Provenance.

Sampling.

Quality control.

Transparency.

Limitations.

Reproducibility.

License.

Author.

Institution.

Structured data helps describe.

It cannot validate research.

This is critical for GEO evidence.

Don't convert internal marketing spreadsheet into “Dataset” just to look scientific.

If data cannot withstand basic questions, don't elevate.

Quality Gates sebelum `@type: Dataset`

Ask:

Is there a real structured information body?

Is it distinct from the article/report?

Does it have stable identity?

Name?

Description?

Creator?

Scope?

Period?

Variables?

Method?

Version?

License/access?

Distribution if public?

Limitations?

Privacy cleared?

Can another person understand what each field means?

If not, fix dataset first.

Then schema.

The strongest Dataset implementation isn't the one with most properties.

It's the one where someone can arrive, understand exactly what the data contains, download or request it appropriately, cite it correctly, and know what conclusions it cannot support.

Only after that does the word “Dataset” really mean something.