Knowledge Graph dan Citation: Hubungan Langsung atau Hanya Korelasi

Sebuah website merapikan knowledge graph.

Organization ID dibuat konsisten.

Entity pages diperbaiki.

Relationship dibersihkan.

Internal link diarahkan ke canonical pages.

Schema graph tidak lagi duplicate.

Tiga minggu kemudian, owned citation di beberapa AI answer naik.

Rapat langsung berubah optimistis.

“Berarti knowledge graph bikin citation naik.”

Kalimat itu mungkin benar.

Mungkin juga tidak.

Yang kita punya baru urutan waktu dan korelasi.

Knowledge graph berubah.

Citation berubah.

Apakah graph menyebabkan citation?

Kita membutuhkan lebih banyak evidence.

Di GEO, ini salah satu area yang paling mudah overclaim karena relationship-nya terasa intuitif. Kalau identity lebih jelas, kita berharap retrieval dan citation membaik.

Expectation itu reasonable sebagai hypothesis.

Tetapi hypothesis bukan causal fact.

Knowledge Graph Punya Direct Function yang Bisa Kita Verifikasi

Mari mulai dari yang pasti.

Knowledge graph internal dapat:

Menjaga canonical entity IDs.

Menyimpan relationships.

Menyimpan provenance.

Mendeteksi conflicts.

Mendukung schema generation.

Membantu internal linking.

Memetakan canonical pages.

Membantu content governance.

Mendukung entity search.

Ini direct, karena organisasi mengontrol system tersebut.

Kalau graph dibuat, duplicate Organization IDs turun dari 5 ke 1.

Direct effect.

Kalau graph digunakan untuk link recommender, 50 orphan entity pages mendapat contextual links.

Direct effect.

Kalau graph memvalidasi CEO current, stale bios ditemukan.

Direct effect.

Citation di external AI system berbeda.

Kita tidak mengontrol selection algorithm.

Itu downstream observation.

Schema juga Tidak Menjamin Citation

Structured data dapat membantu search systems memahami page/entity sesuai documentation tertentu.

Tetapi tidak ada dasar universal untuk mengatakan:

“JSON-LD graph menyebabkan LLM cite.”

Google sendiri membahas structured data untuk specific Search features dan understanding. AI platforms lain punya retrieval mechanism berbeda.

Some systems may not use schema directly.

Some rely on web search indexes.

Some use embeddings.

Some combine.

Some don't expose.

Therefore:

Knowledge graph → cleaner website representation.

Potentially → easier entity understanding/discovery.

Possibly → different source selection.

But arrows beyond first are hypotheses unless measured.

Korelasi yang Plausible Tetap Berguna

Don't swing too far.

If citation rises repeatedly after entity cleanup, that's useful.

We don't need perfect lab proof to act.

Need calibrated language.

Example:

“Owned citation coverage increased from baseline after deployment of canonical entity architecture. The change is consistent with the hypothesis that clearer source and entity relationships may support discoverability, but causal attribution cannot be isolated because content, crawl, model, and external source conditions also changed.”

This is useful.

Not cowardly.

It states:

What happened.

Hypothesis.

Limitation.

Decision can follow.

What Confounds Knowledge Graph Experiments?

Many things.

Content rewritten.

Titles changed.

Internal links changed.

Sitemap updated.

Schema changed.

Page speed improved.

Canonical URLs changed.

Indexation changed.

Media coverage.

Competitor changes.

Model update.

Search index refresh.

User query changes.

Source freshness.

Even if graph is one project, implementation often includes many interventions.

If citation rises, package may matter.

Not one edge.

Need change log.

If team wants stronger causal inference, stage changes.

Baseline Before Graph

Measure first.

Query set.

Repeated runs.

Mention.

Citation.

Recommendation.

Entity accuracy.

Source URLs.

Competitor.

Model/surface.

Dates.

Then deploy.

Without baseline, citation increase cannot be quantified.

“Feels better” not enough.

Also snapshot technical state:

Entity IDs.

Schema duplicates.

Canonical pages.

Internal link graph.

Orphans.

Conflicts.

Now before/after is meaningful.

Control Group Can Help

Suppose 200 comparable entity pages.

Treatment:
100 get graph cleanup.

Control:
100 unchanged during window.

Monitor citation/visibility.

If treatment moves more, evidence strengthens.

But assignment may not random.

Topic popularity differs.

Page authority differs.

Control contaminated by sitewide changes.

Still useful if designed carefully.

Call quasi-experiment.

Not guaranteed A/B.

Need analyst involvement.

Difference-in-Differences Bisa Dipakai

If treatment citation coverage:

20 → 35.

Control:

18 → 25.

Raw treatment +15.

Environment/control +7.

Relative difference +8.

This is simplified.

Real statistical assumptions matter.

Do not publish causal claim just because formula exists.

But framework helps.

If both groups rise equally, graph-specific hypothesis weakens.

If only treatment rises persistently, confidence increases.

Cross-Model Consistency

Graph cleanup.

Model A citation rises.

Model B rises.

Model C flat.

What means?

Potentially some systems benefit, some not.

Or noise.

If all three rise across same query cluster and multiple windows, pattern stronger.

Still not mechanism proof.

Maybe external media event affected all.

Need source analysis.

But cross-model helps robustness.

Don't average away difference.

Citation URL Shift Is More Directly Relevant

One useful metric:

Before:
Owned citations scatter across random articles.

After:
Canonical entity page gets more citations.

If goal was source consolidation, this is meaningful.

Example:
Canonical page share of owned citations:
5% → 42%.

This doesn't prove graph caused it.

But shows source selection moved toward intended canonical asset.

If no content changes except graph/internal relationship changes, hypothesis stronger.

Still external system unknown.

Entity Accuracy Often Has Stronger Logical Link than Citation

Knowledge graph cleanup may primarily improve consistency.

Measure:

Wrong founder.

Wrong ownership.

Wrong product relation.

Wrong location.

False split.

False merge.

If errors decline, graph objective achieved.

Citation may be secondary.

This is important.

Don't judge KG only by citation.

It is identity infrastructure.

Even if citation flat, reducing entity errors has value.

A graph that fixes internal schema duplication and corporate truth is not failure because ChatGPT did not cite more.

Citation obsession distorts architecture.

Correlation Can Be Spurious

Example:

Graph launch July 1.

Citation jumps July 5.

But July 3 major media feature published.

AI cites media and discovers brand more.

Graph unrelated.

Or model update July 4.

All brands see increased citations.

Or query panel changed.

Or competitor site down.

Timeline alone insufficient.

Annotate external events.

Control brands.

Use stable query set.

No causal storytelling by calendar.

Mechanism Evidence Bisa Membantu, tapi Sulit

If platform documentation explicitly states it uses certain structured data for a feature, mechanism more grounded.

But citation selection in general AI systems rarely fully transparent.

We may observe:

Crawler fetched entity page.

Index recognized structured data.

AI answer cited page.

Still chain doesn't prove graph property caused citation.

Need internal platform telemetry unavailable.

So causal ceiling exists.

Accept.

GEO measurement works under partial observability.

The answer is not make stronger claims.

The answer is calibrated confidence.

Three Levels of Claim

Level 1: Observation.

“Citation increased after KG deployment.”

Safe.

Level 2: Association.

“KG deployment is associated with higher observed citation in treatment pages relative to baseline/control.”

Needs data.

Level 3: Causation.

“KG deployment caused citation increase.”

Needs strong experimental/mechanism evidence.

Most public GEO case studies should stay Level 1-2 unless methodology robust.

This ladder is useful for editorial policy.

Don't write Level 3 because it sells better.

Knowledge Graph Can Influence Citation Indirectly Through Website Quality

Possible paths:

Canonical IDs reduce duplicate entity representations.

Graph drives contextual internal links.

Entity hub becomes more complete.

Schema becomes consistent.

Evidence links improve.

Crawler discovery improves.

User navigation improves.

All could indirectly affect discoverability/source fit.

But if citation rises, which path?

Unknown.

This is why “knowledge graph caused” too broad.

Maybe graph project improved content architecture, and content drove result.

The system-level intervention matters.

We can say:
“Entity-first architecture correlated with…”

More honest.

What Does Google Say?

Google's structured data documentation positions markup as a way to give explicit clues about page meaning and enable/qualify certain rich search experiences where requirements are met.

It does not promise general AI citation.

Google's AI features documentation emphasizes standard SEO fundamentals and states no special schema.org markup is required for AI Overviews or AI Mode.

This directly argues against “special KG markup = AI citation hack.”

Entity quality can be good practice.

No guaranteed citation switch.

Other AI Systems

OpenAI, Perplexity, other systems have different search/retrieval stacks.

Do not transfer Google structured-data behavior universally.

If platform does not document schema use for citation, treat as unknown.

Test behavior.

Never write:

“LLMs read your knowledge graph and rank it.”

Some may retrieve structured or graph-like data indirectly.

But generalized statement unsupported.

Platform facts separate from GEO interpretation.

Citation Is Source Selection, Not Entity Understanding Alone

AI can understand Brand X but cite Media Y.

Why?

Media Y supports claim.

More current.

More independent.

More relevant.

More concise.

Better matched query.

Entity understanding does not require owned citation.

This is central.

A perfect entity graph might improve correct identification, while citation still goes to independent source.

That's not failure.

For recommendation query, independent evidence may be appropriate.

Owned citation should not be sole KPI.

Graph + Evidence Is More Important than Graph Alone

If entity page says:

“We are trusted leader.”

No evidence.

Graph relation clean.

AI may still prefer media.

Add:

Public report.

Method.

Case study.

Certification.

Product documentation.

Now source utility improves.

Graph organizes.

Evidence substantiates.

Citation often depends on source content usefulness.

Don't confuse semantics with evidence.

Knowledge graph tells who/what.

Evidence tells why claim is supported.

Both.

But again no guarantee.

How to Design a Better Test

Research question:
Does entity-graph cleanup associate with increased owned citation coverage?

Hypothesis:
Pages with canonical entity IDs, resolved duplicates, and contextual relationships will show higher owned citation coverage than comparable control pages.

Baseline:
4 weeks.

Treatment:
Defined graph changes.

Control:
Matched pages.

No content rewrite if isolating graph.

Query set:
Fixed.

Runs:
Repeated.

Models:
Separate.

Metrics:
Owned citation coverage.
Entity accuracy.
Canonical page share.
Source diversity.
Volatility.

Change log:
Everything else.

Analysis:
Treatment-control difference.

Limitations:
External model changes.

Now result is publishable.

Not perfect.

But credible.

Null Result Is Valuable

Suppose graph cleanup.

Entity accuracy improves.

Citation flat.

That teaches:

KG improved identity but not observed source selection.

Maybe citation is driven more by evidence/content.

Maybe effect too small.

Maybe window short.

Maybe model doesn't use signals.

Don't hide.

This prevents cargo cult.

GEO matures when null results are allowed.

Otherwise every project “works”.

Science theater.

Negative Result

Citation decreases after cleanup.

Panic?

No.

Check:

Indexing.

Canonical error.

Broken links.

Schema invalid.

Page content changed.

Crawler blocked.

Model drift.

Competitor.

If technical regression, fix.

If no issue, could be noise.

Need repeated windows.

Never rollback graph correctness solely due one volatile citation drop unless evidence strong.

Entity truth has intrinsic value.

Knowledge Graph Project Should Have Its Own Success Metrics

Coverage of canonical IDs.

Relationship provenance.

Duplicate node reduction.

Stale relation rate.

Schema consistency.

Entity page completeness.

Internal-link orphan reduction.

Evidence mapping.

Public/private separation.

These are controllable.

AI citation:
external KPI.

Important but secondary.

This separation protects engineering from chasing black-box outcomes.

Executive Reporting

Bad:
“Knowledge graph increased AI citations 58%.”

Unless strong evidence.

Better:
“After entity-graph deployment, owned citation coverage rose from X to Y in the monitored query panel. Entity accuracy also improved. The period included internal-link and content changes, so the result is treated as association rather than isolated causal effect.”

This is decision-grade.

CEO can still act.

No need hype.

Korelasi Bukan Kata Kotor

Marketing often treats “correlation” as weak.

Actually correlation is useful signal.

Repeated, strong, consistent correlations across time/models can guide strategy.

What matters is not pretending.

We operate under uncertainty.

Many business decisions use observational evidence.

The quality comes from:

Method.

Baseline.

Controls.

Transparency.

Replication.

Not from using word “caused”.

GEO Needs Causal Humility

AI search is dynamic.

Black-box elements.

External systems.

Limited telemetry.

That doesn't mean measurement impossible.

It means we should distinguish:

Direct implementation effect.

Observed external association.

Causal inference.

Speculation.

Knowledge graph has direct benefits we control.

Citation is an external behavior we observe.

The relationship between them can be tested.

Sometimes strongly.

Rarely perfectly.

If citation rises after graph cleanup, celebrate the observation.

Then ask the harder question:

“Apakah kita sudah punya cukup evidence untuk mengatakan graph yang menyebabkan?”

Usually answer:

“Belum sepenuhnya.”

That's okay.

A credible GEO practice doesn't need every improvement to be magic.

It needs enough discipline to know what we actually learned.